Developer AI tools, side by side

An AWS service that gives your code access to models from many AI companies through one API, so you can switch models without switching providers. It also includes AgentCore for building and running agents, Knowledge Bases for connecting models to your own data, fine-tuning and Guardrails safety filters.

Free to start: No separate free tier. New AWS customers get up to $200 in Free Tier credits; trying a model in the Bedrock playground is one of the activities that earns them.

Based in: 🇺🇸 United States, Seattle, Washington (Amazon Web Services, Inc.). Source

Terms: RAG · Fine-tuning · Guardrails

Checked Oct 7, 2026 on the company’s own site: Amazon Bedrock · AWS Free Tier · AWS Free Tier credits

Note: Google renamed Vertex AI to Gemini Enterprise Agent Platform.

Google Cloud's platform for building with AI models and agents, and the successor to Vertex AI. Its Model Garden offers more than 200 Google and third-party models, including Gemini, Anthropic's Claude models and open models such as Gemma, with tools to tune, evaluate and deploy them.

Free to start: New Google Cloud customers get up to $300 in free credits to try Agent Platform and other Google Cloud products.

Based in: 🇺🇸 United States, Mountain View, California (Google LLC). Google LLC is organized under Delaware law; its address is in Mountain View, California. Source

Terms: Fine-tuning · Foundation model

Checked Oct 7, 2026 on the company’s own site: Agent Platform product page · Vertex AI name changes

Note: Formerly Azure AI Foundry (and before that Azure AI Studio).

Microsoft's Azure platform for building AI apps and agents, with a model catalog that includes OpenAI, Anthropic, Microsoft, DeepSeek and other models. It adds tracing, evaluations and governance controls, and connects to Azure services and Microsoft Entra ID.

Free to start: You can browse Foundry without an Azure account, but building needs an Azure subscription. Microsoft offers a free Azure trial for up to 30 days.

Based in: 🇺🇸 United States, Redmond, Washington (Microsoft Corporation). Source

Terms: Guardrails

Checked Oct 7, 2026 on the company’s own site: Microsoft Foundry · What is Microsoft Foundry? · Foundry model catalog

OpenAI's API lets your code use the GPT models behind ChatGPT, currently led by GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. It also offers the Agents SDK and Responses API for building agents, with built-in tools such as web search, file search and remote MCP servers.

Free to start: OpenAI's API page lists pay-as-you-go pricing and does not mention a free tier.

Based in: 🇺🇸 United States, San Francisco, California (OpenAI OpCo, LLC). Source

Terms: Context window · Zero data retention

Checked Oct 7, 2026 on the company’s own site: OpenAI API Platform

Anthropic's API gives your code access to its Claude models: currently Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 4.5. It also offers Managed Agents, web search and a sandboxed code-execution tool, each billed by use.

Free to start: Anthropic says new users get a small amount of free credits to test the API.

Based in: 🇺🇸 United States, San Francisco, California (Anthropic PBC). Source

Terms: Context window

Checked Oct 7, 2026 on the company’s own site: Claude pricing · Claude API pricing FAQ

Google's developer API for its Gemini models, with Google AI Studio as a free web workspace for trying prompts. It covers text, images, audio and video, plus embedding and text-to-speech models and tools like Google Search grounding and code execution.

Free to start: Free tier with limited models and lower rate limits. On the free tier Google uses your content to improve its products; on the paid tier it does not.

Based in: 🇺🇸 United States, Mountain View, California (Google LLC). Google LLC is organized under Delaware law; its address is in Mountain View, California. Source

Terms: Embeddings · Context window

Checked Oct 7, 2026 on the company’s own site: Gemini API pricing

Mistral sells API access to its models through Mistral Studio, priced per million tokens, with 50% off for batch jobs. Some of its models are published as open weights that you can download and run on your own servers.

Free to start: Mistral’s Free plan gives limited access to Vibe and includes access to Mistral Studio. Open source: Mistral says its open-weight models (for example Mistral 7B) are Apache 2.0 licensed for research and individual use, and commercial deployments need a Mistral license.

Based in: 🇫🇷 France, Paris (Mistral AI). Source

Terms: Open-weight model

Checked Oct 7, 2026 on the company’s own site: Mistral pricing and FAQ · Mistral Studio

Note: Not the same as Groq. Grok, spelled with a k, is xAI’s family of models (xAI’s site now brands itself SpaceXAI). Groq, with a q, is a separate inference company, listed under Model hosting & inference.

xAI's API for its Grok models, led by Grok 4.7 for code, chat and reasoning, plus Grok Imagine for images and video and a Voice API for speech. It works with the OpenAI SDK by changing the base URL, and offers server-side web search and X search tools.

Free to start: No free API tier or free credits are listed. Every Console account includes a free Playground for testing models; API use is paid from prepaid credits or monthly invoicing. Open source: xAI published the weights of an older model, Grok-1, in March 2024 under Apache 2.0. xAI's site does not list the current API models as open weights.

Based in: 🇺🇸 United States, Austin, Texas (SpaceXAI LLC). SpaceXAI LLC is a Nevada company with its registered offices in Austin, Texas. Source

Terms: Context window · Open-weight model · Zero data retention

Checked Oct 7, 2026 on the company’s own site: xAI API · Models and pricing · Billing · Open release of Grok-1

Note: Meta’s former Llama API address (llama.developer.meta.com) and llama.com now forward to Meta Model API. Its docs list Meta’s Muse models; Llama models are not listed.

Meta’s pay-as-you-go API for its Muse Spark models, plus Muse Image for images, Muse Voice Transcribe for speech-to-text and Segment Anything (SAM 3.1) for finding objects in images and video. It works with the OpenAI and Anthropic SDKs by changing the base URL, and Meta also offers Muse Code, a coding agent for the terminal.

Free to start: Billing is pay-as-you-go. The pricing page mentions “platform free-tier credits” but doesn’t say how much. A Contributor tier costs far less ($0.10 in / $0.20 out per million tokens) in exchange for letting Meta use your prompts and outputs to train future models. Open source: Meta publishes Muse Glimmer, a 30B open-weight model, under Apache 2.0 for running on your own hardware. The hosted Muse Spark models are not open.

Based in: 🇺🇸 United States, Menlo Park, California (Meta Platforms, Inc.). Source

Terms: Context window · Open-weight model

Checked Oct 7, 2026 on the company’s own site: Docs overview · Pricing and rate limits · Open-weight models

Cohere sells business-focused models through its API: Command for generating text, Embed for turning text into embeddings, Rerank for sorting search results, plus Transcribe, Translate and Parse. It also offers dedicated hosting (Model Vault) and private deployments on your own servers or private cloud.

Free to start: Every account gets a free Trial API key. Trial keys are rate-limited (1,000 calls a month) and may not be used for production or commercial work; production keys are pay-as-you-go.

Based in: 🇨🇦 Canada, Toronto, Ontario (Cohere). City from the mailing address in Cohere’s privacy policy. Source

Terms: Embeddings · RAG · On-premises

Checked Oct 7, 2026 on the company’s own site: Cohere pricing · API keys and rate limits · Deployment options

DeepSeek’s API serves its own models, currently DeepSeek-V4.1-Flash, with a 1-million-token context window. It accepts both OpenAI-style and Anthropic-style requests, and prices are half during off-peak hours.

Free to start: No free tier is listed. You top up a balance and usage is deducted from it. Open source: DeepSeek links to the DeepSeek-V4.1-Flash weights on its Hugging Face page, where they are published under the MIT license.

Based in: 🇨🇳 China (Hangzhou DeepSeek Artificial Intelligence Co., Ltd.). Run by Hangzhou DeepSeek Artificial Intelligence Co., Ltd., with its registered address in China. Source

Terms: Context window · Open-weight model

Checked Oct 7, 2026 on the company’s own site: Models and pricing · V4.1-Flash release note · Weights on Hugging Face

Note: AWS says you access Nova models through the Amazon Bedrock API, so Nova is listed as part of Bedrock.

Amazon’s own family of models, including Nova 2 Lite for everyday reasoning tasks and Nova 2.5 Sonic for real-time speech-to-speech voice agents. AWS also offers Nova Act for building agents that operate web browsers, and Nova Forge for building your own model on top of Nova.

Free to start: No separate free tier. New AWS customers get up to $200 in AWS credits to try AWS AI services.

Based in: 🇺🇸 United States, Seattle, Washington (Amazon Web Services, Inc.). Amazon’s model family, offered by AWS through Amazon Bedrock. Source

Terms: Foundation model · Reasoning model

Checked Oct 7, 2026 on the company’s own site: Amazon Nova · Nova models · Nova pricing · Nova user guide

Z.ai sells API access to its GLM models, led by GLM-5.3 for coding and long software tasks, along with vision, image, video and speech models. Several GLM models are also published as open weights.

Free to start: Some smaller models, such as GLM-4.7-Flash and GLM-4.5-Flash, are listed as free; the others are priced per million tokens. Open source: GLM-5.3 weights are on Z.ai’s Hugging Face page under its own GLM-5.3 license: free to use and modify, with extra conditions for very large companies that offer the model as a paid API service. GLM-5.3-Flash is MIT licensed.

Based in: 🇨🇳 China, Beijing (Z.ai (Zhipu)). Z.ai (Zhipu) describes itself as an LLM provider in China and gives its address in Beijing. Operating company for the Z.ai API platform: JINGSHENG HENGXING TECHNOLOGY PTE.LTD, Singapore. Source

Terms: Open-weight model · Context window

Checked Oct 7, 2026 on the company’s own site: Z.ai pricing · GLM-5.3 on Hugging Face · GLM-5.3 license

Moonshot AI’s developer platform for its Kimi models, led by Kimi K3, a flagship model with a 1-million-token context window aimed at software engineering and deep reasoning. It also offers K2.7 Code for coding and built-in tools such as web search and code execution.

Free to start: The platform lists per-token prices (Kimi K3: $3 in / $15 out per million tokens) and doesn’t mention a free tier. Open source: Moonshot publishes the Kimi K3 weights on Hugging Face under its own Kimi K3 license, based on MIT, with extra conditions for very large companies that offer the model as a paid API service.

Based in: 🇨🇳 China, Beijing (Moonshot AI). Moonshot AI’s About page gives its address in Beijing; its Kimi privacy policy names Beijing Moonshot Technology Co., Ltd. and says data is stored in the People’s Republic of China. Operating company for the Kimi API platform: Moonshot AI PTE. LTD., Singapore. Source

Terms: Context window · Open-weight model

Checked Oct 7, 2026 on the company’s own site: Kimi API platform · Kimi K3 on Hugging Face · Kimi K3 license

Qwen is a family of models offered through an API that uses the same request format as the OpenAI API. Many Qwen models are also released as open weights, such as the small Qwen3.5-2B and the image model Qwen-Image-2.1.

Free to start: The Qwen site says its chat app is free; it doesn’t state a free tier for the API. Open source: Many Qwen models are published on Hugging Face; Qwen3.5-2B, for example, is Apache 2.0. Licenses vary by model.

Based in: 🇨🇳 China, Hangzhou (Alibaba Cloud (Qwen team)). Qwen models are developed by the Qwen team at Alibaba Cloud. Qwen’s GitHub organization lists China; city from Alibaba’s GitHub organization (Hangzhou, China). Operating company for qwen.ai: Nth Power Global Tech Singapore Pte. Ltd., Singapore. Source

Terms: Open-weight model

Checked Oct 7, 2026 on the company’s own site: Qwen · Qwen3.5-2B on Hugging Face

Note: OpenAI offers the Decisions API as an endpoint of the OpenAI API (POST /v1/decisions), so it is listed as part of the OpenAI API.

An OpenAI API endpoint for quick, structured answers instead of written replies: the probability that a statement is true, a pick from a fixed list of options, or a score against a rubric, for text and images. OpenAI says it is about 10 times faster than asking the same question through the Responses API. It is in public beta, and gpt-6-luna is the only model.

Free to start: Paid by use: $0.10 per 1 million input tokens with gpt-6-luna, and no charge for output tokens.

Based in: 🇺🇸 United States, San Francisco, California (OpenAI OpCo, LLC). Source

Terms: Inference · Latency

Checked Oct 7, 2026 on the company’s own site: Decisions guide · OpenAI API pricing · Public beta announcement (OpenAI Developer Community)

The site where AI companies and researchers publish open models and datasets; most of the open models in this directory are downloaded from it. Through Inference Providers you can call 200+ models from many hosting companies with one account and no Hugging Face markup, or rent dedicated servers for a model with Inference Endpoints.

Free to start: Free accounts can run Spaces on free CPU hardware. Inference Providers needs purchased credits on a free account; PRO ($9 a month) includes $2 of monthly credits. Open source: Its Transformers library and many other tools are open source (Apache 2.0); the Hub service itself is not.

Based in: 🇺🇸 United States (Hugging Face). Hugging Face says the company is located in the United States; its main EU establishment is Hugging Face SAS in Paris, France. Source

Terms: Open-weight model · Inference · Fine-tuning

Checked Oct 7, 2026 on the company’s own site: Pricing · Inference Providers pricing · Transformers LICENSE

A cloud for running open models: call them through an OpenAI-compatible API and pay per token, rent dedicated GPUs for one model, fine-tune models on your own data, or book H100 and B200 GPU clusters.

Free to start: Together says it does not offer free trials; you need to buy at least $15 of credits to start.

Based in: 🇺🇸 United States, San Francisco, California (Together Computer, Inc.). Together Computer, Inc. is a Delaware corporation; city from the notice address in its terms. Source

Terms: Open-weight model · Fine-tuning · GPU

Checked Oct 7, 2026 on the company’s own site: Docs overview · Billing and credits · Pricing

Note: Not the same as Grok. Groq, spelled with a q, runs fast inference on its own chips. Grok, with a k, is xAI’s (SpaceXAI’s) family of models, listed under Model APIs.

Groq designs its own inference chips, called LPUs, and runs open models on them in GroqCloud, an API that works with the OpenAI SDK. Its newer LPX systems work alongside NVIDIA GPUs.

Free to start: You start on a Free tier. Adding a payment method moves you to the pay-as-you-go Developer tier, which Groq says adds more capacity and features.

Based in: 🇺🇸 United States, Mountain View, California (Groq LLC). City from Groq’s mailing address (a P.O. box). Source

Terms: Inference · Latency · Grok

Checked Oct 7, 2026 on the company’s own site: Groq · GroqCloud docs · Rate limits · Billing FAQs

Note: Replicate has joined Cloudflare. It says it will carry on as a distinct brand and its API isn’t changing.

A site for running models through an API, from thousands of community-contributed open models to commercial ones. Most models are billed for the time they run, and you can package and deploy your own model with Cog, Replicate’s open-source tool.

Free to start: You can run select models for free; after a while you’re asked to set up billing. Open source: Cog, its tool for packaging models, is open source (Apache 2.0); the Replicate service is not.

Based in: 🇺🇸 United States, San Francisco, California (Replicate, LLC). Replicate joined Cloudflare (Replicate blog); both list 101 Townsend St., San Francisco. Source

Terms: Open-weight model · Inference

Checked Oct 7, 2026 on the company’s own site: Pricing · Billing docs · Replicate is joining Cloudflare · Cog LICENSE

Runs AI models on GPUs in Cloudflare’s global network, called from your code without managing servers. It offers 50+ open-source models and fits with Cloudflare’s other developer tools, such as AI Gateway and its vector database.

Free to start: Everyone gets 10,000 Neurons a day free (Neurons are Cloudflare’s usage unit); beyond that you need Workers Paid at $0.011 per 1,000 Neurons.

Based in: 🇺🇸 United States, San Francisco, California (Cloudflare, Inc.). Source

Terms: Inference · Edge computing · GPU

Checked Oct 7, 2026 on the company’s own site: Workers AI docs · Pricing

One API and one bill for 500+ models from 80+ providers. If a provider goes down, requests can fall back to another. OpenRouter passes through each provider’s price without markup and charges a fee when you buy credits.

Free to start: Many models have a free version with low rate limits, and new users get a small free allowance. Paid use draws on credits, with a 5.5% fee on card purchases.

Based in: 🇺🇸 United States, New York, New York (OpenRouter, Inc.). City from the notice address in OpenRouter’s terms. Source

Terms: Inference

Checked Oct 7, 2026 on the company’s own site: OpenRouter · FAQ

One API key for hundreds of text, image, video and audio models. Vercel charges each provider’s list price with no markup, and on paid use you can bring your own provider keys.

Free to start: Every Vercel team gets $5 a month in AI Gateway credits for a subset of models; beyond that you buy credits.

Based in: 🇺🇸 United States, Covina, California (Vercel Inc.). City from the mailing address in Vercel’s privacy notice. Source

Terms: Inference

Checked Oct 7, 2026 on the company’s own site: AI Gateway · AI Gateway pricing

Deploys and serves AI models across Vultr’s data centers on six continents without you managing the servers. It is part of Vultr’s cloud, which also rents GPUs.

Free to start: The product page offers a free account but doesn’t list a free usage tier.

Based in: 🇺🇸 United States, West Palm Beach, Florida (The Constant Company, LLC (Vultr)). Vultr’s contact page gives its location in West Palm Beach, Florida; its terms name The Constant Company, LLC. Source

Terms: Inference · GPU

Checked Oct 7, 2026 on the company’s own site: Vultr Serverless Inference

An open-source AI gateway and Python SDK that you run yourself: one OpenAI-compatible API in front of many model providers, with keys, spending limits, routing and logging for a whole team. A paid Enterprise version adds more admin features.

Free to start: The open-source gateway is free to self-host with no credit card; LiteLLM Enterprise is paid. Open source: Code outside the enterprise/ folder is MIT; the enterprise folder has its own license.

Based in: 🇺🇸 United States, San Francisco, California (Berrie AI Incorporated). Source

Terms: Inference · Latency

Checked Oct 7, 2026 on the company’s own site: LiteLLM · About · Docs · License

Google’s family of open models, sized from small E2B and E4B versions for phones and other devices up to 12B, 26B and 31B versions for personal computers and servers. Gemma 4 reads text, audio and images and has a 256K-token context window. Related models include DiffusionGemma, TranslateGemma and MedGemma.

Free to start: The weights are free to download. Open source: Gemma 4 weights are published on Google’s Hugging Face page under Apache 2.0.

Based in: 🇺🇸 United States, Mountain View, California (Google LLC). Google LLC is organized under Delaware law; its address is in Mountain View, California. Source

Terms: Open-weight model · Context window · Edge computing

Checked Oct 7, 2026 on the company’s own site: Gemma · Gemma docs · Gemma 4 31B on Hugging Face

A small 740M-parameter open model that turns text, images, audio and video into embeddings, so you can search across all of them together. It is built on Gemma 4 and is designed to run locally and offline on consumer GPUs or CPUs.

Free to start: The weights are free to download. Open source: Google publishes the weights under the Apache 2.0 license.

Based in: 🇺🇸 United States, Mountain View, California (Google LLC). Google LLC is organized under Delaware law; its address is in Mountain View, California. Source

Terms: Embeddings · RAG · Open-weight model

Checked Oct 7, 2026 on the company’s own site: EmbeddingGemma docs · Model card on Hugging Face

The Falcon family from TII in Abu Dhabi includes the Falcon-H1 language models, Falcon-H1-Arabic, Falcon Perception for images, and three new models for the UAE: Falcon-Emirati, Falcon-ASR for speech recognition and Falcon-OCR-Arabic for reading Arabic text in images.

Free to start: Open models are free to download; TII also offers web demos of the new models. Open source: TII says it releases its Falcon models as open source or open access. Licenses vary by model; Falcon-OCR, for example, is Apache 2.0 on Hugging Face.

Based in: 🇦🇪 United Arab Emirates, Abu Dhabi (Technology Innovation Institute). TII is part of the Abu Dhabi Government’s Advanced Technology Research Council (TII About page). Source

Terms: Open-weight model

Checked Oct 7, 2026 on the company’s own site: Falcon LLM · Falcon-OCR on Hugging Face

An open-weight model from German company Aleph Alpha for German and English. It is a mixture-of-experts model with 78B total parameters, of which about 3.5B are active for each token, and it can take in up to about 1 million tokens.

Free to start: The weights are free to download. Open source: Published under Apache 2.0 on Aleph Alpha’s Hugging Face page.

Based in: 🇩🇪 Germany, Heidelberg (Aleph Alpha GmbH). Source

Terms: Open-weight model · Parameters · Context window

Checked Oct 7, 2026 on the company’s own site: Kolibri-1 model card · Aleph Alpha

Note: Announced Oct 5, 2026; not downloadable yet. Reflection says it will release the weights later this month and offers early-access sign-up.

Reflection AI’s first open-weight model: a mixture-of-experts model with 501B total parameters, 23B of them active at a time, built for coding, reasoning and agent work. Reflection says it is going through final safety testing.

Free to start: Not available yet; only an early-access sign-up. Open source: Reflection describes Beam as open-weight, with weights due later this month. The license is not stated yet.

Based in: 🇺🇸 United States, Brooklyn, New York (Reflection AI, Inc.). Source

Terms: Open-weight model · Parameters

Checked Oct 7, 2026 on the company’s own site: Introducing Beam · Reflection AI

Clef is a 27B model from Cloudflare that is tuned to make decisions, such as approving, classifying or routing items in business workflows, rather than to chat. A smaller Clef-Flash version is also available. Both are post-trained from Qwen.

Free to start: The weights are free to download. Open source: Published under Apache 2.0 on Cloudflare’s Hugging Face page.

Based in: 🇺🇸 United States, San Francisco, California (Cloudflare, Inc.). Source

Terms: Open-weight model · Fine-tuning

Checked Oct 7, 2026 on the company’s own site: Clef model card · Clef-Flash model card

A small 2B-parameter model that makes quick decisions inside agents, such as which tool to call next, in tens of milliseconds on a local CPU or GPU. The code, weights, training data and scripts are all published.

Free to start: Free to download. Open source: Code is on GitHub and the weights are on Hugging Face under Apache 2.0, along with the training data and scripts.

Based in: 🇺🇸 United States, Seattle, Washington (Amazon Web Services, Inc.). From the Strands Agents project; the strandsagents.com site is © Amazon Web Services, Inc. Source

Terms: Open-source AI · Parameters · Latency

Checked Oct 7, 2026 on the company’s own site: Introducing Strands Decider 2B · Weights on Hugging Face

NVIDIA’s family of open models for AI agents, including the Nemotron 3.5 Lightning models and safety and speech models. NVIDIA publishes the training data as well, and the models run on anything from edge devices to the cloud.

Free to start: The models are openly available to download. Open source: NVIDIA calls the models openly available and publishes training data. Licenses vary by model; many use the NVIDIA Open Model License, which NVIDIA describes as permissive and allowing commercial use.

Based in: 🇺🇸 United States, Santa Clara, California (NVIDIA Corporation). Source

Terms: Open-weight model · GPU

Checked Oct 7, 2026 on the company’s own site: NVIDIA Nemotron

NVIDIA’s pretrained model for classification and regression on tables of data, such as spreadsheets and database records, rather than text.

Free to start: Free to download. Open source: Published on NVIDIA’s Hugging Face page under the OpenMDW 1.1 license.

Based in: 🇺🇸 United States, Santa Clara, California (NVIDIA Corporation). Source

Terms: Open-weight model · Machine learning

Checked Oct 7, 2026 on the company’s own site: Kumo-Tabular model card

Poolside’s open-weight models for coding agents. Laguna XS 2.1 (33B parameters, 3B active, 256K context) is small enough to run on your own device; Laguna S 2.1 (118B, 8B active, 1M context) is its newest. Both are also available on OpenRouter and Vercel AI Gateway.

Free to start: The weights are free to download; hosted use is billed by the provider. Open source: Laguna XS 2.1 and S 2.1 are published under the OpenMDW 1.1 license; the earlier Laguna XS.2 is Apache 2.0.

Based in: 🇺🇸 United States, San Francisco, California (Poolside, Inc.). Poolside, Inc. is a Delaware corporation with offices in San Francisco. Source

Terms: Open-weight model · Parameters · Context window

Checked Oct 7, 2026 on the company’s own site: Poolside · Laguna XS 2.1 on Hugging Face · Laguna S 2.1 on Hugging Face

KT’s Korea-centric open models, in a Base and a smaller Mini version, built to understand Korean language and culture.

Free to start: Free to download. Open source: Published on KT’s Hugging Face page under the MIT license.

Based in: 🇰🇷 South Korea, Seongnam, Gyeonggi-do (KT Corp.). Source

Terms: Open-weight model

Checked Oct 7, 2026 on the company’s own site: Mi:dm 2.0 Base model card

Tencent’s Hy open models (Tencent’s Hunyuan website, hunyuan.tencent.com, now uses the Hy name). Hy3 is a mixture-of-experts language model with 295B total parameters, 21B of them active; the family also includes Hy-MT2 translation models and a Hy4 preview.

Free to start: Free to download. Open source: Hy3 is published on Tencent’s Hugging Face page under Apache 2.0; check each model’s card, as licenses can differ.

Based in: 🇨🇳 China, Shenzhen (Tencent). Source

Terms: Open-weight model · Parameters

Checked Oct 7, 2026 on the company’s own site: Hy3 model card · Tencent Hy

An open-source framework for building apps and agents on top of language models, with ready-made connections to many model providers, databases and tools. The same company makes LangGraph for lower-level agent control, Deep Agents for long-running agents, and LangSmith, a hosted platform for testing, monitoring and deploying agents.

Free to start: The library is free under the MIT license. Open source: Yes, MIT license.

Based in: 🇺🇸 United States (LangChain Inc.). LangChain Inc. is a Delaware corporation; no headquarters city is stated. Source

Terms: RAG · Embeddings

Checked Oct 7, 2026 on the company’s own site: LangChain · LICENSE (MIT)

An open-source Python framework for connecting language models to your own data, commonly used for retrieval-augmented generation (RAG) over documents. The company also runs LlamaParse, a hosted service that turns PDFs, tables, charts and scans into clean text that AI apps can use.

Free to start: The framework is free under the MIT license. LlamaParse has a free plan with 10,000 credits a month (about 1,000 pages). Open source: Yes, MIT license.

Based in: 🇺🇸 United States, San Francisco, California (LlamaIndex, Inc.). LlamaIndex’s GitHub organization lists the United States; city from the office named on its careers page. Its terms and privacy notice give no address. Source

Terms: RAG · Embeddings

Checked Oct 7, 2026 on the company’s own site: LlamaIndex · LICENSE (MIT)

An open-source Python framework for multi-agent workflows, where several AI agents each get a role and work through a task together. The company also sells an enterprise platform for building, running and governing agents.

Free to start: The framework is free under the MIT license. Open source: Yes, MIT license.

Based in: 🇺🇸 United States (CrewAI, Inc.). CrewAI’s GitHub organization lists the United States; its terms say CrewAI, Inc. is a Delaware corporation. No city is stated. Source

Terms: Prompt injection

Checked Oct 7, 2026 on the company’s own site: CrewAI · CrewAI on GitHub

Microsoft’s lightweight, open-source kit for building AI agents and adding AI models to apps written in C#, Python or Java. It sits between your code and the models, handling prompts, plugins (functions the model can call) and memory.

Free to start: Free to use; you pay only for the models you call. Open source: MIT license.

Based in: 🇺🇸 United States, Redmond, Washington (Microsoft Corporation). Source

Terms: Open-source AI

Checked Oct 7, 2026 on the company’s own site: Introduction to Semantic Kernel · LICENSE

An open-source toolkit for building production AI agents in Python or TypeScript. It includes a ready-made agent harness, an SDK for building your own, a safe virtual shell for agents, and evaluation tools, and it works with every major model provider.

Free to start: Free to use; you pay only for the models you call. Open source: Apache 2.0 (the SDK repository is now strands-agents/harness-sdk).

Based in: 🇺🇸 United States, Seattle, Washington (Amazon Web Services, Inc.). Open-source project from AWS; the strandsagents.com site is © Amazon Web Services, Inc. Source

Terms: Open-source AI

Checked Oct 7, 2026 on the company’s own site: Strands Agents · Harness SDK on GitHub

An open-source MCP server from Google that connects AI agents, coding tools and apps directly to your company databases, handling connections and security so you don’t have to.

Free to start: Free to use. Open source: Apache 2.0.

Based in: 🇺🇸 United States, Mountain View, California (Google LLC). Google LLC is organized under Delaware law; its address is in Mountain View, California. Source

Terms: Open-source AI

Checked Oct 7, 2026 on the company’s own site: README on GitHub

Open-source device SDKs and firmware for connecting Meta’s Muse assistant to hardware you build yourself, using an off-the-shelf ESP32 board or a Raspberry Pi with displays, buttons and sensors.

Free to start: The SDKs are free; you need an SDK token from a Muse account. Open source: The device SDKs and firmware are Apache 2.0.

Based in: 🇺🇸 United States, Menlo Park, California (Meta Platforms, Inc.). Muse is Meta’s assistant; gadgets.muse.ai links to Meta’s AI terms. Source

Terms: Open-source AI · Edge computing

Checked Oct 7, 2026 on the company’s own site: Muse Gadgets

Atlassian’s official MCP server, hosted by Atlassian. It lets AI tools such as ChatGPT, Claude, GitHub Copilot CLI or Gemini search and update Jira, Confluence, Jira Service Management, Bitbucket, Loom and other Atlassian Cloud data, within each user’s existing permissions. Atlassian says version 2 went from dozens of tools to more than 200.

Free to start: Each call uses Rovo credits from your organization’s shared pool; allowances depend on your Atlassian plan.

Based in: 🇦🇺 Australia, Sydney (Atlassian). Source

Terms: Prompt injection

Checked Oct 7, 2026 on the company’s own site: Rovo MCP · Getting started · GitHub repository · Team ’26 Europe announcement

An open-source standard for connecting AI apps to outside data and tools, such as files, databases and business software. Claude, ChatGPT, VS Code, Cursor and many other apps support it, so one MCP server can work with all of them.

Free to start: Free to implement. Open source: The specification is Apache 2.0; the project says it is moving its code from MIT to Apache 2.0.

Based in: 🇺🇸 United States (Created by Anthropic; now part of the Agentic AI Foundation (Linux Foundation)). MCP was created by Anthropic, a US company, which donated it to the Agentic AI Foundation under the Linux Foundation. It is set up as a Series of LF Projects, LLC, with contributors from many companies. Source

Terms: Prompt injection · Open-source AI

Checked Oct 7, 2026 on the company’s own site: What is MCP? · LICENSE

Note: Announced Oct 6, 2026. The v0.1 specification and a reference implementation are planned for later in October.

A proposed standard for how people’s personal AI agents deal with companies: handling sign-in, protecting consumers and showing companies what agents do on their websites, APIs or own agents. Meta and Sierra introduced it with several partner companies, and it is open for anyone to implement.

Free to start: Open for anyone to implement.

Based in: 🇺🇸 United States, Menlo Park and San Francisco, California (Meta Platforms, Inc. and Sierra Technologies, Inc.). Introduced by two US companies: Meta (Menlo Park) and Sierra (San Francisco). Source

Checked Oct 7, 2026 on the company’s own site: Introducing Personal Agent Protocol

A fully managed vector database: you store embeddings and Pinecone returns the closest matches quickly, which is the search step in most RAG apps. It also offers hosted embedding and reranking models and Pinecone Assistant for chatting with documents.

Free to start: The Starter plan is free: up to 2 GB of storage, on AWS in one U.S. region.

Based in: 🇺🇸 United States, San Francisco, California (Pinecone Systems, Inc). City from the mailing address in Pinecone’s privacy policy. Source

Terms: Embeddings · RAG

Checked Oct 7, 2026 on the company’s own site: Pinecone pricing

An open-source vector database that can combine meaning-based search with ordinary keyword search (hybrid search). You can run it on your own servers or use Weaviate Cloud, the company's managed service.

Free to start: Weaviate Cloud has an always-free plan: one cluster, up to 100,000 objects. Open source: The core database is BSD-3-Clause; some enterprise features sit under a separate paid license.

Based in: 🇳🇱 Netherlands, Amsterdam (Weaviate B.V.). Source

Terms: Embeddings · RAG

Checked Oct 7, 2026 on the company’s own site: Weaviate pricing · LICENSE

An open-source database for AI search that handles vector, full-text, regex and metadata search. You can run it locally in a few lines of code or use Chroma Cloud, its serverless hosted version.

Free to start: The Chroma Cloud Starter plan is $0 a month plus usage, and comes with $5 in free credits. Running it yourself is free. Open source: Apache 2.0.

Based in: 🇺🇸 United States, San Francisco, California (Chroma Inc.). Source

Terms: Embeddings · RAG

Checked Oct 7, 2026 on the company’s own site: Chroma · Pricing · LICENSE

An open-source vector search engine. You can run it yourself, use the fully managed Qdrant Cloud, or run Qdrant Hybrid Cloud on your own infrastructure while Qdrant manages it.

Free to start: Qdrant Cloud has a Free tier, “free forever”, for testing and prototypes: a single node with 0.5 vCPU, 1 GB RAM and 4 GB disk. Open source: Apache 2.0.

Based in: 🇩🇪 Germany, Berlin (Qdrant Solutions GmbH). Source

Terms: Embeddings · RAG

Checked Oct 7, 2026 on the company’s own site: Pricing · LICENSE

An open-source vector database built to scale to tens of billions of vectors. Milvus Lite installs with pip for quick starts; Zilliz, the company behind it, sells Zilliz Cloud as a fully managed version, including a bring-your-own-cloud option.

Free to start: Zilliz Cloud’s Free plan includes 5 GB of storage and up to 5 collections. Running Milvus yourself is free. Open source: Apache 2.0.

Based in: 🇺🇸 United States, Redwood City, California (Zilliz Inc.). Milvus is a graduated project of the LF AI & Data Foundation, contributed by Zilliz. Source

Terms: Embeddings · RAG

Checked Oct 7, 2026 on the company’s own site: Milvus · Zilliz Cloud pricing · LICENSE

Supabase is a hosted Postgres database with sign-in, file storage and APIs built in. Turning on the pgvector extension lets you store embeddings and run vector similarity search in the same database as the rest of your app’s data.

Free to start: The Free plan includes a 500 MB database and up to 2 active projects; free projects are paused after a week of inactivity. Open source: Supabase is Apache 2.0; pgvector uses the PostgreSQL license.

Based in: 🇺🇸 United States (Supabase). Supabase’s GitHub organization lists the United States; no city is stated. Contracting company in its terms: Supabase Pte. Ltd., Singapore (Supabase, Inc., registered in Delaware, for cloud-marketplace purchases). Source

Terms: Embeddings · RAG

Checked Oct 7, 2026 on the company’s own site: Pricing · pgvector guide · Supabase LICENSE · pgvector LICENSE

A code editor and command-line tool built around AI agents: you hand off a task, and the agent plans it, writes the code and leaves the result for you to review. You can pick models from OpenAI, Anthropic, Google, xAI and Cursor itself, and paid plans add cloud agents and Bugbot code review.

Free to start: The Hobby plan is free with no credit card and a limited number of agent requests.

Based in: 🇺🇸 United States, San Francisco, California (Anysphere, Inc.). City from the contact address in Cursor’s terms. Source

Terms: Context window

Checked Oct 7, 2026 on the company’s own site: Cursor · Cursor pricing · Cursor download

GitHub's AI coding assistant, which works in VS Code, Visual Studio, JetBrains IDEs, Xcode and other editors, on github.com and in the terminal. It suggests code as you type, answers questions in chat, and has an agent you can assign an issue to that opens a pull request.

Free to start: Copilot Free includes 2,000 inline suggestions a month and limited chat and agent use. GitHub may use Free, Pro and Pro+ interactions to train its models unless you opt out.

Based in: 🇺🇸 United States, San Francisco, California (GitHub, Inc.). GitHub’s terms state that Microsoft is an affiliate of GitHub. Source

Terms: Prompt injection

Checked Oct 7, 2026 on the company’s own site: Copilot plans and FAQ

Note: Windsurf has been renamed Devin Desktop. windsurf.com now forwards to devin.ai; Cognition says plans and pricing stay the same, and Windsurf for JetBrains is still available.

An AI code editor built on the Windsurf editor, now with an Agent Command Center for running several coding agents at once, both on your computer and in the cloud (Devin Cloud). It includes Tab completions, inline edits and the Cascade agent.

Free to start: The Free plan includes a light agent quota with limited models, plus unlimited Tab completions and inline edits. Pro is $20 a month, Max $200 a month, and Teams $80 a month plus $40 per full seat.

Based in: 🇺🇸 United States, San Francisco, California (Cognition AI, Inc.). Source

Terms: Context window

Checked Oct 7, 2026 on the company’s own site: Devin Desktop · Plans and pricing

Anthropic’s coding agent: it reads your codebase, edits files, runs commands and works with your development tools. You can use it in the terminal, in VS Code or JetBrains IDEs, in a desktop app, in the browser, or from Slack.

Free to start: No free plan is listed. It is included in Claude Pro ($17 a month billed annually, or $20 monthly) and Max (from $100 a month), Team and Enterprise plans, or you can pay per use with a Claude Console account.

Based in: 🇺🇸 United States, San Francisco, California (Anthropic PBC). Source

Terms: Context window

Checked Oct 7, 2026 on the company’s own site: Claude Code · Claude Code docs · LICENSE

OpenAI’s coding agent. It works in ChatGPT, in the cloud, as an IDE extension, as a desktop app and as the Codex CLI in your terminal.

Free to start: Codex is included in ChatGPT plans: Plus ($20 a month), Pro ($100 a month), Business and Enterprise. It can also be used with an API key. Open source: The Codex CLI is open source under Apache 2.0; the cloud service is not.

Based in: 🇺🇸 United States, San Francisco, California (OpenAI OpCo, LLC). Source

Terms: Context window

Checked Oct 7, 2026 on the company’s own site: Codex · Codex docs · Codex CLI LICENSE

IBM’s AI development partner for your IDE and command line (Bob Shell), which can run agents and subagents in your codebase. Premium packages focus on enterprise modernization, such as upgrading Java apps and mainframe and IBM i development, and Bobalytics reports on usage and cost.

Free to start: A free trial is offered; the price after the trial isn’t shown on the product page.

Based in: 🇺🇸 United States, Armonk, New York (IBM). Source

Checked Oct 7, 2026 on the company’s own site: IBM Bob

xAI’s coding agent and terminal app, now open source. It can run interactively, headless in scripts and CI, or inside editors through the Agent Client Protocol (ACP), and it can run fully local, pointed at your own model server.

Free to start: The code is free; using xAI’s models costs extra. Open source: First-party code is Apache 2.0.

Based in: 🇺🇸 United States, Austin, Texas (SpaceXAI LLC). SpaceXAI LLC is a Nevada company with its registered offices in Austin, Texas. Source

Terms: Grok · Open-source AI

Checked Oct 7, 2026 on the company’s own site: Grok Build is now open source · README on GitHub

Note: MAI-Transcribe-2-Streaming is in public preview in Azure (no service-level agreement; not recommended for production).

Microsoft AI’s own speech models: MAI-Transcribe-2-Streaming turns speech into text in real time in 60 languages, and MAI-Voice-2.1 and its faster Flash version turn text into speech in 23 languages.

Free to start: Pay per use: Transcribe is $0.54 an hour of audio (introductory price through the end of the year); Voice-2.1 is $22 and Flash $15 per million characters.

Based in: 🇺🇸 United States, Redmond, Washington (Microsoft Corporation). Microsoft AI is part of Microsoft; microsoft.ai uses Microsoft’s privacy statement and terms. Source

Terms: Latency

Checked Oct 7, 2026 on the company’s own site: MAI-Transcribe-2-Streaming announcement · MAI-Voice-2.1 · Microsoft Learn: MAI-Transcribe-2-Streaming

A fully managed Azure API for real-time voice agents: you send audio and get back spoken replies, optional avatar visuals and action triggers, without wiring speech recognition, a language model and text-to-speech together yourself.

Free to start: The overview page doesn’t list a free tier.

Based in: 🇺🇸 United States, Redmond, Washington (Microsoft Corporation). Source

Terms: Latency · Cloud computing

Checked Oct 7, 2026 on the company’s own site: What is the Voice Live API?

An Indian AI platform with APIs for speech-to-text in 12 Indic languages, text-to-speech in 11, translation across 23 languages, and turning PDFs and images into structured text. All APIs draw from one prepaid credit balance.

Free to start: You start with free credits (the amount isn’t stated), then pay as you go; speech-to-text is ₹30 an hour.

Based in: 🇮🇳 India, Bengaluru (Axonwise Private Limited (Sarvam AI)). Sarvam AI is the trading name of Axonwise Private Limited, incorporated in India. Source

Checked Oct 7, 2026 on the company’s own site: Sarvam AI · API pricing

An open-source framework for real-time voice, video and text agents in Python or Node.js, with plugins for most major AI providers. LiveKit Cloud hosts and scales the agents and can connect them to phone lines.

Free to start: The Build plan is $0 a month with no credit card, and includes inference credits, 1,000 free agent session minutes a month and one free phone number. Open source: The Agents framework is Apache 2.0.

Based in: 🇺🇸 United States, San Francisco, California (LiveKit Incorporated). Source

Terms: Latency · Open-source AI

Checked Oct 7, 2026 on the company’s own site: Agents framework docs · Pricing · LICENSE

Black Forest Labs makes the FLUX image and video models. The newest, FLUX 3 Image and FLUX 3 Video, are available through its pay-as-you-go API, and some FLUX models are released as open weights you can download and run yourself.

Free to start: Pay as you go with no subscription or seat fees. FLUX 3 Image lists at $0.048 per 1K image. Open source: Some FLUX models are published as open weights; licenses vary by model (for example, FLUX.2-small-decoder is Apache 2.0).

Based in: 🇩🇪 Germany, Freiburg im Breisgau (BFL GmbH). Registered address of BFL GmbH in Germany; Black Forest Labs Inc. in San Francisco handles users outside Europe (privacy policy). Source

Terms: Open-weight model

Checked Oct 7, 2026 on the company’s own site: Black Forest Labs · Pricing · FLUX.2-small-decoder on Hugging Face

Open video models that make 5-second clips with matching 44 kHz sound, including lip-sync, from a text prompt or a starting image. There is a Lite version (3B parameters) and a Pro version (29B), plus an add-on that upscales to Full HD.

Free to start: Free to download. Open source: MIT license on GitHub.

Based in: 🇷🇺 Russia, Moscow (Sber (PJSC Sberbank)). Sber’s developer site lists Kandinsky as one of Sber’s teams; Sber’s address is in Moscow, Russia. kandinskylab.ai itself names no company. Source

Terms: Open-weight model · GPU

Checked Oct 7, 2026 on the company’s own site: Kandinsky 6 on GitHub

Lightricks makes LTX, a family of open video and audio generation models; LTX-2.5 is the newest. You can download the weights and run or fine-tune them on your own hardware, or use the LTX API. Lightricks also offers LTX Desktop and a trainer for custom LoRA add-ons.

Free to start: The open weights are free for companies under $10 million in annual recurring revenue; the API is paid. Open source: Weights are on Lightricks’ Hugging Face page under its own LTX-2.x Community License, not a standard open-source license.

Based in: 🇮🇱 Israel, Jerusalem (Lightricks Ltd.). Jerusalem is the first office on Lightricks’ contact page and the mailing address in the LTX privacy policy. Source

Terms: Open-weight model · Fine-tuning

Checked Oct 7, 2026 on the company’s own site: LTX · LTX-2.5 on Hugging Face · LTX-2.x license

Note: Developers use Nano Banana 2.1 through the Gemini API and Google AI Studio, so it is listed as part of the Gemini API.

Google’s image generation and editing model, based on Gemini 3.6 Flash. In the Gemini API its model id is gemini-nano-banana-2.1; it can mix up to 14 reference images, render text, and make images at 1K, 2K or 4K. Google also uses it in the Gemini app, Flow and Stitch.

Free to start: Google’s pricing page lists no free tier for this model. Paid tier: about $0.034 per 1K image ($30 per 1 million image output tokens).

Based in: 🇺🇸 United States, Mountain View, California (Google LLC). Google LLC is organized under Delaware law; its address is in Mountain View, California. Source

Checked Oct 7, 2026 on the company’s own site: Model card · Image generation docs · Gemini API pricing

A free, open-source tool for downloading and running open models on your own computer, with a REST API on your machine that your apps can call. It also offers optional cloud models, paid with usage credits, for models too big to run at home.

Free to start: Running models on your own hardware is free and unlimited. Cloud models come with a small free starter allowance, then paid credits or plans. Open source: Yes, MIT license.

Based in: Not listed (Ollama Inc.). Ollama’s terms name Ollama Inc. and use California law, but its terms, privacy policy and GitHub organization give no headquarters, address or location.

Terms: Open-weight model · GPU

Checked Oct 7, 2026 on the company’s own site: Ollama pricing · Ollama download · LICENSE (MIT)

A desktop app for finding, downloading and running open models on your computer, with a chat window and a programmable API. Its newer Bionic agent can also edit documents and handle coding tasks, using local models or optional paid cloud models.

Free to start: The Free plan runs local models. Bionic+ ($20 a month) and Pro ($100) add hosted open models. Open source: The app is offered under LM Studio's terms of use; its lms command-line tool is open source (MIT).

Based in: 🇺🇸 United States (Element Labs, Inc.). Element Labs, Inc. is a Delaware corporation with its registered address in Wilmington, Delaware; no headquarters city is stated. Source

Terms: Open-weight model · GPU

Checked Oct 7, 2026 on the company’s own site: LM Studio pricing · LM Studio download · lms LICENSE (MIT)

An open-source engine for serving language models on your own GPUs or other hardware, with a drop-in OpenAI-compatible API. It is built for high throughput, with techniques such as PagedAttention and continuous batching.

Free to start: Free to use; you pay only for your own hardware. Open source: Apache 2.0.

Based in: 🇺🇸 United States (vLLM project). vLLM was first developed in the Sky Computing Lab at UC Berkeley; it is now a community project under the PyTorch Foundation with contributors from many companies. No company is named. Source

Terms: Inference · GPU · Open-weight model

Checked Oct 7, 2026 on the company’s own site: vLLM · Documentation · GitHub

Open-source software that stores and reuses the KV cache (the model’s saved work on a prompt) across model servers such as vLLM, to cut time to first token and raise throughput, especially for long, multi-turn or document-heavy requests.

Free to start: Free to use. Open source: Apache 2.0.

Based in: Not listed (LMCache project). LMCache calls itself an independent open-source project that joined the PyTorch Foundation and is supported in part by Tensormesh. No location is stated. Source

Terms: Inference · GPU

Checked Oct 7, 2026 on the company’s own site: GitHub · Documentation