Model providers
A model provider is a company or research lab that trains large language models and makes them available through a hosted API, downloadable open weights, or both. Providers differ in the modalities they support, how they license their weights, where their models run, and which cloud platforms resell them.
Some providers, such as Anthropic, offer models only through their own API and cloud partners. Others, including DeepSeek, Alibaba's Qwen team, Mistral AI, Moonshot AI, and Z.ai, also publish weights that teams can run on their own hardware, under licenses that range from MIT and Apache-2.0 to custom terms with commercial conditions. US labs such as OpenAI, Reflection AI, and Thinking Machines Lab now publish open weights as well.
This list covers labs that build their own language and multimodal foundation models; image-only and audio-only labs are out of scope. To reach several providers through one endpoint, teams usually add an AI gateway, and independent benchmarks are a better guide to model quality than vendor announcements.
Model providers compared
| Platform | Open source | Deployment | Best for |
|---|---|---|---|
| Ai2 (Olmo) | No | Self-hosted | Fully open models for research and auditing |
| AI21 Labs | No | API, Self-hosted | Custom post-trained models and hybrid-architecture open weights |
| Alibaba (Qwen) | No | API, Self-hosted | Open weights across a wide range of model sizes |
| Amazon Nova | No | API | Amazon's own models for teams building on AWS |
| Anthropic | No | API | Agentic and coding workloads across several clouds |
| Arcee AI | No | Self-hosted | US-built open weights for customization |
| Baidu (ERNIE) | No | API, Self-hosted | Chinese-language models with open and hosted options |
| ByteDance Seed | No | API, Self-hosted | Seed models through ByteDance's cloud platforms |
| Cohere | No | API, Self-hosted | Enterprise retrieval, reranking, and private deployments |
| DeepSeek | No | API, Self-hosted | MIT-licensed open weights for self-hosting or API use |
| Google DeepMind (Gemini and Gemma) | No | API, Self-hosted | Multimodal and long-context work in the Google ecosystem |
| IBM Granite | No | API, Self-hosted | Apache-licensed enterprise models with vendor support |
| Inception (Mercury) | No | API | Low-latency generation with diffusion LLMs |
| inclusionAI (Ant Group) | No | Self-hosted | Permissively licensed open models from Ant Group |
| LG AI Research (EXAONE) | No | Self-hosted | Korean-English models from an industrial research lab |
| Liquid AI | No | API, Self-hosted, Local | On-device and edge models |
| Meituan LongCat | No | API, Self-hosted | MIT-licensed open weights from a consumer tech company |
| Meta (Muse and Llama) | No | API, Self-hosted | Meta models through an API or as open weights |
| Microsoft (MAI and Phi) | No | API, Self-hosted | Small open models and Microsoft's in-house frontier models |
| MiniMax | No | API, Self-hosted | Open-weight agentic LLMs plus speech and video models |
| Mistral AI | No | API, Self-hosted | Self-deployable models from a European provider |
| Moonshot AI (Kimi) | No | API, Self-hosted | Open-weight mixture-of-experts models for agents and coding |
| Nous Research | No | Self-hosted | Open-weight Hermes models and agent tooling |
| NVIDIA Nemotron | No | API, Self-hosted | Open models with published data, tuned for NVIDIA hardware |
| OpenAI | No | API, Self-hosted | One API for text, voice, and image models |
| Reflection AI | No | API, Self-hosted | Large Apache-licensed open weights from a US lab |
| Reka | No | API, Self-hosted | Multimodal video and audio understanding |
| Sarvam AI | No | API, Self-hosted | Indian-language text and voice applications |
| SpaceXSI (formerly xAI) | No | API | Grok models for text, code, voice, and media |
| StepFun | No | API, Self-hosted | Open-weight Step models and speech APIs |
| Tencent Hy (Hunyuan) | No | API, Self-hosted | Open-weight text, image, video, and 3D models |
| Thinking Machines Lab | No | API, Self-hosted | Fine-tuning open-weight models through a managed API |
| TII (Falcon) | No | Self-hosted | Open-weight models from a government research institute |
| TypeSafe AI | No | API | Typed decisions inside automation workflows |
| Upstage (Solar) | No | API, Self-hosted | Korean and Japanese models and document parsing |
| Writer (Palmyra) | No | API | Enterprise agents on a vendor's own models |
| Xiaomi MiMo | No | API, Self-hosted | MIT-licensed reasoning and coding models |
| Z.ai (GLM) | No | API, Self-hosted | MIT-licensed GLM models for coding and agents |
Profiles

Ai2 (Olmo)
Allen Institute for AI · allenai.org · Seattle, United States · founded 2014
Ai2, the nonprofit research institute founded by Paul Allen, releases fully open models such as Olmo and Molmo together with their training data, code, and checkpoints, mostly under the Apache-2.0 license.

AI21 Labs
ai21.com · Tel Aviv, Israel · founded 2017
AI21 Labs developed the Jamba family of hybrid Transformer-Mamba models, with Jamba2 weights released under Apache-2.0, and now centers its business on post-training specialized models for enterprises with forward-deployed researchers.

Alibaba (Qwen)
Alibaba Cloud · qwen.ai · Hangzhou, China
Alibaba's Qwen team releases open-weight language, coding, vision, and image models, many under the Apache-2.0 license, and offers additional API-only models through Alibaba Cloud Model Studio.

Amazon Nova
Amazon · nova.amazon.com
Amazon Nova is Amazon's family of foundation models, including Nova 2 Lite and Nova 2 Sonic, offered through Amazon Bedrock and a Nova API key, with Nova Forge for building custom models and Nova Act for browser agents.
Anthropic
anthropic.com · San Francisco, United States · founded 2021
Anthropic develops the Claude model family, available through the Claude API and on Amazon Bedrock, Google Cloud, and Microsoft Foundry, with support for tool use, prompt caching, extended thinking, and computer use.
Arcee AI
arcee.ai
Arcee AI is a US-based open model lab that builds the Trinity family in Nano, Mini, and Large sizes, released as open weights designed to be transparent, customizable, and deployable on the infrastructure a team chooses.
Baidu (ERNIE)
Baidu · ernie.baidu.com · Beijing, China
Baidu develops the ERNIE model family, offers it through the Qianfan platform on Baidu AI Cloud, and has released open-weight ERNIE models, including the ERNIE 4.5 family under the Apache-2.0 license.
ByteDance Seed
ByteDance · seed.bytedance.com · founded 2023
ByteDance Seed, the company's AI research team established in 2023, develops the Seed language, vision, speech, and video models, serves them through Volcano Engine and BytePlus ModelArk, and publishes open-weight releases such as Seed-OSS.

Cohere
cohere.com · Toronto, Canada · founded 2019
Cohere builds enterprise models, including the Command generative models and the Embed and Rerank retrieval models, and offers them through its API, the North agent platform, Model Vault dedicated inference, and private deployments.
DeepSeek
deepseek.com · Hangzhou, China · founded 2023
DeepSeek, an AI lab based in Hangzhou, publishes its DeepSeek-V4 language models and earlier R1 reasoning models as open weights under the MIT license and serves them through an OpenAI-compatible API.
Google DeepMind (Gemini and Gemma)
Google · ai.google.dev · London, United Kingdom
Google DeepMind builds the Gemini models, which Google offers through the Gemini API, Google AI Studio, and Gemini Enterprise Agent Platform on Google Cloud, and it publishes the Gemma family of open-weight models.

IBM Granite
IBM · ibm.com · Armonk, United States
IBM Granite is a family of open models for business that covers language, vision, speech, time series, and guardrail tasks, released under the Apache-2.0 license and available on watsonx.ai, Hugging Face, and other platforms.

Inception (Mercury)
Inception Labs · inceptionlabs.ai
Inception develops Mercury, a family of diffusion large language models that generate many tokens in parallel instead of one at a time, offered through its API and enterprise deployments for latency-sensitive applications.

inclusionAI (Ant Group)
Ant Group · inclusion-ai.org · Hangzhou, China
inclusionAI, the open AI initiative backed by Ant Group, develops the Ling language models and related reasoning and multimodal models, and publishes most of them as open weights under the MIT or Apache-2.0 license.

LG AI Research (EXAONE)
LG AI Research · lgresearch.ai · Seoul, South Korea
LG AI Research develops EXAONE, a family of Korean and English language models, publishing open weights that include the K-EXAONE 2.0 series under Apache-2.0 and earlier EXAONE versions under LG's own model license.
Liquid AI
liquid.ai · Cambridge, Massachusetts, United States · founded 2023
Liquid AI builds Liquid Foundation Models (LFMs), compact open-weight language, vision, and audio models designed for low latency on phones, laptops, and other edge hardware as well as in the cloud.
Meituan LongCat
Meituan · longcat.ai · Beijing, China
Meituan's LongCat team develops LongCat language, reasoning, audio, and video models, releases them as open weights under the MIT license, and offers them through LongCat AI products, including a coding agent.
Meta (Muse and Llama)
Meta Platforms · dev.meta.ai · Menlo Park, United States
Meta serves its Muse models, including Muse Spark and Muse Wonder, through the OpenAI-compatible Meta Model API, and publishes open-weight models such as Muse Spark Open and the Llama family under Meta's own licenses.

Microsoft (MAI and Phi)
Microsoft · microsoft.ai · Redmond, United States
Microsoft AI develops the MAI models for reasoning, coding, cybersecurity, voice, and transcription, used in Copilot and GitHub Copilot, while Microsoft also publishes the Phi small language models as open weights, many under the MIT license.

MiniMax
minimax.io · Shanghai, China · founded 2021
MiniMax develops the MiniMax M-series language models, released as open weights, along with video, speech, and music models, and offers them through its API platform and the MiniMax Code agent.
Mistral AI
mistral.ai · Paris, France · founded 2023
Mistral AI releases open-weight models, many under the Apache-2.0 license, alongside commercial models, and offers them through its API, Mistral Studio, the Forge customization service, and on-premises or VPC deployments.
Moonshot AI (Kimi)
Moonshot AI · moonshot.ai · Beijing, China · founded 2023
Moonshot AI develops the Kimi models, publishes open weights for its large Kimi mixture-of-experts models on Hugging Face, and serves them through the Kimi API platform for coding, agentic, and long-context work.

Nous Research
nousresearch.com
Nous Research is an open AI research lab that releases the Hermes family of open-weight models, often post-trained from other open base models, and maintains Hermes Agent, an MIT-licensed open-source AI agent.

NVIDIA Nemotron
NVIDIA · nvidia.com · Santa Clara, United States
NVIDIA's Nemotron family includes open-weight language, reasoning, speech, and multimodal models released with training datasets under the NVIDIA Nemotron Open Model License, and packaged as NIM microservices for deployment.

OpenAI
openai.com · San Francisco, United States · founded 2015
OpenAI develops the GPT model family, offered through its API for text, reasoning, tool calling, image generation, audio, and realtime voice, and publishes the gpt-oss open-weight models under the Apache-2.0 license.
Reflection AI
reflection.ai · New York, United States · founded 2024
Reflection AI released Beam, a 501-billion-parameter mixture-of-experts model with weights under Apache-2.0, available through the Reflection API, the Mirror CLI, and self-hosting with SGLang, vLLM, or TokenSpeed.

Reka
Reka AI · reka.ai
Reka builds natively multimodal models for video, image, audio, and text, including the Spark, Edge, Flash, and Core lines, has released Reka Flash 3 weights under Apache-2.0, and now focuses on omni models for physical AI.
Sarvam AI
sarvam.ai · Bengaluru, India · founded 2023
Sarvam AI builds models for Indian languages, including the open-weight Sarvam-30B and Sarvam-105B under Apache-2.0, plus speech-to-text, text-to-speech, and document parsing APIs on its sovereign AI platform.

SpaceXSI (formerly xAI)
SpaceX · x.ai
SpaceXSI, the AI company formerly named xAI, develops the Grok models and offers them through the Grok API for text, reasoning, coding, function calling, voice, and image and video generation.
xAI now operates as SpaceXSI (SpaceX Super Intelligence), while the Grok models, the Grok API, and the x.ai and docs.x.ai domains keep their names.
StepFun
stepfun.com · Shanghai, China · founded 2023
StepFun develops the Step language, vision, and audio models, publishes many of them as open weights under the Apache-2.0 license, and serves them through the StepFun Open Platform API.
Tencent Hy (Hunyuan)
Tencent · aistudio.tencent.com · Shenzhen, China
Tencent's Hunyuan team, whose newer models carry the Hy name, develops language, image, video, and 3D models, publishes many as open weights on Hugging Face, including the Apache-2.0 Hy4 preview, and offers them through Tencent Cloud.

Thinking Machines Lab
thinkingmachines.ai · San Francisco, United States · founded 2025
Thinking Machines Lab released Inkling, a multimodal mixture-of-experts model with controllable reasoning effort, as open weights under Apache-2.0, and runs Tinker, an API for fine-tuning open-weight models.

TII (Falcon)
Technology Innovation Institute · falconllm.tii.ae · Abu Dhabi, United Arab Emirates · founded 2020
The Technology Innovation Institute in Abu Dhabi develops the Falcon models, including the hybrid Falcon-H1 series and Falcon Perception vision models, and releases them as open weights under its Falcon license or Apache-2.0.

TypeSafe AI
typesafe.ai · San Francisco, United States
TypeSafe AI builds Jev, a proprietary System One model served through its own API that answers typed questions about program state with structured values and calibrated probabilities instead of generated text.
Upstage (Solar)
Upstage · upstage.ai · Yongin, South Korea · founded 2020
Upstage, a South Korean AI company, develops the Solar language models, including Solar Pro, Solar Mini, and the open-weight Solar Open line, alongside Document Parse and other document intelligence APIs.
Writer (Palmyra)
Writer · writer.com · San Francisco, United States · founded 2020
Writer is an enterprise AI platform company that trains its own Palmyra models, including the Palmyra X6 model for agentic work, offers them inside its agent platform and through an API, and has released smaller Palmyra models openly.
Xiaomi MiMo
Xiaomi · mimo.xiaomi.com · Beijing, China
Xiaomi's MiMo team develops the MiMo language, reasoning, and speech models, releases them as open weights under the MIT license, and serves them through the Xiaomi MiMo API Open Platform and the MiMo Code agent.
Z.ai (GLM)
Zhipu AI · z.ai · Beijing, China
Z.ai, the international brand of Zhipu AI, develops the GLM language and vision models, releases many of them as open weights under the MIT license, and serves them through its API and the Z.ai assistant.
How to choose
- 01Model access
- Whether models are available only through a hosted API or also as downloadable weights that run on your own infrastructure.
- 02Weight licensing
- For open-weight models, whether the license is OSI-approved, such as MIT or Apache-2.0, or a custom license with usage or revenue conditions.
- 03Capabilities and modalities
- Support for reasoning, tool calling, structured outputs, long context, and image, audio, or video input and output.
- 04Platform availability
- Whether the models are also sold through Amazon Bedrock, Google Cloud, Microsoft Foundry, or inference providers, which affects procurement and data residency.
- 05Data handling
- Retention policies, regional processing, and options for zero data retention or private deployment.
- 06SDKs and API compatibility
- Official SDK languages and whether the API accepts OpenAI-format requests, which lowers the cost of switching.
Frequently asked questions
What is an LLM provider?
- An LLM provider is a company that trains large language models and offers them to developers, usually through an API billed per token. OpenAI, Anthropic, Google, and Mistral AI are examples. Some providers also release open weights, and many of their models are resold through cloud platforms such as Amazon Bedrock, Microsoft Foundry, and Google Cloud.
What is an open weight model?
- An open-weight model is a model whose trained parameters are published for anyone to download and run, usually on Hugging Face. Its training data and training code may stay private, and its license may add conditions. The Open Source Initiative's definition of open source AI also asks for enough information about training data to rebuild the system, so many MIT- or Apache-2.0-licensed models count only as open-weight.
Open source vs closed source LLM: how do you choose?
- Closed models served through an API require no infrastructure to operate, while open-weight models can run on your own hardware, be fine-tuned, and keep data inside your network. A common pattern combines both: a hosted proprietary model for the hardest tasks and a self-hosted open-weight model for high-volume, latency-sensitive, or privacy-restricted workloads. Evaluate both on your own tasks before committing.
Which Chinese companies make LLMs?
- Chinese model developers include DeepSeek, Alibaba's Qwen team, Moonshot AI, Z.ai (Zhipu AI), MiniMax, ByteDance Seed, Baidu, Tencent, StepFun, Xiaomi, Meituan, and Ant Group's inclusionAI. Many of them publish open weights on Hugging Face, often under MIT or Apache-2.0 licenses, and serve their models through APIs that accept OpenAI-format requests, so they can be tested alongside US models with little code change.
How do you use multiple LLM providers in one app?
- Most teams place an AI gateway or routing layer between the application and the providers. The gateway exposes one API, usually in the OpenAI format, translates requests for each provider, and adds failover, rate limits, cost tracking, and caching. Provider-agnostic SDKs are a lighter option. In both cases, keeping prompts and evaluations portable turns a model switch into a configuration change rather than a rewrite.
Related guides
LLM models and benchmarks hubOther categories
- Inference providers
- GPU clouds
- Serving engines
- Vector databases
- Embedding and reranking models
- Document parsing and data ingestion
- Fine-tuning and post-training
- Agent frameworks
- Agent infrastructure
- Web search APIs
- MCP platforms
- Voice AI
- AI gateways
- LLM observability and evaluation
- Guardrails and AI security
- AI coding tools
Product details change quickly; check each vendor’s site for current features, licensing and pricing.