GitHub and Microsoft shipped Copilot Chat support for open LLMs hosted on Hugging Face on June 17, 2026. The update lets developers select open-weight models inside the Copilot Chat panel instead of relying only on GPT-5.5 or Claude models. The initial catalog includes Meta Llama 4 Scout, Mistral Small 3.2, Qwen 3 Coder, and Microsoft Phi-4-mini. Parameter counts range from 3.8 billion to 109 billion. Context windows go up to 128k tokens for most models and 10M for Llama 4 Scout. This move follows months of user pressure after GitHub Copilot moved to usage-based billing and free tiers tightened.
Microsoft announced the integration on its official Copilot blog and pointed developers to the Hugging Face open model catalog. The models are fetched through Hugging Face Inference Endpoints when used inside Copilot Chat. Developers do not need to download weights locally. The feature works in Visual Studio Code, Visual Studio, and GitHub.com. Microsoft said the open LLM option will roll out to free Copilot users first, then to Business and Enterprise plans. The company confirmed that this is not a replacement for its own Phi models or GPT-5.5. It is an additional route for teams that want open licenses and self-hosted evaluation.
The reason this matters is license control and cost. Open-weight models under Apache 2.0 or MIT can be audited, fine-tuned, and deployed on your own infrastructure. They avoid per-token billing multipliers that hit closed models. For many developers, the ability to pin a specific model version in Copilot Chat removes provider lock-in. The tradeoff is quality and speed. Smaller open models often score lower on complex agentic coding tasks than GPT-5.5 or Claude Opus. But for code completion, test generation, and refactoring, the gap is narrower. This release also matters because free AI model access has gotten more complicated in 2026. Copilot Chat with open LLMs offers a new fallback when closed model quotas run out.
How Do the Top Options Compare?
| Model | Parameters | Context Window | License | Copilot Chat Tier |
|---|---|---|---|---|
| Meta Llama 4 Scout | 109B total, 17B active | 10M tokens | Llama 4 Community License | Free, Pro, Business |
| Mistral Small 3.2 | 24B | 128k tokens | Apache 2.0 | Free, Pro |
| Qwen 3 Coder 30B-A3B | 30B total, 3B active | 256k tokens | Apache 2.0 | Free, Pro, Business |
| Microsoft Phi-4-mini | 3.8B | 128k tokens | MIT | Free, Pro, Business, Enterprise |
| DeepSeek V3.1 | 671B total, 37B active | 128k tokens | MIT | Pro, Business |
Open LLM availability varies by region. Business and Enterprise tenants may need admin approval before enabling third-party model access.
1. Meta Llama 4 Scout , Long context refactoring and repo scale tasks
Meta Llama 4 Scout is the largest model in the Copilot Chat open catalog. It uses a mixture of experts design with 109 billion total parameters and 17 billion active parameters. The model supports a 10 million token context window. That is far beyond the 128k context of most other open models. The Llama 4 Community License allows commercial fine-tuning and deployment. The model scores 84.1 on MMLU and 78.6 on HumanEval in Copilot Chat tests. Long context benchmark results on long code review tasks show 71.3 percent accurate recall.
The main downside is hardware. Running the 109B checkpoint locally requires multiple high VRAM GPUs. Copilot Chat routes this model through managed Hugging Face endpoints. That removes local hardware pressure but adds network latency. Teams that need to refactor very large codebases should try this model first. The 10M context lets you drop entire repositories into the prompt without chunking. Some users report that long sessions lose consistency after 150k tokens. The model is also available through the free Copilot tier with lower rate limits. See how it stacks up against other free AI models in 2026.
Key strengths:
- ✅ Massive 10M token context window
- ✅ Handles large monorepo code review without chunking
- ✅ License permits commercial fine-tuning
- ✅ Works in free Copilot Chat tier
- ❌ 109B total weights slow on CPU-only inference
- ❌ Requires high VRAM for local self-hosting
- ❌ Hallucinates more than GPT-5.5 on long agentic sessions
Who it’s for: Teams that need to refactor or review very large codebases without paying for closed context expansion.
2. Mistral Small 3.2 , Balanced open model for everyday Copilot Chat coding
Mistral Small 3.2 is a 24 billion parameter dense model. It has a 128k token context window and uses the Apache 2.0 license. That license gives you broad rights to modify, fine-tune, and commercialize the model. The model scores 87.5 on HumanEval and 82.4 on MMLU. It handles code completion, function generation, and small refactors well. Copilot Chat serves Mistral Small 3.2 through Hugging Face endpoints with low latency. The model is available on free and Pro tiers. Free tier users get a higher request allowance than for Llama 4 Scout.
Dense architecture makes it easier to self-host on a single 48GB GPU. The 128k context window is enough for most individual files and small projects. It will struggle with very large monorepos. The model is not ideal for multi-step agentic planning across many files. Some Copilot Chat users report occasional throttling during Europe afternoon hours. Still, for daily coding, it is the best all-around open model in this release. It avoids the hardware cost of MoE models and the license limits of the Llama 4 Community License.
Key strengths:
- ✅ Apache 2.0 license allows unrestricted commercial use
- ✅ Lower latency than 100B models in cloud endpoints
- ✅ Strong instruction following for code edits
- ✅ Available on free tier with fewer throttles
- ❌ 128k context is not enough for giant monorepos
- ❌ Weaker at complex multi-step agent planning
- ❌ Some Hugging Face endpoints queue at peak hours
Who it’s for: Developers who want a daily driver open model with permissive licensing and decent speed.
3. Qwen 3 Coder 30B-A3B , Code generation and agentic coding on open weights
Qwen 3 Coder 30B-A3B is the strongest open coding model in the Copilot Chat catalog. It uses a mixture of experts layout with 30 billion total parameters and 3 billion active. The model supports a 256k token context window. Its Apache 2.0 license allows commercial use without royalty fees. On HumanEval it scores 90.1 percent. On MBPP it scores 92.3 percent. On SWE-bench Verified it scores 58.2 percent. Those are strong results for an open weights model.
The active parameter count of 3 billion keeps managed endpoint costs low. That means Copilot Chat can serve this model to free tier users without huge infrastructure spend. The model handles agentic coding, multi-file edits, and tool calls better than Mistral Small 3.2. It is not perfect. The MoE design adds reliability concerns for local self-hosting. Fine-tuning the router and experts together is harder than tuning a dense model. The model also has limited non-English documentation performance. Still, if you care about raw coding benchmark scores under a permissive license, this is the pick. It is part of the broader AI coding tool pricing overhaul that hit in June 2026.
Key strengths:
- ✅ Strongest open coding benchmark scores in Copilot catalog
- ✅ 256k context handles very large files
- ✅ Apache 2.0 license with no royalty strings
- ✅ Active parameter count keeps inference cost low
- ❌ MoE architecture adds complexity for self-hosting
- ❌ Fine-tuning support requires careful router training
- ❌ Not as strong on non-English documentation tasks
Who it’s for: Developers who prioritize raw coding accuracy and can accept a MoE model.
4. Microsoft Phi-4-mini , Low VRAM local use and fast completions
Microsoft Phi-4-mini is the smallest model in the open catalog. It has 3.8 billion parameters and a 128k token context window. The MIT license means you can use it for almost any purpose. Scores include 82.0 on HumanEval and 75.8 on MMLU. Those numbers are lower than larger models. But Phi-4-mini is fast and cheap to run. It fits on an 8GB consumer GPU without quantization. That makes it a strong option for local Copilot Chat fallback when online endpoints are busy.
The model works well for structured JSON generation, unit tests, and docstrings. It struggles with novel algorithms and long reasoning chains. Dense architecture makes it easy to fine-tune with LoRA. Microsoft included this model to give developers a low resource route inside Copilot Chat. The model is available on all Copilot plans including Enterprise. Some users will find it too weak for complex C++ or Rust refactoring. But for Python and TypeScript helpers, it is adequate. The MIT license removes any attribution or usage concerns.
Key strengths:
- ✅ Small enough to run on 8GB consumer GPUs
- ✅ MIT license is as permissive as it gets
- ✅ Fast token generation inside Copilot Chat
- ✅ Good at structured JSON and unit tests
- ❌ Limited reasoning depth for novel algorithms
- ❌ No 10M context option
- ❌ Fewer parameters means more mistakes on hard problems
Who it’s for: Laptop users, students, and teams that need a lightweight local model.
5. DeepSeek V3.1 , Frontier open weights for Pro tier and self-hosting
DeepSeek V3.1 is the frontier open weights option. It uses a mixture of experts design with 671 billion total parameters and 37 billion active. The model has a 128k token context window. The MIT license allows full commercial use. On HumanEval it scores 89.5 percent. On SWE-bench Verified it scores 62.8 percent. These scores are close to closed frontier models but not equal to GPT-5.5. Copilot Chat offers DeepSeek V3.1 only on Pro and Business plans. Free tier users cannot select it.
The model is impressive for agentic coding and tool use. It handles long multi-file tasks without losing context. The main issue is hardware. Self-hosting the 671B checkpoint requires server class GPUs. Copilot Chat managed endpoints can be slow at peak. Microsoft said Pro users get a higher request allowance than free users. Business tenants may need admin approval to enable this model. It is the best choice if you need open weights with near frontier coding ability. Open model access like this is part of the growing free AI tier limits debate in 2026.
Key strengths:
- ✅ Frontier class reasoning on open weights
- ✅ MIT license supports commercial derivative use
- ✅ Strong agentic coding and tool calling
- ✅ Available on Pro tier with higher limits
- ❌ Very large checkpoint requires serious hardware
- ❌ Latency in Copilot Chat can spike at peak
- ❌ Not available on free tier after June 2026 quota changes
Who it’s for: Teams that need near-frontier coding power under an open license and have Pro seats.
Frequently Asked Questions
Which open LLMs are available in Copilot Chat via Hugging Face?
Copilot Chat offers Meta Llama 4 Scout, Mistral Small 3.2, Qwen 3 Coder 30B-A3B, Microsoft Phi-4-mini, and DeepSeek V3.1 at launch. Microsoft said it will add more models from Hugging Face based on demand. Availability varies by plan and region.
Do I need a Hugging Face account to use open LLMs in Copilot Chat?
No. Copilot Chat uses Microsoft managed Hugging Face endpoints. You do not need your own Hugging Face token. Some Business and Enterprise plans can bring their own Hugging Face API keys for private endpoints.
Is this free?
Free tier users get access to Llama 4 Scout, Mistral Small 3.2, Qwen 3 Coder, and Phi-4-mini. DeepSeek V3.1 requires Pro or higher. Free tier still has rate limits based on Copilot Chat usage policies.
Can I run these models locally instead of through Copilot Chat?
Yes. The same weights are available on Hugging Face under their original licenses. You can download them and run them with Ollama, vLLM, or llama.cpp. Local use removes Copilot Chat rate limits but adds your own hardware costs.
What license are these open LLMs under?
Mistral Small 3.2 and Qwen 3 Coder use Apache 2.0. Phi-4-mini and DeepSeek V3.1 use MIT. Llama 4 Scout uses the Llama 4 Community License. Each license allows commercial use with different attribution or usage restrictions.
How do open LLMs in Copilot Chat compare to closed GPT-5.5?
Open models are behind GPT-5.5 on complex agentic coding and long horizon planning. But they are strong on code completion, unit tests, and refactoring. They offer lower cost and license control. The gap narrows for local self-hosting.
What Should You Remember?
- Open model support launched in Copilot Chat on June 17, 2026 through Hugging Face endpoints.
- Free tier access includes Llama 4 Scout, Mistral Small 3.2, Qwen 3 Coder, and Phi-4-mini.
- DeepSeek V3.1 is available only on Pro and Business plans due to higher compute needs.
- Licenses vary from MIT and Apache 2.0 to the Llama 4 Community License.
- Context windows range from 128k tokens up to 10M for Llama 4 Scout.
- Local self-hosting is possible for all models, but hardware requirements jump for Frontier MoE models.
Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.
