fix(cli): expand model capability detection to include Llama, Nemotron, and Mistral

[aaronlippold]

Summary

Fixes false capability warnings for Llama, Nemotron, and Mistral model families in Continue CLI.

Problem

The isModelCapable function was showing warnings like:

"The model 'Llama 3.3 Nemotron 49B' is not recommended for use with cn due to limited reasoning and tool calling capabilities"

This warning was incorrect - these models have excellent capabilities.

Solution

Added detection patterns for three major model families:

• /llama/ - Meta Llama models (3.1, 3.3, Code Llama, etc.)
• /nemotron/ - NVIDIA Nemotron models
• /mistral/ - Mistral AI models (Small, Large, Codestral, etc.)

Research Validation

Llama 3.3 / Nemotron:

• Arena Hard: 85.0
• AlpacaEval 2 LC: 57.6
• Prompt engineer a faster method of code editing #1 on automatic alignment benchmarks (as of Oct 2024)
• Proven tool calling support

Mistral:

• MMLU: 81.2% (near GPT-4)
• Native function calling and JSON mode
• Strong code generation and reasoning

Testing

• ✅ All 26 tests passing
• ✅ Updated tests to reflect intentional behavior change
• ✅ Lint and format checks passing
• ✅ Verified with nvidia/Llama-3_3-Nemotron-Super-49B-v1

Impact

• Removes false warnings for widely-used model families
• Enables multiEdit tool for capable models
• Improves user experience with open-weight models
• Aligns capability detection with real-world performance

Checklist

• Tests added/updated and passing
• Code follows Continue coding standards
• Changes are backward compatible
• Research-validated change

Authored by: Aaron Lippold lippold@gmail.com

---

Summary by cubic

Fixes false capability warnings in the Continue CLI by recognizing Llama, Nemotron, and Mistral models as capable. This makes capability checks accurate and enables tools like multiEdit for these models.

• Bug Fixes
  • Added /llama/, /nemotron/, /mistral/ patterns to isModelCapable (matches name or model).
  • Updated tests to mark these families capable across providers; 26/26 passing.

^{Written for commit d494bc1. Summary will update automatically on new commits.}

[cubic-dev-ai]

No issues found across 2 files

[RomneyDa]

@aaronlippold we've not included 7B models because they often have poor outputs and are (debatably of course) relatively on the less capable end. In our experience the CLI performs poorly with most llama models and we'd want users to be warned

Thoughts on an environment variable that hides this warning instead? Or perhaps, making it show for only a couple consecutive sessions?

[aaronlippold]

Sure, I’d be happy to do it that way. I will take a look at it tomorrow I am also a contributor that doesn’t mind if you just cherry pick what you like and tell me why you left the other stuff up. It’s your code base after all :-)

…

-------- Aaron Lippold ***@***.*** 260-255-4779 twitter/aim/yahoo,etc. 'aaronlippold'

On Tue, Nov 25, 2025 at 23:15 Dallin Romney ***@***.***> wrote: ***@***.**** requested changes on this pull request. @aaronlippold <https://github.com/aaronlippold> we've not included 7B models because they often have poor outputs and are (debatably of course) relatively on the less capable end Thoughts on an environment variable that hides this warning instead? — Reply to this email directly, view it on GitHub <#8845 (review)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AALK42H533ZDAU23QLTGV7L36USOXAVCNFSM6AAAAACM4WM632VHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMZTKMBYGU2DQNRUHA> . You are receiving this because you were mentioned.Message ID: ***@***.***>