Skip to content
localmodel.run

Leaderboard · BFCL

Best local LLMs for tool use and agents

The Berkeley Function-Calling Leaderboard (BFCL) scores whether a model picks the right tool and calls it with the right arguments, across simple, parallel and multi-call invocations. We use its Overall Acc column, taking the native function-calling (FC) result where a model has one. It is the closest open, sourced measure of how agent-ready a local model is.

19 models · ranked by BFCL runs on
  1. 1 59.1% KI Kimi K2 Instruct 1000B MoE
  2. 2 48.7% Qwen3 32B 32B
  3. 3 46.5% CR Command A 111B
  4. 4 42.6% Qwen3 8B 8B
  5. 5 41% Qwen3 14B 14B
  6. 6 37.3% Llama 4 Maverick 400B MoE
  7. 7 31.9% Llama 3.3 70B 70B
  8. 8 30.4% Gemma 3 12B 12B
  9. 9 29.5% Gemma 3 27B 27B
  10. 10 28.8% Phi-4 14B 14B
  11. 11 28.4% Qwen3 1.7B 1.7B
  12. 12 28.1% Llama 4 Scout 109B MoE
  13. 13 27% FN Falcon3 10B 10B
  14. 14 25.8% Llama 3.1 8B 8B
  15. 15 23.9% Qwen3 0.6B 600M
  16. 16 22% Llama 3.2 3B 3B
  17. 17 19.6% Gemma 3 4B 4B
  18. 18 10.8% Llama 3.2 1B 1B
  19. 19 7.2% Gemma 3 1B 1B

Score is BFCL (overall acc on the berkeley function-calling leaderboard.), sourced from gorilla.cs.berkeley.edu/leaderboard. The "runs on" column is the lightest tracked device that runs the model at Q4_K_M; tap a model for the full hardware list. Coverage is partial: only models we verified against the canonical board appear.

On agentic leaderboards

BFCL tests a concrete capability: does the model call the right tool with the right arguments? Newer agent-focused boards such as Agent Arena rank whole-task orchestration, where today's leaders are still frontier cloud models and open-weight models are well behind. BFCL is the open, sourced signal that covers the local models you can actually run.

Other boards

Best local LLMs for coding Percent of 225 Exercism exercises solved across six languages. Best local LLMs by human preference Human-preference Elo from LMArena's blind head-to-head votes.

Or pick by what you need with use-case picks, or browse the full model list.

FAQ

What is the best local model for tool use?

Kimi K2 Instruct tops this board at 59.1% on BFCL. It needs ~586.6 GB at Q4_K_M, more than any single tracked device, so it wants a high-memory rig.

How is this tool use leaderboard scored?

The Berkeley Function-Calling Leaderboard (BFCL) scores whether a model picks the right tool and calls it with the right arguments, across simple, parallel and multi-call invocations. We use its Overall Acc column, taking the native function-calling (FC) result where a model has one. It is the closest open, sourced measure of how agent-ready a local model is. We only list a model after verifying its score against the canonical leaderboard for that exact model, so coverage is partial by design. Source: gorilla.cs.berkeley.edu/leaderboard.

Can I run these models locally?

Every row shows the lightest tracked device that runs the model at Q4_K_M and links to the full hardware list, so you see the ranking and your hardware fit together.

Benchmark scores are third-party, sourced from gorilla.cs.berkeley.edu/leaderboard, and reflect the full-precision model, not a specific quant. Memory and fit are computed at Q4_K_M, validated 2026-08-03. See methodology.