Capacité d’agent des LLM locaux (appel d’outils)
Réponse rapide
Un LLM local qui discute bien peut quand même être inutile pour piloter un agent — émettre de façon fiable un appel d'outil valide est une compétence différente, entraînée. À la mi-2026, la couronne open-weight de l'appel d'outils appartient à Qwen, qui occupe environ 10 des 13 places du Berkeley Function-Calling Leaderboard ; pour le travail agentique plus lourd (tau-bench, SWE-bench), GLM-4.5/4.6 et Kimi K2.6 mènent, tous sous licences permissives. La distinction qui décide si un modèle fonctionne comme cerveau d'agent est s'il possède un modèle d'outil NATIF entraîné (Qwen, Mistral, Llama, Hermes, Granite, GLM, Phi-4-mini l'ont tous) ou s'il a besoin d'un décodage contraint par grammaire ajouté (Gemma 2/3, Phi-4 de base et DeepSeek-R1 ne l'ont pas) — et la fiabilité s'effondre sous environ 4 milliards de paramètres. Cette base note chaque famille sur les deux, chaque chiffre de benchmark étant daté car ce classement évolue chaque mois.
Le plus petit pilote d'agent local sûr : le trio ~4B — Phi-4-mini (MIT, jetons d'outil natifs, le choix unique le plus sûr), Qwen3-4B-2507, Gemma 4 E4B — toujours jumelé à un décodage contraint par grammaire. Bête de somme plus grande sur un seul GPU : Devstral Small 2 (24B, code) ou Qwen3 14B/32B. Poids lourd agentique propre pour le commercial : GLM-4.5-Air (MIT). Évitez Cohere Command en production (toute la gamme est non-commerciale). Et revérifiez chaque score après août 2026 — cette liste pourrit vite.
🔴 Fraîcheur : 2026-07-18 — ce classement évolue chaque mois. BFCL v3 et v4 sont des suites différentes et NON comparables entre elles. Revérifiez à gorilla.cs.berkeley.edu avant de vous fier à un chiffre après ~août 2026.
| Famille de modèle | Niveau d'outil | Format natif ? | Benchmark (daté) | Plus petit fiable | Licence |
|---|---|---|---|---|---|
| Qwen3.5 (open sizes)Alibaba Qwen · MoE 397B-A17B / 122B-A10B / 35B-A3B; dense 27B/9B/4B/2B | excellent | natif | BFCL-v4 (2026-07-18): 397B-A17B 0.729, 27B 0.685, 9B 0.661, 4B 0.503, 2B 0.436 — highest-scoring OPEN family | 9B usable, 27B recommended; 4B marginal, <=2B unreliable | permissiveApache-2.0 |
| Idéal pour : Top open-weight general tool-caller; 27B dense = single-GPU sweet spot. Qwen owns ~10 of 13 BFCL-v4 slots. Score collapse 9B(0.661)->4B(0.503) marks the practical agent floor. The board-topping Qwen3.7 Max is proprietary API. | |||||
| Qwen3-Next-80B-A3B / Qwen3-235B-A22BAlibaba Qwen · 80B-A3B (3B active); 235B-A22B (2507) | excellent | natif | BFCL-v3 (2026-07-18): Qwen3-Next-80B-A3B-Thinking 0.720 (top open), 235B-Thinking-2507 0.719 | 80B-A3B (3B active) best perf/cost; 235B for max accuracy | permissiveApache-2.0 |
| Idéal pour : Highest-accuracy open general tool-caller; 3B-active MoE keeps inference cheap. For Thinking models avoid ReAct/stopword tool templates (emits stopwords mid-thought, breaks parsing — Qwen docs). | |||||
| Qwen3-Coder-Next (80B-A3B) / Qwen3-Coder (480B)Alibaba Qwen · 480B-A35B; Coder-Next 80B-A3B | excellent (coding agents) | natif | Coder-Next 80B-A3B >70% SWE-bench Verified w/ SWE-Agent (2026-03-03); 480B ~Claude-Sonnet-4-class | 80B-A3B (3B active) delivers flagship agentics at low cost | permissiveApache-2.0 |
| Idéal pour : Long-horizon coding agents with heavy tool use + failure recovery. 3B-active MoE is the standout cost/performance pick for local coding agents; trained via large-scale agentic RL. | |||||
| Qwen3 dense (4B–32B)Alibaba Qwen · 0.6B/1.7B/4B/8B/14B/32B (dual-mode) | good (14B/32B), fair (8B), poor (<=1.7B) | natif | tau2-Telecom: 14B 32%, 4B-2507 27% (ToolMind, arXiv 2511.15718). Param count a WEAK predictor <8B (non-monotonic) | 4B usable (esp 2507), 8B+ recommended; 1.7B unreliable | permissiveApache-2.0 |
| Idéal pour : Edge/on-device tool agents — 4B-2507 is the smallest genuinely reliable Qwen tool-caller. Non-thinking mode + xgrammar is the standard recipe for 4B/8B agents. Always grammar-constrain below 8B. | |||||
| GLM-4.5 / GLM-4.5-AirZhipu / Z.ai · 355B-A32B; Air 106B-A12B | excellent (ARC: agentic/reasoning/coding) | natif | tau-bench 70.1%, BFCL-v3 ~76.7% best-overall-among-baselines, SWE-bench Verified 64.2% (arXiv 2508.06471, Jul/Aug 2025). Beats Gemini 2.5 Pro on tau-bench | Air (106B/12B active) — multi-GPU or heavily-quantized big single GPU; NOT a laptop model | permissiveMIT |
| Idéal pour : Open-weight coding/agentic workhorse for multi-step tool + browser + code-exec loops. THE 2026 standout for open agentic tool use — near-frontier tau-bench/BFCL with an MIT license. Only cost is VRAM. | |||||
| GLM-4.6Z.ai · ~357B MoE, 200K ctx | excellent | natif | near Claude-Sonnet-class on coding/agentic (Z.ai release 2025-09-30); exact tau/BFCL deltas over 4.5 unverified | large MoE only (use 4.5-Air if smaller needed) | permissiveMIT |
| Idéal pour : Long-context coding + search/browser agent, self-hosted, commercial-OK. Coding/agent-focused refresh of 4.5. Later GLM-4.7/GLM-5 references exist in 2026 sources but are only partly verified. | |||||
| Kimi K2 / K2.6Moonshot AI · 1T-A32B MoE | excellent | natif | tau2-Telecom 96% (K2.6); SOTA on Humanity's Last Exam / BrowseComp with tools (Moonshot 2026). Stable across 200-300 sequential calls | 1T MoE (32B active) — hosting is the barrier, needs a serious multi-GPU rig | permissiveModified MIT |
| Idéal pour : Long-horizon agentic coding + deep tool loops (200-300 calls) where you can host a 1T MoE. K2.6 released 2026-04-20. Frontier-class open agentic model, but the 1T footprint is out of reach for typical self-hosters despite the permissive license. | |||||
| DeepSeek-V3.2DeepSeek-AI · 671B-A37B MoE | excellent | natif | tau2-bench: Airline 63.8, Retail 81.1, Telecom 96.2 (arXiv 2512.02556, Dec 2025) | 671B only — server-class multi-GPU | permissiveMIT (verify per download) |
| Idéal pour : Reasoning-heavy multi-turn agents interleaving thinking + tool calls. Best-documented DeepSeek agent numbers. Watch inference-engine compat for thinking-mode tool calls — template changed vs V3.1. | |||||
| Devstral Small 2 / Devstral 2Mistral (w/ All Hands AI) · Small 2 24B; Devstral 2 123B | excellent (coding agents) | natif | SWE-bench Verified: Devstral 2 (123B) 72.2%, Small 2 (24B) 68.0% (2025-12-09) — among best open-weight SWE-bench | Small 2 (24B) runs on one RTX 4090 / 32GB Mac, still 68% SWE-bench | permissiveApache-2.0 (Small; verify 123B card) |
| Idéal pour : THE open-weight choice for self-hosted coding agents (repo-level bug-fix / PR automation). 256K context; ships open-source Mistral Vibe CLI. Strongest agentic-coding story in the Mistral camp. | |||||
| Mistral Small 3.x / 4 (24B)Mistral AI · 24B (3.1/3.2 add vision) | good (3.2 improved FC) | natif | no clean public BFCL v3/v4 entry (2026-07-18); vendor 'improved FC' notes only — unverified | 24B is the family floor, reliable for tools | permissiveApache-2.0 |
| Idéal pour : Best single-GPU (24B) open agent worker: native format + permissive license. Mistral 3 (Large 3 + Ministral 3) all Apache-2.0 too — replaces the non-commercial Large 2 for blocked self-hosters. | |||||
| Hermes 4 (14B / 70B / 405B)Nous Research · 14B (Qwen3 base); 70B/405B (Llama-3.1 base) | good (14B) -> excellent (70B/405B) | natif | Hermes 4 released 2025-08-26, SOTA-among-open on Nous suite; no canonical dated BFCL/tau row (unverified) | 14B — Apache-2.0, no MAU cap (Qwen3 base) | permissive (14B) / restricted (70B+)MIXED: 14B Apache-2.0; 70B/405B Llama Community License |
| Idéal pour : Neutral-aligned, low-refusal local tool agent; the 14B is the best small commercial pick here. The Hermes tool format is now a de-facto standard reused across Qwen/others. Hermes 4-14B / Apache-2.0 is a standout. | |||||
| IBM Granite 4.xIBM · 4.1 up to ~30B, 512K ctx; 4.0 H-Small 32B-A9B, H-Tiny 7B, Micro 3B, Nano 350M/1B | good (8B) -> excellent (Small/30B) | natif | BFCL-v3 (2026): 4.1 8B 68.27%, 4.1 30B 73.68% (ahead of Gemma-3-27B-class) | ~7-8B reliably; 3B Micro for edge; 350M-1B Nano for constrained edge tool calling | permissiveApache-2.0 (signed, ISO-42001 certified) |
| Idéal pour : Memory-efficient, commercially-clean enterprise/edge agent (Mamba hybrid = low KV-cache for long tool loops). The safest COMMERCIAL self-host pick: permissive license + efficiency + dated BFCL numbers. | |||||
| Llama 3.3 70BMeta · 70B only | good | natif | BFCL v2 tool-use 77.3% (2024-12-06; top open at the time). Llama no longer on the LIVE BFCL v3 top table | 70B (only size) — ~405B-class tool use at 70B footprint | restrictedLlama 3.3 Community License (>700M MAU cap) |
| Idéal pour : Best-value self-hosted Llama agent brain. The practical Llama pick for agents in 2026, but its benchmark lead has been overtaken by Qwen/GLM. Meta's 2026 frontier work went CLOSED-weight. | |||||
| Phi-4-miniMicrosoft · 3.8B (+ multimodal ~5.6B) | good (for its size) | natif | BFCL v4 low-to-mid 80s (ertas third-party, May 2026) — no official Berkeley row | IT IS the small option (3.8B), fits an 8GB GPU | permissiveMIT |
| Idéal pour : On-device MIT-licensed tool agent — the safest single small pick. Real native tool template + MIT (fewer license strings). Note the LARGER base Phi-4 (14B) LACKS this template — only the mini has it. | |||||
| Gemma 4 (E4B / 26B-A4B / 31B)Google · E2B/E4B; 26B-A4B MoE; 31B dense (2026-03-31) | good | natif | BFCL v4: 26B-A4B ~89% non-live, E4B mid-to-high 80s — both THIRD-PARTY (unsloth/ertas), not official Berkeley | E4B (4B effective) smallest reliable general tool-caller | permissive (unverified)Apache-2.0 claimed on release page — UNVERIFIED (conflicts with Gemma's historical custom license; confirm on the weight download) |
| Idéal pour : On-device/edge agent needing native structured tool calls with a small footprint. Resolves Gemma's no-native-format history (Gemma 2/3 had NO tool template — grammar-constrain only). Verify the Apache-2.0 claim — material if true. | |||||
| Gemma 2 / 3Google · Gemma 2: 2B/9B/27B; Gemma 3: 1B/4B/12B/27B + 3n | poor (2/3) — no native tool support | grammaire seulement | none official | 27B, and only with heavy prompt scaffolding | restrictedGemma Terms of Use (custom, non-OSI) |
| Idéal pour : Avoid for tools; chat/summarization with grammar-constrained JSON if forced. THE baseline 'no native tool template' case — every tool call is prompt-engineered + parsed, YOU supply the structure via grammar. Google shipped FunctionGemma-270M and Gemma 4 tokens specifically to fix this. | |||||
| Command-R / R+ / R7B / Command ACohere · R 35B; R+ 104B; R7B 7B; Command A 111B | good (R/R7B) -> excellent (R+/Command A) | natif | R+ beat GPT-4 Turbo on ToolTalk(Hard) + topped BFCL single-turn FC (Cohere eval 2024, ~84 BFCL secondary — exact unverified) | R7B (7B) smallest agent-capable, single 12-16GB GPU; good tool-ABSTENTION | non-commercialCC-BY-NC-4.0 + AUP — NON-COMMERCIAL across the ENTIRE open lineup |
| Idéal pour : On-prem RAG agent with grounded citations + multi-step/parallel tools (research/non-commercial only). 🔴 Capability-strong RAG+tool combo, but the non-commercial license disqualifies most production self-hosting. R7B's know-when-NOT-to-call is a rare, valuable property. | |||||
| DeepSeek-R1 + distillsDeepSeek-AI · R1 671B; distills 1.5B-70B (Qwen/Llama bases) | poor -> fair (community templates) | partiel | none reliable — R1 targeted reasoning (AIME/MATH), not function-calling | 14B distill smallest marginally usable for tools; <8B weak | permissiveMIT (R1 + all six distills) |
| Idéal pour : Reasoning/planning LAYER feeding a separate tool-executor — not a direct tool-caller. Tool calling was a known R1 weak point, fixed in V3.1/V3.2 not R1. <think> tags break naive tool-call parsers. Same caveat applies to base Phi-4 and Phi-4-reasoning. | |||||
Le plus petit pilote fiable. The smallest genuinely reliable, commercially-clean local agent driver in 2026 is the ~3.8-4B on-device trio: Phi-4-mini (3.8B, MIT, native tool tokens), Qwen3-4B-Instruct-2507 (Apache-2.0, native template, strong with grammar constraints), and Gemma 4 E4B (native tokens). Phi-4-mini is the safest single pick — real trained-in tool tokens + MIT + fits an 8GB GPU. Pair ANY of them with grammar-constrained decoding and per-task eval; below ~4B, drop to a fine-tuned specialist (FunctionGemma-270M) rather than expect general tool-calling. For a bigger single-GPU workhorse, Devstral Small 2 (24B) or Qwen3 14B/32B is the reliability step-up.
Format natif vs décodage contraint par grammaire. Grammar-constrained decoding restricts the allowed tokens at each step so the output physically cannot leave your schema — llama.cpp GBNF, Outlines (Pydantic), and XGrammar (now the DEFAULT backend in vLLM and SGLang, ~100x faster than earlier libs; exposed via guided_json/guided_grammar/guided_regex) let even models with NO native tool template (Gemma 2/3, base Phi-4, DeepSeek-R1) emit valid tool calls and kill the validate-retry loop. Key distinction: a NATIVE trained-in format makes the model reliably CHOOSE the right call; grammar-constraining only guarantees the SHAPE is valid — you want both, and on weak/small models grammar is what makes them usable at all.
Données ouvertes (CC BY 4.0) : CSV · JSON · API : /wp-json/dc/v1/agent-capability
Produits, réparations et guides connexes
- hub IA auto-hébergée pour Bitcoiners
- guide des plebs pour l’IA auto-hébergée
- installer Ollama en 10 minutes
- LM Studio vs Ollama vs llama.cpp
- connecter l’IA locale à Home Assistant et Obsidian
- dépannage de l’IA auto-hébergée
- convertir du matériel de minage en hashcenter IA
- classements des modèles d’IA locale
Dernière révision: 27 juillet 2026.
