{
    "meta": {
        "title": "D-Central — Local LLM Agent-Capability Database (Tool-Calling)",
        "description": "Which of 18 self-hostable open-weight LLM families can reliably drive agents and tools: tool-call reliability tier (BFCL / tau-bench, dated), native trained-in tool template vs grammar-constrained decoding, parallel calls, smallest reliable size, and a license commercial-safety flag. A monthly-moving leaderboard — every score is dated.",
        "generated": "2026-07-18T17:40:38+00:00",
        "as_of": "2026-07-18",
        "version": "1.0",
        "license": "https://creativecommons.org/licenses/by/4.0/",
        "license_name": "CC BY 4.0",
        "source": "https://d-central.tech/local-llm-agent-capability/",
        "benchmark_of_record": "Berkeley Function-Calling Leaderboard (BFCL v3/v4; v3 and v4 NOT cross-comparable) + tau-bench/tau2-bench; SWE-bench Verified for coding agents.",
        "method": "The agent-capability layer on the local-LLM model database. The load-bearing distinction is NATIVE trained-in tool template (model reliably CHOOSES the right call) vs grammar-constrained decoding bolted on (guarantees only the SHAPE is valid — required for models with no native format). Every benchmark claim is DATED because this leaderboard moves monthly. Sub-10B BFCL-v4 figures that are third-party (not the official Berkeley harness) are flagged; non-commercial licenses are flagged.",
        "provenance": "gorilla.cs.berkeley.edu (BFCL v3/v4), tau-bench/tau2-bench, SWE-bench Verified, model cards, release notes; verified 2026-07-18.",
        "disclaimer": "FRESHNESS: re-verify every score after ~August 2026 — models change monthly. License flags (Cohere Command CC-BY-NC entire lineup; Llama Community License >700M MAU; Qwen2.5-3B research-only) must be re-checked at source before commercial self-hosting. Not legal advice on licensing."
    },
    "rows": [
        {
            "model": "Qwen3.5 (open sizes)",
            "developer": "Alibaba Qwen",
            "sizes": "MoE 397B-A17B / 122B-A10B / 35B-A3B; dense 27B/9B/4B/2B",
            "tool_tier": "excellent",
            "native_format": "yes (Hermes-style, in tokenizer)",
            "grammar": "yes (hermes parser; xgrammar/GBNF)",
            "parallel": "yes",
            "benchmark": "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",
            "min_size": "9B usable, 27B recommended; 4B marginal, <=2B unreliable",
            "license": "Apache-2.0",
            "commercial": "permissive",
            "best_use": "Top open-weight general tool-caller; 27B dense = single-GPU sweet spot",
            "grade": "A",
            "notes": "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."
        },
        {
            "model": "Qwen3-Next-80B-A3B / Qwen3-235B-A22B",
            "developer": "Alibaba Qwen",
            "sizes": "80B-A3B (3B active); 235B-A22B (2507)",
            "tool_tier": "excellent",
            "native_format": "yes (Hermes-style)",
            "grammar": "yes",
            "parallel": "yes",
            "benchmark": "BFCL-v3 (2026-07-18): Qwen3-Next-80B-A3B-Thinking 0.720 (top open), 235B-Thinking-2507 0.719",
            "min_size": "80B-A3B (3B active) best perf/cost; 235B for max accuracy",
            "license": "Apache-2.0",
            "commercial": "permissive",
            "best_use": "Highest-accuracy open general tool-caller; 3B-active MoE keeps inference cheap",
            "grade": "A",
            "notes": "For Thinking models avoid ReAct/stopword tool templates (emits stopwords mid-thought, breaks parsing — Qwen docs)."
        },
        {
            "model": "Qwen3-Coder-Next (80B-A3B) / Qwen3-Coder (480B)",
            "developer": "Alibaba Qwen",
            "sizes": "480B-A35B; Coder-Next 80B-A3B",
            "tool_tier": "excellent (coding agents)",
            "native_format": "yes (qwen3_coder parser)",
            "grammar": "yes",
            "parallel": "yes",
            "benchmark": "Coder-Next 80B-A3B >70% SWE-bench Verified w/ SWE-Agent (2026-03-03); 480B ~Claude-Sonnet-4-class",
            "min_size": "80B-A3B (3B active) delivers flagship agentics at low cost",
            "license": "Apache-2.0",
            "commercial": "permissive",
            "best_use": "Long-horizon coding agents with heavy tool use + failure recovery",
            "grade": "A",
            "notes": "3B-active MoE is the standout cost/performance pick for local coding agents; trained via large-scale agentic RL."
        },
        {
            "model": "Qwen3 dense (4B–32B)",
            "developer": "Alibaba Qwen",
            "sizes": "0.6B/1.7B/4B/8B/14B/32B (dual-mode)",
            "tool_tier": "good (14B/32B), fair (8B), poor (<=1.7B)",
            "native_format": "yes (Hermes-style, every size)",
            "grammar": "yes (relied on heavily at small sizes)",
            "parallel": "yes (reliability drops with size)",
            "benchmark": "tau2-Telecom: 14B 32%, 4B-2507 27% (ToolMind, arXiv 2511.15718). Param count a WEAK predictor <8B (non-monotonic)",
            "min_size": "4B usable (esp 2507), 8B+ recommended; 1.7B unreliable",
            "license": "Apache-2.0",
            "commercial": "permissive",
            "best_use": "Edge/on-device tool agents — 4B-2507 is the smallest genuinely reliable Qwen tool-caller",
            "grade": "A",
            "notes": "Non-thinking mode + xgrammar is the standard recipe for 4B/8B agents. Always grammar-constrain below 8B."
        },
        {
            "model": "GLM-4.5 / GLM-4.5-Air",
            "developer": "Zhipu / Z.ai",
            "sizes": "355B-A32B; Air 106B-A12B",
            "tool_tier": "excellent (ARC: agentic/reasoning/coding)",
            "native_format": "yes (GLM agentic template)",
            "grammar": "yes",
            "parallel": "yes",
            "benchmark": "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",
            "min_size": "Air (106B/12B active) — multi-GPU or heavily-quantized big single GPU; NOT a laptop model",
            "license": "MIT",
            "commercial": "permissive",
            "best_use": "Open-weight coding/agentic workhorse for multi-step tool + browser + code-exec loops",
            "grade": "A",
            "notes": "THE 2026 standout for open agentic tool use — near-frontier tau-bench/BFCL with an MIT license. Only cost is VRAM."
        },
        {
            "model": "GLM-4.6",
            "developer": "Z.ai",
            "sizes": "~357B MoE, 200K ctx",
            "tool_tier": "excellent",
            "native_format": "yes (GLM template)",
            "grammar": "yes",
            "parallel": "yes",
            "benchmark": "near Claude-Sonnet-class on coding/agentic (Z.ai release 2025-09-30); exact tau/BFCL deltas over 4.5 unverified",
            "min_size": "large MoE only (use 4.5-Air if smaller needed)",
            "license": "MIT",
            "commercial": "permissive",
            "best_use": "Long-context coding + search/browser agent, self-hosted, commercial-OK",
            "grade": "B",
            "notes": "Coding/agent-focused refresh of 4.5. Later GLM-4.7/GLM-5 references exist in 2026 sources but are only partly verified."
        },
        {
            "model": "Kimi K2 / K2.6",
            "developer": "Moonshot AI",
            "sizes": "1T-A32B MoE",
            "tool_tier": "excellent",
            "native_format": "yes (native tool-call; OpenAI-compatible)",
            "grammar": "yes",
            "parallel": "yes (multi-agent swarm variants)",
            "benchmark": "tau2-Telecom 96% (K2.6); SOTA on Humanity's Last Exam / BrowseComp with tools (Moonshot 2026). Stable across 200-300 sequential calls",
            "min_size": "1T MoE (32B active) — hosting is the barrier, needs a serious multi-GPU rig",
            "license": "Modified MIT",
            "commercial": "permissive",
            "best_use": "Long-horizon agentic coding + deep tool loops (200-300 calls) where you can host a 1T MoE",
            "grade": "B",
            "notes": "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."
        },
        {
            "model": "DeepSeek-V3.2",
            "developer": "DeepSeek-AI",
            "sizes": "671B-A37B MoE",
            "tool_tier": "excellent",
            "native_format": "yes (revised template; tool calls inside thinking mode)",
            "grammar": "yes (server strict mode + xgrammar)",
            "parallel": "unverified",
            "benchmark": "tau2-bench: Airline 63.8, Retail 81.1, Telecom 96.2 (arXiv 2512.02556, Dec 2025)",
            "min_size": "671B only — server-class multi-GPU",
            "license": "MIT (verify per download)",
            "commercial": "permissive",
            "best_use": "Reasoning-heavy multi-turn agents interleaving thinking + tool calls",
            "grade": "A",
            "notes": "Best-documented DeepSeek agent numbers. Watch inference-engine compat for thinking-mode tool calls — template changed vs V3.1."
        },
        {
            "model": "Devstral Small 2 / Devstral 2",
            "developer": "Mistral (w/ All Hands AI)",
            "sizes": "Small 2 24B; Devstral 2 123B",
            "tool_tier": "excellent (coding agents)",
            "native_format": "yes ([TOOL_CALLS], trained on agentic traces)",
            "grammar": "yes",
            "parallel": "yes (agent-loop)",
            "benchmark": "SWE-bench Verified: Devstral 2 (123B) 72.2%, Small 2 (24B) 68.0% (2025-12-09) — among best open-weight SWE-bench",
            "min_size": "Small 2 (24B) runs on one RTX 4090 / 32GB Mac, still 68% SWE-bench",
            "license": "Apache-2.0 (Small; verify 123B card)",
            "commercial": "permissive",
            "best_use": "THE open-weight choice for self-hosted coding agents (repo-level bug-fix / PR automation)",
            "grade": "A",
            "notes": "256K context; ships open-source Mistral Vibe CLI. Strongest agentic-coding story in the Mistral camp."
        },
        {
            "model": "Mistral Small 3.x / 4 (24B)",
            "developer": "Mistral AI",
            "sizes": "24B (3.1/3.2 add vision)",
            "tool_tier": "good (3.2 improved FC)",
            "native_format": "yes ([TOOL_CALLS], needs mistral parser)",
            "grammar": "yes",
            "parallel": "yes",
            "benchmark": "no clean public BFCL v3/v4 entry (2026-07-18); vendor 'improved FC' notes only — unverified",
            "min_size": "24B is the family floor, reliable for tools",
            "license": "Apache-2.0",
            "commercial": "permissive",
            "best_use": "Best single-GPU (24B) open agent worker: native format + permissive license",
            "grade": "B",
            "notes": "Mistral 3 (Large 3 + Ministral 3) all Apache-2.0 too — replaces the non-commercial Large 2 for blocked self-hosters."
        },
        {
            "model": "Hermes 4 (14B / 70B / 405B)",
            "developer": "Nous Research",
            "sizes": "14B (Qwen3 base); 70B/405B (Llama-3.1 base)",
            "tool_tier": "good (14B) -> excellent (70B/405B)",
            "native_format": "yes — ORIGIN of the ecosystem-standard <tool_call>{json}</tool_call> format",
            "grammar": "yes (first-class JSON mode)",
            "parallel": "yes (post-reasoning in H4)",
            "benchmark": "Hermes 4 released 2025-08-26, SOTA-among-open on Nous suite; no canonical dated BFCL/tau row (unverified)",
            "min_size": "14B — Apache-2.0, no MAU cap (Qwen3 base)",
            "license": "MIXED: 14B Apache-2.0; 70B/405B Llama Community License",
            "commercial": "permissive (14B) / restricted (70B+)",
            "best_use": "Neutral-aligned, low-refusal local tool agent; the 14B is the best small commercial pick here",
            "grade": "A",
            "notes": "The Hermes tool format is now a de-facto standard reused across Qwen/others. Hermes 4-14B / Apache-2.0 is a standout."
        },
        {
            "model": "IBM Granite 4.x",
            "developer": "IBM",
            "sizes": "4.1 up to ~30B, 512K ctx; 4.0 H-Small 32B-A9B, H-Tiny 7B, Micro 3B, Nano 350M/1B",
            "tool_tier": "good (8B) -> excellent (Small/30B)",
            "native_format": "yes (Granite tool-turn template)",
            "grammar": "yes (Nano runs in-browser via WebGPU)",
            "parallel": "yes (robustness unverified at 2B)",
            "benchmark": "BFCL-v3 (2026): 4.1 8B 68.27%, 4.1 30B 73.68% (ahead of Gemma-3-27B-class)",
            "min_size": "~7-8B reliably; 3B Micro for edge; 350M-1B Nano for constrained edge tool calling",
            "license": "Apache-2.0 (signed, ISO-42001 certified)",
            "commercial": "permissive",
            "best_use": "Memory-efficient, commercially-clean enterprise/edge agent (Mamba hybrid = low KV-cache for long tool loops)",
            "grade": "A",
            "notes": "The safest COMMERCIAL self-host pick: permissive license + efficiency + dated BFCL numbers."
        },
        {
            "model": "Llama 3.3 70B",
            "developer": "Meta",
            "sizes": "70B only",
            "tool_tier": "good",
            "native_format": "yes (<|python_tag|> / ipython / JSON blocks; broadest runtime support)",
            "grammar": "yes (widest ecosystem)",
            "parallel": "partial/unverified",
            "benchmark": "BFCL v2 tool-use 77.3% (2024-12-06; top open at the time). Llama no longer on the LIVE BFCL v3 top table",
            "min_size": "70B (only size) — ~405B-class tool use at 70B footprint",
            "license": "Llama 3.3 Community License (>700M MAU cap)",
            "commercial": "restricted",
            "best_use": "Best-value self-hosted Llama agent brain",
            "grade": "B",
            "notes": "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."
        },
        {
            "model": "Phi-4-mini",
            "developer": "Microsoft",
            "sizes": "3.8B (+ multimodal ~5.6B)",
            "tool_tier": "good (for its size)",
            "native_format": "yes (<|tool|> wrapped JSON schema — the model that ADDED Phi FC)",
            "grammar": "yes (native + external)",
            "parallel": "unverified (slightly weaker raw mapping vs Qwen3-4B)",
            "benchmark": "BFCL v4 low-to-mid 80s (ertas third-party, May 2026) — no official Berkeley row",
            "min_size": "IT IS the small option (3.8B), fits an 8GB GPU",
            "license": "MIT",
            "commercial": "permissive",
            "best_use": "On-device MIT-licensed tool agent — the safest single small pick",
            "grade": "B",
            "notes": "Real native tool template + MIT (fewer license strings). Note the LARGER base Phi-4 (14B) LACKS this template — only the mini has it."
        },
        {
            "model": "Gemma 4 (E4B / 26B-A4B / 31B)",
            "developer": "Google",
            "sizes": "E2B/E4B; 26B-A4B MoE; 31B dense (2026-03-31)",
            "tool_tier": "good",
            "native_format": "yes — FIRST mainline Gemma with native tool tokens (cuts ~15-20% tokens vs JSON prompting)",
            "grammar": "yes (native + xgrammar/GBNF)",
            "parallel": "unverified",
            "benchmark": "BFCL v4: 26B-A4B ~89% non-live, E4B mid-to-high 80s — both THIRD-PARTY (unsloth/ertas), not official Berkeley",
            "min_size": "E4B (4B effective) smallest reliable general tool-caller",
            "license": "Apache-2.0 claimed on release page — UNVERIFIED (conflicts with Gemma's historical custom license; confirm on the weight download)",
            "commercial": "permissive (unverified)",
            "best_use": "On-device/edge agent needing native structured tool calls with a small footprint",
            "grade": "B",
            "notes": "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."
        },
        {
            "model": "Gemma 2 / 3",
            "developer": "Google",
            "sizes": "Gemma 2: 2B/9B/27B; Gemma 3: 1B/4B/12B/27B + 3n",
            "tool_tier": "poor (2/3) — no native tool support",
            "native_format": "NO — delimiter-based prompt engineering + text parsing only",
            "grammar": "external ESSENTIAL — GBNF/xgrammar/outlines are the ONLY way to get reliable structured output",
            "parallel": "no",
            "benchmark": "none official",
            "min_size": "27B, and only with heavy prompt scaffolding",
            "license": "Gemma Terms of Use (custom, non-OSI)",
            "commercial": "restricted",
            "best_use": "Avoid for tools; chat/summarization with grammar-constrained JSON if forced",
            "grade": "A",
            "notes": "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."
        },
        {
            "model": "Command-R / R+ / R7B / Command A",
            "developer": "Cohere",
            "sizes": "R 35B; R+ 104B; R7B 7B; Command A 111B",
            "tool_tier": "good (R/R7B) -> excellent (R+/Command A)",
            "native_format": "yes (Cohere multi-hop 'tool plan' schema)",
            "grammar": "yes (GGUF -> GBNF/outlines)",
            "parallel": "yes (headline feature)",
            "benchmark": "R+ beat GPT-4 Turbo on ToolTalk(Hard) + topped BFCL single-turn FC (Cohere eval 2024, ~84 BFCL secondary — exact unverified)",
            "min_size": "R7B (7B) smallest agent-capable, single 12-16GB GPU; good tool-ABSTENTION",
            "license": "CC-BY-NC-4.0 + AUP — NON-COMMERCIAL across the ENTIRE open lineup",
            "commercial": "non-commercial",
            "best_use": "On-prem RAG agent with grounded citations + multi-step/parallel tools (research/non-commercial only)",
            "grade": "B",
            "notes": "🔴 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."
        },
        {
            "model": "DeepSeek-R1 + distills",
            "developer": "DeepSeek-AI",
            "sizes": "R1 671B; distills 1.5B-70B (Qwen/Llama bases)",
            "tool_tier": "poor -> fair (community templates)",
            "native_format": "partial — base R1/distills shipped WITHOUT robust tool support",
            "grammar": "yes — often the ONLY way to get reliable JSON from R1",
            "parallel": "no/unverified",
            "benchmark": "none reliable — R1 targeted reasoning (AIME/MATH), not function-calling",
            "min_size": "14B distill smallest marginally usable for tools; <8B weak",
            "license": "MIT (R1 + all six distills)",
            "commercial": "permissive",
            "best_use": "Reasoning/planning LAYER feeding a separate tool-executor — not a direct tool-caller",
            "grade": "B",
            "notes": "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."
        }
    ],
    "context": {
        "reality_2026": "As of mid-2026 the open-weight tool-calling crown belongs to Chinese labs: Qwen owns the Berkeley Function-Calling Leaderboard (~10 of 13 BFCL-v4 slots; open Qwen3.5-397B-A17B leads at 0.729), while GLM-4.5/4.6 (MIT) and Kimi K2.6 lead the heavyweight AGENTIC (tau-bench/SWE-bench) tier. Llama and Mistral have slipped off the LIVE BFCL top table — their strong numbers are v1/v2-era (2024-25) — and Hermes contributed the de-facto tool-call FORMAT now reused across the ecosystem. Small models CAN drive agents, but only with a native trained-in tool template PLUS grammar-constrained decoding: reliability falls off a cliff below ~4B, and parameter count is a poor predictor under 8B.",
        "smallest_driver": "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.",
        "grammar_note": "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.",
        "freshness_warning": "🔴 This is a MONTHLY-moving leaderboard: every BFCL/tau-bench/SWE-bench score here is dated (mostly 2025-08 to 2026-07). Re-verify against gorilla.cs.berkeley.edu, model cards and tau-bench before trusting any figure after ~August 2026. BFCL v3 and v4 are DIFFERENT suites — scores are NOT cross-comparable. Sub-10B BFCL-v4 figures are often third-party (not the official Berkeley harness) and are flagged."
    }
}