LuminaBench
Tool use benchmark / Berkeley Gorilla team

Berkeley Function Calling Leaderboard v4

Lumina VerifiedLumina checked the published sources, versions, model labels, score extraction and conversion. External evaluations were run by their publishers; Lumina has not independently rerun or reproduced them.

Overall accuracy

Version BFCL-v4-f7cf735Verified 6 November 2026109 published resultsOfficial source

Berkeley Function Calling Leaderboard v4

LuminaBench
Show
20 of 84 models · best variant per model
03507001,0501,4001,211Claude-Opus-4-5-202511011,189Claude Sonnet 4.5 (20250929)1,185Gemini 3 Pro Preview1,185GLM-4.6 (FC thinking)1,170Grok 4.1 Fast (Reasoning)1,166Claude Haiku 4.5 (20251001)1,137o3-2025-04-161,136Grok 4 (0709)1,116Moonshotai-Kimi-K2-Instruct1,112Grok-4-1-fast-non-reasoning1,106Command A Reasoning1,104DeepSeek-V3.2-Exp (Prompt + Thinking)1,102Gemini 2.5 Flash1,100gpt-5.2-2025-12-111,098gpt-5-mini-2025-08-071,094xLAM-2-32b-fc-r1,091DeepSeek-V3.2-Exp1,090GPT-4.1-2025-04-141,087o4-mini-2025-04-161,086xLAM-2-70b-fc-r

No comparable per-task resource measurements were published. Missing metrics are omitted from their graph; quality scores remain available.

Results table

Method & coverage

Original BFCL v4 overall function-calling accuracy; native FC and Prompt settings remain harnesses. Multi-turn, search and memory tasks are text/tool-based. Do not import component scores as independent overall indexes.

No comparable per-task resource measurements were published.

How the Lumina conversion works

1,000 + 75 × ((original score − 36.362976190476175) / 14.626367297075404). The result is rounded once. Scores may tie.

Frozen initial observed text-only cohort; one mean per exact publisher release/API, balancing repeated settings/routes within that release. Original score and protocol remain separate. No inferred variants or future refitting.

Calibration
bfcl-v4-z75-v1 · experimental
Reference mean
36.362976190476175
Population SD
14.626367297075404
Coverage
Standalone benchmark.
Identity, exclusions & attribution

Original measurements and exact source settings retained; unknown API, effort and harness versions remain unknown.

Text function signatures, arguments, conversations and tool results; no screenshot input.

Standalone text benchmark; no new combined index. Source identity/effort/protocol qualifications remain in Data and CSV.

All published reasoning settings, including Claude Max, are included by default. Display labels use Low and xHigh; source labels remain in the retained records.

Official source

Publisher terms. Scores and resource figures are attributed to the original publisher.