--- title: ALL Bench Leaderboard emoji: π colorFrom: indigo colorTo: purple sdk: static pinned: false license: apache-2.0 short_description: ALL Bench Leaderboard models: # ===== LLM - Open Source ===== # --- Flagship / Large --- - Qwen/Qwen3.5-397B-A17B # π μΆκ° β μ€νμμ€ μ΅λ νλκ·Έμ (17B/397B MoE) - Qwen/Qwen3.5-122B-A10B - nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 - openai/gpt-oss-120b # π μΆκ° β Nemotron μ§μ κ²½μ (5.1B/116.8B MoE) - upstage/Solar-Open-100B - moonshotai/Kimi-K2 # π μΆκ° β FINAL Bench 1μ (68.71), Union Eval 62.4 - Qwen/Qwen3-Next-80B-A3B-Thinking # --- Medium --- - Qwen/Qwen3.5-35B-A3B - Qwen/Qwen3.5-27B - mistralai/Mistral-Large-Instruct-2411 # π μΆκ° β μ λ½ λν μ€ν λͺ¨λΈ - LGAI-EXAONE/EXAONE-Deep-32B # π μΆκ° β νκ΅ μλ²λ¦° AI (K-EXAONE) - MiniMaxAI/MiniMax-M2.5 - openai/gpt-oss-20b # π μΆκ° β AIME 98.7% μν MoE μ±νΌμΈ # --- Small / Edge --- - Qwen/Qwen3.5-9B - Qwen/Qwen3.5-4B - microsoft/phi-4 - deepseek-ai/DeepSeek-V3 - deepseek-ai/DeepSeek-R1 - zai-org/GLM-5 - meta-llama/Llama-4-Scout-17B-16E-Instruct - meta-llama/Llama-4-Maverick-17B-128E-Instruct - K-intelligence/Midm-2.0-Base-Instruct - Nanbeige/Nanbeige4.1-3B - stepfun-ai/Step-3.5-Flash # ===== VLM - Open Source ===== - OpenGVLab/InternVL3.5-241B-A28B # π μΆκ° β VLM μ€νμμ€ 1μ (MMMU 77.7) - OpenGVLab/InternVL3-78B - Qwen/Qwen2.5-VL-72B-Instruct - Qwen/Qwen3-VL-30B-A3B - moonshotai/Kimi-VL-A3B-Thinking # π μΆκ° β VLM κ²½λ μΆλ‘ λν # ===== Image Generation ===== - black-forest-labs/FLUX.1-dev - stabilityai/stable-diffusion-3.5-large # ===== Video Generation ===== - Lightricks/LTX-Video # ===== Music Generation ===== - facebook/musicgen-large - facebook/jasco-chords-drums-melody-1B datasets: - FINAL-Bench/Metacognitive - FINAL-Bench/ALL-Bench-Leaderboard - ginigen-ai/smol-worldcup --- # π ALL Bench Leaderboard 2026 **The only AI benchmark dataset covering LLM Β· VLM Β· Agent Β· Image Β· Video Β· Music in a single unified file.**
  ## Dataset Summary ALL Bench Leaderboard aggregates and cross-verifies benchmark scores for **90+ AI models** across 6 modalities. Every numerical score is tagged with a confidence level (`cross-verified`, `single-source`, or `self-reported`) and its original source. The dataset is designed for researchers, developers, and decision-makers who need a trustworthy, unified view of the AI model landscape. | Category | Models | Benchmarks | Description | |----------|--------|------------|-------------| | **LLM** | 41 | 32 fields | MMLU-Pro, GPQA, AIME, HLE, ARC-AGI-2, Metacog, SWE-Pro, IFEval, LCB, **Union Eval**, etc. | | **VLM Flagship** | 11 | 10 fields | MMMU, MMMU-Pro, MathVista, AI2D, OCRBench, MMStar, HallusionBench, etc. | | **Agent** | 10 | 8 fields | OSWorld, ΟΒ²-bench, BrowseComp, Terminal-Bench 2.0, GDPval-AA, SWE-Pro | | **Image Gen** | 10 | 7 fields | Photo realism, text rendering, instruction following, style, aesthetics | | **Video Gen** | 10 | 7 fields | Quality, motion, consistency, text rendering, duration, resolution | | **Music Gen** | 8 | 6 fields | Quality, vocals, instrumental, lyrics, duration |    ## What's New β v2.2.1 ### π Union Eval β NEW **ALL Bench's proprietary integrated benchmark.** Fuses the discriminative core of 10 existing benchmarks (GPQA, AIME, HLE, MMLU-Pro, IFEval, LiveCodeBench, BFCL, ARC-AGI, SWE, FINAL Bench) into a single 1000-question pool with a season-based rotation system. **Key features:** - **100% JSON auto-graded** β every question requires mandatory JSON output with verifiable fields. Zero keyword matching. - **Fuzzy JSON matching** β tolerates key name variants, fraction formats, text fallback when JSON parsing fails. - **Season rotation** β 70% new questions each season, 30% anchor questions for cross-season IRT calibration. - **8 rounds of empirical testing** β v2 (82.4%) β v3 (82.0%) β Final (79.5%) β S2 (81.8%) β S3 (75.0%) β Fuzzy (69.9/69.3%). **Key discovery:** *"The bottleneck in benchmarking is not question difficulty β it's grading methodology."* **Empirically confirmed LLM weakness map:** - π΄ Poetry + code cross-constraints: 18-28% - π΄ Complex JSON structure (10+ constraints): 0% - π΄ Pure series computation (Ξ£kΒ²/3α΅): 0% - π’ Metacognitive reasoning (Bayes, proof errors): 95% - π’ Revised science detection: 86% **Current scores (S3, 20Q sample, Fuzzy JSON):** | Model | Union Eval | |-------|-----------| | Claude Sonnet 4.6 | **69.9** | | Claude Opus 4.6 | **69.3** | ### Other v2.2 changes - Fair Coverage Correction: composite scoring ^0.5 β ^0.7 - +7 FINAL Bench scores (15 total) - Columns sorted by fill rate - Model Card popup (click model name) Β· FINAL Bench detail popup (click Metacog score) - π₯ Heatmap, π° Price vs Performance scatter tools ## Live Leaderboard π **[https://huggingface.co/spaces/FINAL-Bench/all-bench-leaderboard](https://huggingface.co/spaces/FINAL-Bench/all-bench-leaderboard)** Interactive features: composite ranking, dark mode, advanced search (`GPQA > 90 open`, `price < 1`), Model Finder, Head-to-Head comparison, Trust Map heatmap, Bar Race animation, Model Card popup, FINAL Bench detail popup, and downloadable Intelligence Report (PDF/DOCX). ## Data Structure ``` data/ βββ llm.jsonl # 41 LLMs Γ 32 fields (incl. unionEval β NEW) βββ vlm_flagship.jsonl # 11 flagship VLMs Γ 10 benchmarks βββ agent.jsonl # 10 agent models Γ 8 benchmarks βββ image.jsonl # 10 image gen models Γ S/A/B/C ratings βββ video.jsonl # 10 video gen models Γ S/A/B/C ratings βββ music.jsonl # 8 music gen models Γ S/A/B/C ratings ``` ## LLM Field Schema | Field | Type | Description | |-------|------|-------------| | `name` | string | Model name | | `provider` | string | Organization | | `type` | string | `open` or `closed` | | `group` | string | `flagship`, `open`, `korean`, etc. | | `released` | string | Release date (YYYY.MM) | | `mmluPro` | float \| null | MMLU-Pro score (%) | | `gpqa` | float \| null | GPQA Diamond (%) | | `aime` | float \| null | AIME 2025 (%) | | `hle` | float \| null | Humanity's Last Exam (%) | | `arcAgi2` | float \| null | ARC-AGI-2 (%) | | `metacog` | float \| null | FINAL Bench Metacognitive score | | `swePro` | float \| null | SWE-bench Pro (%) | | `bfcl` | float \| null | Berkeley Function Calling (%) | | `ifeval` | float \| null | IFEval instruction following (%) | | `lcb` | float \| null | LiveCodeBench (%) | | `sweV` | float \| null | SWE-bench Verified (%) β deprecated | | `mmmlu` | float \| null | Multilingual MMLU (%) | | `termBench` | float \| null | Terminal-Bench 2.0 (%) | | `sciCode` | float \| null | SciCode (%) | | `unionEval` | float \| null | **β NEW** Union Eval S3 β ALL Bench integrated benchmark (100% JSON auto-graded) | | `priceIn` / `priceOut` | float \| null | USD per 1M tokens | | `elo` | int \| null | Arena Elo rating | | `license` | string | `Prop`, `Apache2`, `MIT`, `Open`, etc. |    ## Composite Score ``` Score = Avg(confirmed benchmarks) Γ (N/10)^0.7 ``` 10 core benchmarks across the **5-Axis Intelligence Framework**: Knowledge Β· Expert Reasoning Β· Abstract Reasoning Β· Metacognition Β· Execution. **v2.2 change:** Exponent adjusted from 0.5 to 0.7 for fairer coverage weighting. Models with 7/10 benchmarks receive Γ0.79 (was Γ0.84), while 4/10 receives Γ0.53 (was Γ0.63). ## Confidence System Each benchmark score in the `confidence` object is tagged: | Level | Badge | Meaning | |-------|-------|---------| | `cross-verified` | ββ | Confirmed by 2+ independent sources | | `single-source` | β | One official or third-party source | | `self-reported` | ~ | Provider's own claim, unverified | Example: ```json "Claude Opus 4.6": { "gpqa": { "level": "cross-verified", "source": "Anthropic + Vellum + DataCamp" }, "arcAgi2": { "level": "cross-verified", "source": "Vellum + llm-stats + NxCode + DataCamp" }, "metacog": { "level": "single-source", "source": "FINAL Bench dataset" }, "unionEval": { "level": "single-source", "source": "Union Eval S3 β ALL Bench official" } } ``` ## Usage ```python from datasets import load_dataset # Load LLM data ds = load_dataset("FINAL-Bench/ALL-Bench-Leaderboard", "llm") df = ds["train"].to_pandas() # Top 5 LLMs by GPQA ranked = df.dropna(subset=["gpqa"]).sort_values("gpqa", ascending=False) for _, m in ranked.head(5).iterrows(): print(f"{m['name']:25s} GPQA={m['gpqa']}") # Union Eval scores union = df.dropna(subset=["unionEval"]).sort_values("unionEval", ascending=False) for _, m in union.iterrows(): print(f"{m['name']:25s} Union Eval={m['unionEval']}") ```    ## Union Eval β Integrated AI Assessment Union Eval is ALL Bench's proprietary benchmark designed to address three fundamental problems with existing AI evaluations: 1. **Contamination** β Public benchmarks leak into training data. Union Eval rotates 70% of questions each season. 2. **Single-axis measurement** β AIME tests only math, IFEval only instruction-following. Union Eval integrates arithmetic, poetry constraints, metacognition, coding, calibration, and myth detection. 3. **Score inflation via keyword matching** β Traditional rubric grading gives 100% to "well-written" answers even if content is wrong. Union Eval enforces mandatory JSON output with zero keyword matching. **Structure (S3 β 100 Questions from 1000 Pool):** | Category | Questions | Role | Expected Score | |----------|-----------|------|---------------| | Pure Arithmetic | 10 | Confirmed Killer #1 | 0-57% | | Poetry/Verse IFEval | 8 | Confirmed Killer #2 | 18-28% | | Structured Data IFEval | 7 | JSON/CSV verification | 0-70% | | FINAL Bench Metacognition | 20 | Core brand | 50-95% | | Union Complex Synthesis | 15 | Extreme multi-domain | 40-73% | | Revised Science / Myths | 5 | Calibration traps | 50-86% | | Code I/O, GPQA, HLE | 19 | Expert + execution | 50-100% | | BFCL Tool Use, Anchors | 16 | Cross-season calibration | varies | Note: The 100-question dataset is **not publicly released** to prevent contamination. Only scores are published. ## FINAL Bench β Metacognitive Benchmark FINAL Bench measures AI self-correction ability. Error Recovery (ER) explains 94.8% of metacognitive performance variance. 15 frontier models evaluated. - 𧬠[FINAL-Bench/Metacognitive Dataset](https://huggingface.co/datasets/FINAL-Bench/Metacognitive) - π [FINAL-Bench/Leaderboard](https://huggingface.co/spaces/FINAL-Bench/Leaderboard) ## Changelog | Version | Date | Changes | |---------|------|---------| | **v2.2.1** | 2026-03-10 | π **Union Eval β NEW** β integrated benchmark column (`unionEval` field). Claude Opus 4.6: 69.3 Β· Sonnet 4.6: 69.9 | | v2.2 | 2026-03-10 | Fair Coverage (^0.7), +7 Metacog scores, Model Cards, FINAL Bench popup, Heatmap, Price-Perf | | v2.1 | 2026-03-08 | Confidence badges, Intelligence Report, source tracking | | v2.0 | 2026-03-07 | All blanks filled, Korean AI data, 42 LLMs cross-verified | | v1.9 | 2026-03-05 | +3 LLMs, dark mode, mobile responsive | ## Citation ```bibtex @misc{allbench2026, title={ALL Bench Leaderboard 2026: Unified Multi-Modal AI Evaluation}, author={ALL Bench Team}, year={2026}, url={https://huggingface.co/spaces/FINAL-Bench/all-bench-leaderboard} } ``` --- `#AIBenchmark` `#LLMLeaderboard` `#GPT5` `#Claude` `#Gemini` `#ALLBench` `#FINALBench` `#Metacognition` `#UnionEval` `#VLM` `#AIAgent` `#MultiModal` `#HuggingFace` `#ARC-AGI` `#AIEvaluation` `#VIDRAFT.net`