Datasets:
benchmark stringlengths 4 32 | metric stringlengths 4 40 | torki_2 float64 38 99.4 | items stringlengths 3 7 ⌀ | setting stringclasses 3
values | sample stringclasses 2
values | note stringlengths 23 100 ⌀ |
|---|---|---|---|---|---|---|
GPQA Diamond | accuracy, average of 4 runs | 89.65 | 198 x 4 | thinking on (Stellar/Cosmos default) | no | options shuffled per run; unanswered (ran out of room) counted wrong |
AIME 2025 | accuracy, average of 8 attempts | 96.25 | 30 x 8 | thinking on (Stellar/Cosmos default) | no | null |
AIME 2024 | accuracy, average of 8 attempts | 95.83 | 30 x 8 | thinking on (Stellar/Cosmos default) | no | null |
AIME 2026 | accuracy, average of 4 attempts | 95 | 30 x 4 | thinking on (Stellar/Cosmos default) | no | null |
MATH-500 | accuracy | 99.4 | 500 | thinking on (Stellar/Cosmos default) | no | null |
MMLU-Pro | accuracy, weighted to full subject mix | 88.07 | 788 | thinking on (Stellar/Cosmos default) | yes | fixed random 100 per subject for 6 subjects, 30-48 for 8; about +/-2 points |
Humanity's Last Exam (text-only) | accuracy, weighted to full subject mix | 38.01 | 400 | thinking on (Stellar/Cosmos default) | yes | first 400 of 2,158 text-only questions; official grading prompt run on Torki 2; about +/-4.8 points |
HumanEval+ | pass@1 (base + extended tests) | 94.51 | 164 | thinking on (Stellar/Cosmos default) | no | null |
MBPP+ | pass@1 (base + extended tests) | 83.86 | 378 | thinking on (Stellar/Cosmos default) | no | null |
LiveCodeBench | pass@1, v6 window 2025-02..2025-04 | 77.1 | 131 | thinking on (Stellar/Cosmos default) | no | 65,536-token answer cap |
LiveCodeBench | pass@1, v5 window 2024-08..2025-01 | 79.57 | 323 | thinking on (Stellar/Cosmos default) | no | null |
IFEval | prompt-level strict accuracy | 90.94 | 541 | thinking on (Stellar/Cosmos default) | no | null |
IFBench | prompt-level strict accuracy | 73.33 | 300 | thinking on (Stellar/Cosmos default) | no | null |
MMMU | accuracy | 84.67 | 900 | thinking on (Stellar/Cosmos default) | no | rule-based answer checking |
MathVista | accuracy | 89.1 | 1000 | thinking on (Stellar/Cosmos default) | no | rule-based answer checking |
ChartQA | relaxed accuracy | 85.24 | 2500 | thinking off | no | null |
DocVQA | ANLS | 94.57 | 5349 | thinking off | no | null |
MGSM | accuracy, reference scorer | 85.96 | 2750 | thinking on (Stellar/Cosmos default) | no | 93.35 if spelling variants of the Telugu/Bengali word for 'Answer' are accepted |
RULER (32K-128K) | average of 13 tasks x 3 lengths | 94.14 | 1950 | thinking off | yes | 50 samples per task per length (reference uses 500) |
MILU (11 Indian languages) | accuracy | 85.09 | 550 | thinking on (Stellar/Cosmos default) | yes | fixed random 50 per language; about +/-3 points |
MMMLU (14 languages) | accuracy | 75.25 | 2800 | thinking off, letter only | yes | not comparable with thinking-on numbers from other labs |
BFCL v4 | accuracy, non-live, official AST average | 85.23 | null | thinking on (Stellar/Cosmos default) | no | function-calling mode; BFCL's own scorer; web-search and memory sections not run, so no BFCL overall |
BFCL v4 | accuracy, live (real-world calls) | 75.35 | null | thinking on (Stellar/Cosmos default) | no | function-calling mode; BFCL's own scorer; web-search and memory sections not run, so no BFCL overall |
BFCL v4 | accuracy, multi-turn | 50.75 | null | thinking on (Stellar/Cosmos default) | no | function-calling mode; BFCL's own scorer; web-search and memory sections not run, so no BFCL overall |
BFCL v4 | accuracy, irrelevance detection | 74.64 | null | thinking on (Stellar/Cosmos default) | no | function-calling mode; BFCL's own scorer; web-search and memory sections not run, so no BFCL overall |
Torki 2 — Evaluation Results
Torki 2 is the model behind Torki, the AI assistant built by TorkeHub. This dataset holds its results on 22 public benchmarks, measured between 30 September and 1 October 2026 on the same production model that Torki users get.
Launch page: torkehub.com/torki-2 · Interactive results: TorkeHub/torki-2-benchmarks
Results
| Area | Benchmark | Metric | Torki 2 | Items | Setting |
|---|---|---|---|---|---|
| Reasoning and knowledge | GPQA Diamond | accuracy, average of 4 runs | 89.7 | 198 x 4 | thinking on (Stellar/Cosmos default) |
| Humanity's Last Exam (text-only) ¹ | accuracy, weighted to full subject mix | 38.0 | 400 | thinking on (Stellar/Cosmos default) | |
| MMLU-Pro ¹ | accuracy, weighted to full subject mix | 88.1 | 788 | thinking on (Stellar/Cosmos default) | |
| Mathematics | AIME 2026 | accuracy, average of 4 attempts | 95.0 | 30 x 4 | thinking on (Stellar/Cosmos default) |
| AIME 2025 | accuracy, average of 8 attempts | 96.3 | 30 x 8 | thinking on (Stellar/Cosmos default) | |
| AIME 2024 | accuracy, average of 8 attempts | 95.8 | 30 x 8 | thinking on (Stellar/Cosmos default) | |
| MATH-500 | accuracy | 99.4 | 500 | thinking on (Stellar/Cosmos default) | |
| Coding | LiveCodeBench | pass@1, v6 window 2025-02..2025-04 | 77.1 | 131 | thinking on (Stellar/Cosmos default) |
| LiveCodeBench | pass@1, v5 window 2024-08..2025-01 | 79.6 | 323 | thinking on (Stellar/Cosmos default) | |
| HumanEval+ | pass@1 (base + extended tests) | 94.5 | 164 | thinking on (Stellar/Cosmos default) | |
| MBPP+ | pass@1 (base + extended tests) | 83.9 | 378 | thinking on (Stellar/Cosmos default) | |
| Instruction following | IFEval | prompt-level strict accuracy | 90.9 | 541 | thinking on (Stellar/Cosmos default) |
| IFBench | prompt-level strict accuracy | 73.3 | 300 | thinking on (Stellar/Cosmos default) | |
| Vision | MMMU | accuracy | 84.7 | 900 | thinking on (Stellar/Cosmos default) |
| MathVista | accuracy | 89.1 | 1000 | thinking on (Stellar/Cosmos default) | |
| DocVQA | ANLS | 94.6 | 5349 | thinking off | |
| ChartQA | relaxed accuracy | 85.2 | 2500 | thinking off | |
| Languages | MILU (11 Indian languages) ¹ | accuracy | 85.1 | 550 | thinking on (Stellar/Cosmos default) |
| MGSM | accuracy, reference scorer | 86.0 | 2750 | thinking on (Stellar/Cosmos default) | |
| MMMLU (14 languages) ¹ | accuracy | 75.3 | 2800 | thinking off, letter only | |
| Long context | RULER (32K-128K) ¹ | average of 13 tasks x 3 lengths | 94.1 | 1950 | thinking off |
| Tool use | BFCL v4 | accuracy, non-live, official AST average | 85.2 | thinking on (Stellar/Cosmos default) | |
| BFCL v4 | accuracy, live (real-world calls) | 75.4 | thinking on (Stellar/Cosmos default) | ||
| BFCL v4 | accuracy, multi-turn | 50.8 | thinking on (Stellar/Cosmos default) | ||
| BFCL v4 | accuracy, irrelevance detection | 74.6 | thinking on (Stellar/Cosmos default) |
¹ Fixed random sample, not the full benchmark, because of the time available; margins of error are in
data/summary.csv. Unanswered questions, including answers that ran out of room, count as wrong everywhere.
Compared with other models
Other models' numbers are their makers' own published figures, or an independent measurement (marked †: Artificial Analysis, Epoch AI or Vals AI) where the maker publishes none, at the highest stated reasoning setting and without tools. Every number's source link is in data/comparisons.csv. The best number in each row is in bold, whoever holds it; a dash means no comparable number is published.
Against Anthropic and OpenAI
| Benchmark | Torki 2 | Claude Opus 5.5 | Claude Fable 5.1 | Claude Sonnet 5.5 | Claude Haiku 4.5 | GPT-6 Astra | GPT-6.1 Sol | GPT-6 Luna | gpt-oss-120b |
|---|---|---|---|---|---|---|---|---|---|
| GPQA Diamond | 89.7 | 90.6† | 93.7† | 95.6† | 73.0 | 96.0 | 95.4† | 90.5† | 80.1 |
| Humanity's Last Exam (text-only) | 38.0 | 61.4† | 59.1† | 55.0† | 10.4† | 54.7† | 52.9† | 38.5† | 19.6† |
| MMLU-Pro | 88.1 | — | 92.4† | — | 78.7† | — | — | — | 79.2† |
| AIME 2025 | 96.3 | — | — | — | 80.7 | — | — | — | 92.5 |
| MATH-500 | 99.4 | — | — | — | — | — | — | — | 94.8† |
| LiveCodeBench v6 | 77.1 | — | 90.5† | — | 61.5† | — | — | — | 83.2† |
| IFBench | 73.3 | — | — | — | 54.3† | — | — | — | 69.0† |
| MMMU | 84.7 | — | — | — | 73.2 | — | — | — | — |
| MILU | 85.1 | 93.1 | 93.0 | 91.6 | — | — | — | — | — |
| MGSM | 86.0 | — | — | — | 92.2† | — | — | — | 92.0† |
Against Google, Meta and xAI
| Benchmark | Torki 2 | Gemini 3.1 Pro | Gemini 3 Flash | Gemini 3.1 Flash-Lite | Gemma 4 31B | Muse Spark 1.3 | Muse Glimmer-30B | Llama 4 Maverick | Grok 4.7 | Grok 4.3 |
|---|---|---|---|---|---|---|---|---|---|---|
| GPQA Diamond | 89.7 | 94.3 | 90.4 | 86.9 | 84.3 | 93.5† | 83.5 | 69.8 | — | 90.1† |
| AIME 2025 | 96.3 | — | 95.2 | — | — | — | — | — | — | — |
| LiveCodeBench v6 | 77.1 | — | — | 72.0 | 80.0 | — | — | — | — | — |
| MMLU-Pro | 88.1 | — | — | — | 85.2 | — | — | 80.5 | — | — |
| AIME 2026 | 95.0 | — | — | — | 89.2 | — | 94.7 | — | — | — |
| Humanity's Last Exam (text-only) | 38.0 | — | — | — | — | 48.7† | 22.0 | — | 43.1† | 37.2† |
| IFBench | 73.3 | — | — | — | — | — | 77.0 | — | — | 81.3† |
| MMMU | 84.7 | — | — | — | — | — | — | 73.4 | — | — |
| MathVista | 89.1 | — | — | — | — | — | — | 73.7 | — | — |
| ChartQA | 85.2 | — | — | — | — | — | — | 90.0 | — | — |
| DocVQA | 94.6 | — | — | — | — | — | — | 94.4 | — | — |
| MGSM | 86.0 | — | — | — | — | — | — | 92.3 | — | — |
Against China and India
| Benchmark | Torki 2 | DeepSeek-V4-Pro | DeepSeek-V4.1-Flash | DeepSeek-V4-Pro (preview) | Kimi K3 | GLM-5.3 | MiniMax-M3 | Seed 2.0 Pro | Hunyuan Hy4 (preview) | ERNIE 5.1 | Step 5 (preview) | Sarvam-105B | Sarvam-30B |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GPQA Diamond | 89.7 | 92.4 | 90.9 | — | 93.5 | 91.7† | 92.9† | 88.9 | 92.3 | 91.0 | 93.5 | 78.7 | 66.5 |
| Humanity's Last Exam (text-only) | 38.0 | 42.7 | 39.1 | — | 46.9† | 42.3† | 39.0† | 32.4 | 43.4 | — | 46.5† | — | — |
| AIME 2026 | 95.0 | — | 100.0 | — | — | — | — | 94.2 | — | 95.0 | — | — | — |
| MMLU-Pro | 88.1 | — | — | 87.5 | 88.0† | 86.8† | 84.2† | 87.0 | — | 84.3 | — | 81.7 | 80.0 |
| LiveCodeBench v6 | 77.1 | — | — | 93.5 | — | — | — | 87.8 | — | — | — | 71.7 | 70.0 |
| IFBench | 73.3 | — | — | — | — | — | 82.9† | — | — | — | — | — | — |
| AIME 2025 | 96.3 | — | — | — | — | — | — | 98.3 | — | — | — | 88.3 | 80.0 |
| MMMU | 84.7 | — | — | — | — | — | — | 85.4 | — | — | — | — | — |
| MathVista | 89.1 | — | — | — | — | — | — | 89.8 | — | — | — | — | — |
| MATH-500 | 99.4 | — | — | — | — | — | — | — | — | — | — | 98.6 | 97.0 |
| IFEval | 90.9 | — | — | — | — | — | — | — | — | — | — | 84.8 | — |
| MILU | 85.1 | — | — | — | — | — | — | — | — | — | — | — | 76.8 |
Torki 2's Humanity's Last Exam, MMLU-Pro and MILU numbers are fixed samples; MGSM is 93.4 if Telugu and Bengali spelling variants of 'Answer' are accepted. Claude results include Anthropic's safeguard fallback to another Claude model, as Anthropic's system cards describe. Llama 4 Maverick has no thinking mode.
How we measured
- Same model as the product. Every request went to the production Torki 2 model with thinking on at the depth Torki's Stellar and Cosmos modes use, temperature 0.7, and an answer limit of 65,536 tokens (Torki's thinking modes allow more). Exceptions are marked "thinking off" in the table.
- No tools. No web search, code execution or calculator, on any benchmark.
- Reference prompts and scorers. Prompts and answer checks follow each benchmark's public reference implementation; AIME, GPQA and MATH-500 follow OpenAI's simple-evals, LiveCodeBench, EvalPlus, IFEval, IFBench, BFCL and RULER use their own scorers.
- Repeats. GPQA Diamond is the average of 4 runs with the options reshuffled each time; AIME 2024 and 2025 are averaged over 8 attempts per problem, AIME 2026 over 4.
- Every score was checked by hand from the raw responses before publishing. Scoring bugs we found were fixed by re-scoring the saved responses; none of the fixes re-ran the model.
Notes on individual benchmarks:
- Humanity's Last Exam: the first 400 of the 2,158 text-only questions in published order, weighted to the full subject mix. The benchmark's reference grader is a commercial model from another lab, so we ran the official grading prompt on Torki 2 itself; 16 of 16 grader decisions checked by hand matched the official answers. A strict rule-based reading that only accepts an exact or mathematically equal answer gives 27.0%, the floor.
- MMLU-Pro: 788 of 12,032 questions (fixed random 100 per subject for six subjects, 30–48 for eight), weighted to the full subject mix.
- MGSM: OpenAI's reference scorer needs the exact word for "Answer" in each language. In Telugu, and occasionally Bengali, Torki 2 wrote a common spelling variant, which the scorer rejects; accepting it gives 93.35%.
- RULER: 13 tasks at 32K, 64K and 128K tokens, 50 samples per task and length (the reference uses 500), thinking off.
- MMMLU: thinking off and a one-letter answer, on 200 questions per language. Not comparable with thinking-on numbers from other labs.
- BFCL v4: function-calling mode with BFCL's own scorer. The web-search and memory sections were not run, so no BFCL overall score is given.
Not run: SWE-bench Verified and Terminal-Bench (they need sandbox infrastructure we have not opened up), and τ²-bench (its simulated customer is a commercial model from another lab).
Files
| File | Contents |
|---|---|
data/summary.csv |
One row per benchmark and metric: score, items, setting, whether it is a sample, notes |
data/per_item.csv |
One row per question and attempt: benchmark, item id, repeat, category, correct, score, finish reason, answer length in tokens |
data/comparisons.csv |
Other models' numbers used for comparison, each with its source link and whether it is official |
Question text, reference answers and model responses are not included. Several benchmarks, including GPQA and Humanity's Last Exam, ask that their questions not be republished online.
Contact
support@torkehub.com · © 2026 TorkeHub Inc.
- Downloads last month
- 64