MiniMax M2.5 vs Qwen3.8 27B
Qwen3.8 27B leads MiniMax M2.5 on coding, 74.2 to 37.2 on the Vibe Coding Index, and MiniMax M2.5 costs 2.1x less per output token ($1.20 vs $2.55 per million). Updated October 22, 2026; 46 benchmarks have a result for both.
MiniMax M2.5 or Qwen3.8 27B: which to pick
Pick Qwen3.8 27B for coding: it scores 74.2 on the Vibe Coding Index against 37.2 for MiniMax M2.5.
Pick MiniMax M2.5 for output tokens 2.1x cheaper ($1.20 vs $2.55 per million) and a lower cost per solved task ($0.62 against $1.26).
- Shared results
- 46
- MiniMax M2.5 only
- 25
- Qwen3.8 27B only
- 39
- Like-for-like categories
- 6 of 9
Only benchmarks that tested both models count toward a pick; a result only one model has is shown but not compared.
Which one for your work
Each card is decided by one number both models carry. Where only one has it, the card says so and makes no call.
Coding
Vibe Coding Index, six public coding benchmarks
Qwen3.8 27B
Qwen3.8 27B scores 74.2 against 37.2.
3 and 4 benchmarks behind the scores
Intelligence
AA Intelligence Index
Qwen3.8 27B
Qwen3.8 27B leads, 33.7 against 22.8.
Same board, both measured
Agentic work
AA Agentic Index
No call
Only Qwen3.8 27B has a AA Agentic Index result, 45.8.
Directional only
Reasoning
GPQA Diamond
Qwen3.8 27B
Qwen3.8 27B leads, 88.9% against 82.1%.
Same board, both measured
Math
AIME
No call
Only MiniMax M2.5 has a AIME result, 88.8%.
Directional only
Knowledge
MMLU Pro
Qwen3.8 27B
Qwen3.8 27B leads, 84.3% against 80.1%.
Same board, both measured
Web design
Design Arena: websites
No call
Only MiniMax M2.5 has a Design Arena: websites result, 1222.
Directional only
Long documents
Listed context window
Qwen3.8 27B
Qwen3.8 27B takes 1M tokens against 205K.
Listed by OpenRouter
Everyday cost
1K input + 500 output tokens
MiniMax M2.5
MiniMax M2.5 $0.0009, Qwen3.8 27B $0.0017 per turn.
List prices
Cost per solved task
One SWE-bench Verified task, solved
MiniMax M2.5
MiniMax M2.5 $0.62, Qwen3.8 27B $1.26.
Published attempt cost and pass rate
At a glance
- Coding index
- MiniMax M2.5 37.2 from 3 benchmarks, Qwen3.8 27B 74.2 from 4 benchmarks.
- Widest category
- Coding, 24.1 points apart on average over 8 shared benchmarks.
- Widest single gap
- Vibe Code Bench (Vals AI): MiniMax M2.5 14.8% against Qwen3.8 27B 64.8%, 50.0 points apart on a 0 to 100 scale.
- Output price
- MiniMax M2.5 $1.20 against Qwen3.8 27B $2.55 per million output tokens, 2.1x apart.
- Context window
- MiniMax M2.5 takes 205K tokens, Qwen3.8 27B 1M.
Benchmark by benchmark
First the score: boards each model leads, per category, where both have a result. Then every board, a category to a fold; bars are each result on a 0 to 100 scale, so compare the two in a row, not one row with another.
| Category | MiniMax M2.5 | Bars | Qwen3.8 27B | Ties |
|---|---|---|---|---|
| Coding | 0 | 8 | ||
| Intelligence | 0 | 2 | ||
| Reasoning | 0 | 8 | ||
| Math | 0 | 3 | ||
| Knowledge | 0 | 7 | ||
| General | 0 | 18 | ||
| All shared boards | 0 | 46 |
CodingLike-for-likeSix of these boards feed the Vibe Coding Index08
| Benchmark | MiniMax M2.5 | Qwen3.8 27B | Leader |
|---|---|---|---|
| SWE-bench Verified | 75.8% | Not measured | Only MiniMax M2.5 measured |
| SWE-bench Multilingual | 68.3% | Not measured | Only MiniMax M2.5 measured |
| LiveBench coding | Not measured | 86.8% | Only Qwen3.8 27B measured |
| LMArena WebDev | 1386 | 1593 | Qwen3.8 27B |
| AA Coding Index | Not measured | 68.1 | Only Qwen3.8 27B measured |
| LiveCodeBench (Vals AI) | 79.2% | 84.0% | Qwen3.8 27B |
| SWE-bench Verified (Vals AI) | 74.2% | 86.0% | Qwen3.8 27B |
| Vibe Code Bench (Vals AI) | 14.8% | 64.8% | Qwen3.8 27B |
| Code migration (Vals AI) | Not measured | 14.2% | Only Qwen3.8 27B measured |
| ProgramBench (Vals AI) | Not measured | 0.0% | Only Qwen3.8 27B measured |
| IOI (Vals AI) | Not measured | 39.1% | Only Qwen3.8 27B measured |
| Terminal-Bench 2.1 (Vals AI) | Not measured | 58.4% | Only Qwen3.8 27B measured |
| SkillsBench (Vals AI) | Not measured | 38.1% | Only Qwen3.8 27B measured |
| CyberBench (Vals AI) | Not measured | 67.8% | Only Qwen3.8 27B measured |
| SciCode | Not measured | 46.6% | Only Qwen3.8 27B measured |
| Terminal-Bench Hard (Artificial Analysis run) | 34.8% | Not measured | Only MiniMax M2.5 measured |
| Terminal-Bench | 42.7% | Not measured | Only MiniMax M2.5 measured |
| ALE-Bench | 618.2 | Not measured | Only MiniMax M2.5 measured |
| LMArena Text: coding | 1380 | 1483 | Qwen3.8 27B |
| LMArena Text: software and IT | 1383 | 1478 | Qwen3.8 27B |
| LMArena WebDev: React | 1374 | 1602 | Qwen3.8 27B |
| LMArena WebDev: HTML | 1414 | 1570 | Qwen3.8 27B |
| LMArena WebDev: from an image | Not measured | 1577 | Only Qwen3.8 27B measured |
| SWE Atlas: Codebase QnA | 10.3% | Not measured | Only MiniMax M2.5 measured |
| SWE Atlas: Refactoring | 19.5% | Not measured | Only MiniMax M2.5 measured |
| SWE Atlas: Test Writing | 18.6% | Not measured | Only MiniMax M2.5 measured |
| Neither model has a result on the other 17 boards here | |||
IntelligenceLike-for-likePublic boards, shown as published, kept out of the index02
| Benchmark | MiniMax M2.5 | Qwen3.8 27B | Leader |
|---|---|---|---|
| AA Intelligence Index | 22.8 | 33.7 | Qwen3.8 27B |
| Epoch Capabilities Index | 146.6 | 151.7 | Qwen3.8 27B |
Agentic workDirectional onlyPublic boards, shown as published, kept out of the index12 boards, none shared
| Benchmark | MiniMax M2.5 | Qwen3.8 27B | Leader |
|---|---|---|---|
| AA Agentic Index | Not measured | 45.8 | Only Qwen3.8 27B measured |
| LiveBench research | Not measured | 59.4% | Only Qwen3.8 27B measured |
| LiveBench browser | Not measured | 32.0% | Only Qwen3.8 27B measured |
| LiveBench long agentic | Not measured | 27.9% | Only Qwen3.8 27B measured |
| Finance Agent (Vals AI) | Not measured | 48.5% | Only Qwen3.8 27B measured |
| Tax Agent Bench (Vals AI) | Not measured | 30.9% | Only Qwen3.8 27B measured |
| Legal Research Bench (Vals AI) | Not measured | 36.1% | Only Qwen3.8 27B measured |
| Excel Modeling Benchmark (Vals AI) | Not measured | 59.7% | Only Qwen3.8 27B measured |
| Terminal-Bench-Science (Vals AI) | Not measured | 0.0% | Only Qwen3.8 27B measured |
| τ²-Bench Telecom | 95.3% | Not measured | Only MiniMax M2.5 measured |
| APEX-Agents | Not measured | 47.5% | Only Qwen3.8 27B measured |
| Vending-Bench 2 | -$23 | Not measured | Only MiniMax M2.5 measured |
| Neither model has a result on the other 30 boards here | |||
ReasoningLike-for-likePublic boards, shown as published, kept out of the index08
| Benchmark | MiniMax M2.5 | Qwen3.8 27B | Leader |
|---|---|---|---|
| GPQA Diamond | 82.1% | 88.9% | Qwen3.8 27B |
| LiveBench reasoning | Not measured | 80.0% | Only Qwen3.8 27B measured |
| MMMU Pro (Vals AI) | Not measured | 83.9% | Only Qwen3.8 27B measured |
| GPQA Diamond (Artificial Analysis run) | 84.8% | 90.5% | Qwen3.8 27B |
| Humanity's Last Exam (Artificial Analysis run) | 20.5% | 33.9% | Qwen3.8 27B |
| ARC-AGI | 63.7% | 87.5% | Qwen3.8 27B |
| ARC-AGI-2 | 4.9% | 42.4% | Qwen3.8 27B |
| LMArena Text: hard prompts | 1372 | 1459 | Qwen3.8 27B |
| LMArena Text: expert | 1378 | 1479 | Qwen3.8 27B |
| LMArena Text: hard prompts in English | 1384 | 1461 | Qwen3.8 27B |
| LMArena Vision: diagrams | Not measured | 1285 | Only Qwen3.8 27B measured |
| LMArena Vision: homework | Not measured | 1295 | Only Qwen3.8 27B measured |
| Kagi LLM Benchmark | 55.2% | Not measured | Only MiniMax M2.5 measured |
| Neither model has a result on the other 10 boards here | |||
MathLike-for-likePublic boards, shown as published, kept out of the index03
| Benchmark | MiniMax M2.5 | Qwen3.8 27B | Leader |
|---|---|---|---|
| AIME | 88.8% | Not measured | Only MiniMax M2.5 measured |
| ProofBench | 4.0% | 16.0% | Qwen3.8 27B |
| LMArena Text: math | 1378 | 1455 | Qwen3.8 27B |
| LMArena Text: mathematical work | 1378 | 1458 | Qwen3.8 27B |
| Neither model has a result on the other 10 boards here | |||
KnowledgeLike-for-likePublic boards, shown as published, kept out of the index07
| Benchmark | MiniMax M2.5 | Qwen3.8 27B | Leader |
|---|---|---|---|
| MMLU Pro | 80.1% | 84.3% | Qwen3.8 27B |
| LegalBench (Vals AI) | 80.0% | 82.4% | Qwen3.8 27B |
| CaseLaw (Vals AI) | 53.5% | Not measured | Only MiniMax M2.5 measured |
| CorpFin (Vals AI) | 59.6% | Not measured | Only MiniMax M2.5 measured |
| TaxEval (Vals AI) | 68.2% | 70.8% | Qwen3.8 27B |
| MortgageTax (Vals AI) | Not measured | 64.9% | Only Qwen3.8 27B measured |
| MedQA (Vals AI) | 92.5% | Not measured | Only MiniMax M2.5 measured |
| MedCode (Vals AI) | Not measured | 28.7% | Only Qwen3.8 27B measured |
| LMArena Text: science | 1369 | 1467 | Qwen3.8 27B |
| LMArena Text: medicine and healthcare | 1357 | 1437 | Qwen3.8 27B |
| LMArena Text: legal and government | 1366 | 1447 | Qwen3.8 27B |
| LMArena Text: business and finance | 1356 | 1438 | Qwen3.8 27B |
| Neither model has a result on the other 8 boards here | |||
Web designDirectional onlyPublic boards, shown as published, kept out of the index8 boards, none shared
| Benchmark | MiniMax M2.5 | Qwen3.8 27B | Leader |
|---|---|---|---|
| Design Arena: websites | 1222 | Not measured | Only MiniMax M2.5 measured |
| Design Arena: UI components | 1170 | Not measured | Only MiniMax M2.5 measured |
| Design Arena: game dev | 1176 | Not measured | Only MiniMax M2.5 measured |
| Design Arena: data viz | 1172 | Not measured | Only MiniMax M2.5 measured |
| Design Arena: 3D | 1170 | Not measured | Only MiniMax M2.5 measured |
| Design Arena: all code categories | 1207 | Not measured | Only MiniMax M2.5 measured |
| Design Arena: SVG | 1150 | Not measured | Only MiniMax M2.5 measured |
| LiveBench design | Not measured | 75.9% | Only Qwen3.8 27B measured |
| Neither model has a result on the other 1 board here | |||
GeneralLike-for-likePublic boards, shown as published, kept out of the index018
| Benchmark | MiniMax M2.5 | Qwen3.8 27B | Leader |
|---|---|---|---|
| LiveBench data analysis | Not measured | 76.6% | Only Qwen3.8 27B measured |
| SAGE (Vals AI) | Not measured | 52.4% | Only Qwen3.8 27B measured |
| MedScribe (Vals AI) | Not measured | 83.8% | Only Qwen3.8 27B measured |
| IFBench | 71.6% | Not measured | Only MiniMax M2.5 measured |
| Long context reasoning (AA-LCR) | 73.3% | 82.0% | Qwen3.8 27B |
| DTBench | Not measured | 88.0% | Only Qwen3.8 27B measured |
| LMCA | Not measured | 41.4% | Only Qwen3.8 27B measured |
| CL-bench | 11.4% | Not measured | Only MiniMax M2.5 measured |
| Surface Evolver Bench | Not measured | 45.0% | Only Qwen3.8 27B measured |
| CL-bench Life | 6.3% | Not measured | Only MiniMax M2.5 measured |
| LMArena Text | 1359 | 1440 | Qwen3.8 27B |
| LMArena Text: instruction following | 1353 | 1436 | Qwen3.8 27B |
| LMArena Text: creative writing | 1331 | 1383 | Qwen3.8 27B |
| LMArena Vision | Not measured | 1269 | Only Qwen3.8 27B measured |
| LMArena Text: English | 1380 | 1448 | Qwen3.8 27B |
| LMArena Text: non-English | 1338 | 1428 | Qwen3.8 27B |
| LMArena Text: Chinese | 1393 | 1501 | Qwen3.8 27B |
| LMArena Text: German | 1361 | 1443 | Qwen3.8 27B |
| LMArena Text: Japanese | 1171 | 1387 | Qwen3.8 27B |
| LMArena Text: Korean | 1233 | 1384 | Qwen3.8 27B |
| LMArena Text: French | 1363 | 1465 | Qwen3.8 27B |
| LMArena Text: Spanish | 1353 | 1449 | Qwen3.8 27B |
| LMArena Text: Russian | 1359 | 1415 | Qwen3.8 27B |
| LMArena Text: Polish | 1334 | 1439 | Qwen3.8 27B |
| LMArena Text: multi-turn | 1364 | 1440 | Qwen3.8 27B |
| LMArena Text: longer queries | 1366 | 1448 | Qwen3.8 27B |
| LMArena Text: writing and language | 1345 | 1405 | Qwen3.8 27B |
| LMArena Text: entertainment and media | 1329 | 1393 | Qwen3.8 27B |
| LMArena Vision: English | Not measured | 1267 | Only Qwen3.8 27B measured |
| LMArena Vision: Chinese | Not measured | 1326 | Only Qwen3.8 27B measured |
| LMArena Vision: reading text | Not measured | 1281 | Only Qwen3.8 27B measured |
| LMArena Vision: humor | Not measured | 1235 | Only Qwen3.8 27B measured |
| LMArena Vision: creative writing | Not measured | 1237 | Only Qwen3.8 27B measured |
| Neither model has a result on the other 15 boards here | |||
Image generationNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
| Benchmark | MiniMax M2.5 | Qwen3.8 27B | Leader |
|---|---|---|---|
| Neither model has a result on 5 boards here | |||
What each workload costs
Three fixed token mixes at each model's listed price, so per-token rates become a bill.
Chat turn
1K input + 500 output tokens
MiniMax M2.5 costs less
Repository review
50K input + 3K output tokens
MiniMax M2.5 costs less
Agent loop
200K input + 10K output tokens
MiniMax M2.5 costs less
Listed OpenRouter rates, without cache or batch discounts. A workload that does not fit a model's context window would need splitting.
Specification differences
The leading value in a row is marked; a row with a missing value names no leader. Our two rows are marked in red.
| Specification | MiniMax M2.5 | Qwen3.8 27B |
|---|---|---|
| Maker | MiniMax | Qwen |
| Weights | Open weights | Open weights |
| Released | February 12, 2026 | August 14, 2026 |
| Context window | 205K | 1M |
| Input $/Mtok | $0.30 | $0.425 |
| Output $/Mtok | $1.20 | $2.55 |
| Coding index | 37.2 | 74.2 |
| Independent measurements behind it | 3 | 4 |
| Cost per solved task | $0.62 | $1.26 |
| Source | OpenRouter listing | OpenRouter listing |
Questions
Is MiniMax M2.5 better than Qwen3.8 27B for coding?
Pick Qwen3.8 27B for coding: it scores 74.2 on the Vibe Coding Index against 37.2 for MiniMax M2.5. Pick MiniMax M2.5 for output tokens 2.1x cheaper ($1.20 vs $2.55 per million) and a lower cost per solved task ($0.62 against $1.26).
Which is cheaper, MiniMax M2.5 or Qwen3.8 27B?
MiniMax M2.5 costs $0.30 per million input tokens and $1.20 per million output tokens; Qwen3.8 27B costs $0.425 and $2.55. Per solved SWE-bench Verified task, MiniMax M2.5 costs $0.62 and Qwen3.8 27B $1.26.
Which has the bigger context window, MiniMax M2.5 or Qwen3.8 27B?
Qwen3.8 27B, with 1M tokens against 205K.
How many benchmarks is the MiniMax M2.5 vs Qwen3.8 27B comparison based on?
46 benchmarks have a result for both models as of October 22, 2026. MiniMax M2.5's coding index rests on 3 independent benchmarks and Qwen3.8 27B's on 4. Where only one model has a result, the page says so instead of naming a winner.