Compare / head-to-head
Qwen3 235B-A22B Thinking 2507vsQwen3 4B Instruct 2507
Qwen3 235B-A22B Thinking 2507 leads 15 of 15 shared benchmarks. Both offer a 262K-token context window.
Benchmarks from cited public sources; pricing from official pages; status from official provider feeds.
Shared benchmarks
15 – 0
Qwen3 235B-A22B Thinking 2507 leads
Cheaper per token
—
list price, input + output
Larger context
Tie
both 262K tokens
Providers
Qwen
same provider
Qwen3 235B-A22B Thinking 2507
Qwen · released 2025-07-25
text
Qwen3 4B Instruct 2507
Qwen · released 2025-08-06
text
- Context
- 262K
- Max out
- —
- Input /1M
- —
- Output /1M
- —
- Cached /1M
- —
- Scores
- 18 · 7 core
Quality
Benchmark matrix
15 shared · 8 only Qwen3 235B-A22B Thinking 2507 · 3 only Qwen3 4B Instruct 2507| Benchmark | Qwen3 235B-A22B Thinking 2507 | Qwen3 4B Instruct 2507 | Δ | Edge |
|---|---|---|---|---|
| Reported by both · 15 | ||||
| GPQA Diamond | 80.1% ↗epoch run | 45.8% ↗epoch run | +34.3 pt | Qwen3 235B-A22B Thinking 2507 |
| MMLU-Pro | 84.4% ↗ | 69.6% ↗ | +14.8 pt | Qwen3 235B-A22B Thinking 2507 |
| AIME25 | 92.3% ↗ | 47.4% ↗ | +44.9 pt | Qwen3 235B-A22B Thinking 2507 |
| Arena-Hard v2 | 79.7% ↗ | 43.4% ↗ | +36.3 pt | Qwen3 235B-A22B Thinking 2507 |
| BFCL-v3 | 71.9% ↗ | 61.9% ↗ | +10 pt | Qwen3 235B-A22B Thinking 2507 |
| Chess Puzzles (Epoch AI run) | 12% ↗epoch run | 4% ↗epoch run | +8 pt | Qwen3 235B-A22B Thinking 2507 |
| GPQA | 81.1% ↗ | 62% ↗ | +19.1 pt | Qwen3 235B-A22B Thinking 2507 |
| HMMT25 | 83.9% ↗ | 31% ↗ | +52.9 pt | Qwen3 235B-A22B Thinking 2507 |
| IFEval | 87.8% ↗ | 83.4% ↗ | +4.4 pt | Qwen3 235B-A22B Thinking 2507 |
| LiveBench 20241125 | 78.4% ↗ | 63% ↗ | +15.4 pt | Qwen3 235B-A22B Thinking 2507 |
| LiveCodeBench v6 (25.02-25.05) | 74.1% ↗ | 35.1% ↗ | +39 pt | Qwen3 235B-A22B Thinking 2507 |
| MMLU-Redux | 93.8% ↗ | 84.2% ↗ | +9.6 pt | Qwen3 235B-A22B Thinking 2507 |
| OTIS Mock AIME 2024-2025 (Epoch AI run) | 86.7% ↗epoch run | 52.2% ↗epoch run | +34.5 pt | Qwen3 235B-A22B Thinking 2507 |
| SuperGPQA | 64.9% ↗ | 42.8% ↗ | +22.1 pt | Qwen3 235B-A22B Thinking 2507 |
| TAU2-Retail | 71.9% ↗ | 40.4% ↗ | +31.5 pt | Qwen3 235B-A22B Thinking 2507 |
| Only Qwen3 235B-A22B Thinking 2507 reports · 8 | ||||
| LMArena Elo | 1400 ↗ | not reported | — | — |
| HLE (text-only) | 18.2% ↗ | not reported | — | — |
| Mystery Game Puzzles (Epoch AI run) | 9% ↗epoch run | not reported | — | — |
| OJBench | 32.5% ↗ | not reported | — | — |
| SimpleQA Verified | 40.4% ↗epoch run | not reported | — | — |
| TAU2-Airline | 58% ↗ | not reported | — | — |
| Vending-Bench 2 (Andon Labs) | -11.34 ↗ | not reported | — | — |
| WeirdML (Håvard Tveit Ihle) | 41% ↗ | not reported | — | — |
| Only Qwen3 4B Instruct 2507 reports · 3 | ||||
| Aider-Polyglot | not reported | 12.9% ↗ | — | — |
| MultiPL-E | not reported | 76.8% ↗ | — | — |
| ZebraLogic | not reported | 80.2% ↗ | — | — |
Scores tagged "epoch" or "matharena" are independent runs, used only where the lab has not published its own; ⚠ marks rows MathArena flags as released after the competition. Higher is better on every row. Δ is Qwen3 235B-A22B Thinking 2507 minus Qwen3 4B Instruct 2507 in the benchmark's own unit. "Not reported" means the lab has not published that figure; it is not a zero. ↗ opens the source.
Specs & pricing
Side by side
Official pricing pages and model cards| Spec | Qwen3 235B-A22B Thinking 2507 | Qwen3 4B Instruct 2507 | Edge |
|---|---|---|---|
| Context window | 262K tokens | 262K tokens | Tie |
| Max output | — | — | — |
| Input price / 1M | $0.23 ↗ | — | — |
| Output price / 1M | $2.3 ↗ | — | — |
| Cached input / 1M | — | — | — |
| Input + output / 1M Lower is cheaper. List prices; batch, tool and regional fees excluded. | $2.53 | — | — |
| Modalities | text | text | Tie |
| Released | 2025-07-25 | 2025-08-06 | — |
| Cited benchmark scores | 23 | 18 | — |
Reliability
Qwen status
Live from /statusMore matchups
Qwen3 235B-A22B Thinking 2507 vs …
Models sharing the most benchmarksMore matchups
Qwen3 4B Instruct 2507 vs …
Models sharing the most benchmarksBuilt by Respan
Which one wins on your data?
Public benchmarks are a starting point. Run Qwen3 235B-A22B Thinking 2507 and Qwen3 4B Instruct 2507 on your own prompts with Respan evals, or route to either through one gateway key with automatic failover.