Compare / head-to-head
Qwen3 235B-A22B Instruct 2507vsQwen3-VL 235B-A22B Thinking
Qwen3-VL 235B-A22B Thinking leads 5 of 8 shared benchmarks. Qwen3 235B-A22B Instruct 2507 is 3.8x cheaper per token. Qwen3 235B-A22B Instruct 2507 has the larger context window (262K tokens).
Benchmarks from cited public sources; pricing from official pages; status from official provider feeds.
Shared benchmarks
3 – 5
Qwen3-VL 235B-A22B Thinking leads
Cheaper per token
Qwen3 235B-A22B Instruct 2507
3.8x cheaper, input + output
Larger context
Qwen3 235B-A22B Instruct 2507
262K tokens
Providers
Qwen
same provider
Qwen3 235B-A22B Instruct 2507
Qwen · released 2025-07-21
text
Qwen3-VL 235B-A22B Thinking
Qwen · released 2025-09-23
textvisionvideo
Quality
Benchmark matrix
8 shared · 9 only Qwen3 235B-A22B Instruct 2507 · 9 only Qwen3-VL 235B-A22B Thinking| Benchmark | Qwen3 235B-A22B Instruct 2507 | Qwen3-VL 235B-A22B Thinking | Δ | Edge |
|---|---|---|---|---|
| Reported by both · 8 | ||||
| MMLU-Pro | 83% ↗ | 83.8% ↗ | -0.8 pt | Qwen3-VL 235B-A22B Thinking |
| LMArena Elo | 1422.4 ↗ | 1395 ↗ | +27.4 | Qwen3 235B-A22B Instruct 2507 |
| AIME25 | 70.3% ↗ | 89.7% ↗ | -19.4 pt | Qwen3-VL 235B-A22B Thinking |
| HMMT25 | 55.4% ↗ | 77.4% ↗ | -22 pt | Qwen3-VL 235B-A22B Thinking |
| IFEval | 88.7% ↗ | 88.2% ↗ | +0.5 pt | Qwen3 235B-A22B Instruct 2507 |
| MMLU-Redux | 93.1% ↗ | 93.7% ↗ | -0.6 pt | Qwen3-VL 235B-A22B Thinking |
| SimpleQA | 54.3% ↗ | 44.4% ↗ | +9.9 pt | Qwen3 235B-A22B Instruct 2507 |
| SuperGPQA | 62.6% ↗ | 64.3% ↗ | -1.7 pt | Qwen3-VL 235B-A22B Thinking |
| Only Qwen3 235B-A22B Instruct 2507 reports · 9 | ||||
| Aider-Polyglot | 57.3% ↗ | not reported | — | — |
| ARC-AGI | 41.8% ↗ | not reported | — | — |
| Arena-Hard v2 | 79.2% ↗ | not reported | — | — |
| BFCL-v3 | 70.9% ↗ | not reported | — | — |
| GPQA | 77.5% ↗ | not reported | — | — |
| LiveBench 20241125 | 75.4% ↗ | not reported | — | — |
| LiveCodeBench v6 (25.02-25.05) | 51.8% ↗ | not reported | — | — |
| TAU2-Retail | 74.6% ↗ | not reported | — | — |
| WeirdML (Håvard Tveit Ihle) | 38.7% ↗ | not reported | — | — |
| Only Qwen3-VL 235B-A22B Thinking reports · 9 | ||||
| BFCL | not reported | 71.9% ↗ | — | — |
| HLE | not reported | 13.6% ↗ | — | — |
| LCB v6 | not reported | 70.1% ↗ | — | — |
| LMArena Vision (LMArena) | not reported | 1188 ↗ | — | — |
| MathVista (mini) | not reported | 85.8% ↗ | — | — |
| MMMU (val) | not reported | 80.6% ↗ | — | — |
| MMMU_Pro | not reported | 69.3% ↗ | — | — |
| OSWorld | not reported | 38.1% ↗ | — | — |
| VideoMME (w/o sub) | not reported | 79% ↗ | — | — |
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 Instruct 2507 minus Qwen3-VL 235B-A22B Thinking 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 Instruct 2507 | Qwen3-VL 235B-A22B Thinking | Edge |
|---|---|---|---|
| Context window | 262K tokens | 256K tokens | Qwen3 235B-A22B Instruct 2507 |
| Max output | — | — | — |
| Input price / 1M | $0.23 ↗ | $0.4 ↗ | Qwen3 235B-A22B Instruct 2507 |
| Output price / 1M | $0.92 ↗ | $4 ↗ | Qwen3 235B-A22B Instruct 2507 |
| Cached input / 1M | — | — | — |
| Input + output / 1M Lower is cheaper. List prices; batch, tool and regional fees excluded. | $1.15 | $4.4 | Qwen3 235B-A22B Instruct 2507 |
| Modalities | text | text · vision · video | Qwen3-VL 235B-A22B Thinking |
| Released | 2025-07-21 | 2025-09-23 | — |
| Cited benchmark scores | 17 | 17 | — |
Reliability
Qwen status
Live from /statusMore matchups
Qwen3 235B-A22B Instruct 2507 vs …
Models sharing the most benchmarksMore matchups
Qwen3-VL 235B-A22B Thinking vs …
Models sharing the most benchmarksBuilt by Respan
Which one wins on your data?
Public benchmarks are a starting point. Run Qwen3 235B-A22B Instruct 2507 and Qwen3-VL 235B-A22B Thinking on your own prompts with Respan evals, or route to either through one gateway key with automatic failover.