Model

Qwen3-235B-A22B-Thinking-2507

OPEN

Alibaba · Qwen3 family · released Jul 25, 2025

Paired always-on reasoning ('Thinking-only') sibling of Qwen3-235B-A22B-Instruct-2507, released 4 days later. Enforces <think> mode, extends max output to 81,920 tokens, native 262,144-token context. Same 235B/22B-active MoE backbone as the Instruct-2507 release already in corpus, but a distinct fine-tune (not a shared-weights effort tier), so tracked as its own standalone entry.

ReasoningCodingVisionFunction callingTool useAgentic

1842.4

Elo · rank #76

Parameters

235B

Active params

22B (MoE)

Context

256K tokens

Architecture

MoE, 128 experts / 8 active

License

Apache 2.0

Languages

119+

API price (in/out)

$0.23 / $2.3

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

AIMEMath92.3%#36

best: GPT-5.2 · 100.0%

Arena-Hard v2Human preference79.7%#2

best: Qwen3-Max · 86.1%

Arena EloHuman preference1399#23

best: Claude Fable 5 · 1505

BFCL v3Agents71.9%#4

best: Hunyuan-A13B · 78.3%

CodeforcesCoding2134#12

best: DeepSeek-V4-Pro (Think Max) · 3206

GPQA DiamondReasoning81.1%#72

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning18.2%#66

best: Claude Sonnet 5 · 57.4%

IFEvalReasoning87.8%#36

best: Gemma 4 26B A4B · 98.5%

LiveCodeBenchCoding74.1%#59

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MMLU-ProKnowledge84.4%#32

best: Claude Fable 5 · 91.5%

MMLU-ReduxKnowledge93.8%#9

best: Qwen3.7-Max · 95.0%

SuperGPQAReasoning64.9%#10

best: Qwen3.7-Max · 73.6%

τ²-Bench TelecomAgents45.6%#32

best: Claude Opus 4.6 · 99.3%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Qwen3 family

Elo progression across releases

Qwen3-14B (Non-Thinking)Apr 20251094.6 Qwen3-14B (Thinking)Apr 20251396.5 Qwen3-235B-A22B (Non-Thinking)Apr 20251235.2 Qwen3-235B-A22B (Thinking)Apr 20251536.2 Qwen3-30B-A3B (Non-Thinking)Apr 20251101.1 Qwen3-30B-A3B (Thinking)Apr 20251483.2 Qwen3-32B (Non-Thinking)Apr 20251121.8 Qwen3-32B (Thinking)Apr 20251509.2 Qwen3-8B (Non-Thinking)Apr 2025922.1 Qwen3-8B (Thinking)Apr 20251323.0 Qwen3-235B-A22B-Instruct-2507Jul 20251591.0 Qwen3-235B-A22B-Thinking-2507Jul 20251842.4

Resources

HFQwen/Qwen3-235B-A22B-Thinking-2507 DOCSqwen.readthedocs.io APIAvailable via Alibaba API

API price $0.23/$2.3 · each benchmark row carries its own source badge (see methodology)