Model

Qwen3.6-35B-A3B

OPEN

Alibaba · Qwen3.6 family · released Apr 16, 2026

First open-weight model of the Qwen3.6 generation; sparse MoE (35B total/3B active) that substantially beats its direct predecessor Qwen3.5-35B-A3B and rivals much larger dense models on agentic coding, per the launch blog. This is the single real model behind the leads' 'Qwen3.6 (~35B, open-weights)' and 'Qwen3.6-35B-A3B' entries, which were duplicates of each other (see excluded).

ReasoningCodingVisionFunction callingTool useAgentic

1973.2

Elo · rank #79

Parameters

35B

Active params

3B (MoE)

Context

256K tokens

Architecture

Hybrid-attention MoE (Qwen3-Next lineage): 40 layers, 10x(3x(Gated DeltaNet -> MoE) -> 1x(Gated Attention -> MoE)); 256 experts, 8 routed + 1 shared activated per token; hidden dim 2048; natively multimodal (Causal LM with built-in vision encoder)

License

Apache 2.0

Languages

—

API price (in/out)

No hosted API

Modalities

text · vision

Benchmark results

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

AI2DVision92.7%#16

best: Molmo 72B · 96.3%

AIMEMath92.7%#35

best: GPT-5.2 · 100.0%

C-EvalKnowledge90.0%#13

best: Qwen3.6-Plus · 93.3%

CharXivVision78.0%#22

best: Qwen3.8-Flash-Next · 90.6%

GPQA DiamondReasoning86.0%#59

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning21.4%#68

best: Claude Opus 5 · 64.7%

HumanEvalCoding93.3%#10

best: Claude Opus 4.5 · 99.4%

LiveCodeBenchCoding80.4%#49

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

MathVistaVision86.4%#7

best: Seed 2.1 Pro · 90.7%

MMBench (English)Vision92.8%#4

best: Qwen3.5-397B-A17B · 93.7%

MMLU-ProKnowledge85.2%#25

best: Claude Fable 5 · 91.5%

MMLU-ReduxKnowledge93.3%#13

best: Qwen3.7-Max · 95.0%

MMMU-ProVision75.3%#28

best: Claude Opus 4.7 · 85.5%

MMMUVision81.7%#19

best: Claude Fable 5 · 89.3%

RefCOCOVision92.0%#9

best: Qwen3.5-Omni-Plus · 95.0%

SuperGPQAReasoning64.7%#11

best: Qwen3.7-Max · 73.6%

SWE-bench ProCoding49.5%#50

best: Claude Fable 5.1 · 81.2%

SWE-bench VerifiedCoding73.4%#51

best: Claude Opus 5 · 96.0%

Terminal-Bench 2.0Coding51.5%#59

best: Gemini 3.8 Flash · 89.4%

Video-MMEVision86.6%#7

best: Seed 2.1 Pro · 89.2%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Permissive

QLoRA24.1 GB1× RTX 5090 32GB

LoRA74.8 GB1× A100 80GB

Full fine-tune562.0 GB4× H200 141GB

QLoRA SFT on ~10k samples ≈ $17.07 (1× RTX 5090 32GB)

Plan a fine-tune →

Qwen3.6 family

Elo progression across releases

Qwen3.6-PlusApr 20262290.1 Qwen3.6-35B-A3BApr 20261973.2 Qwen3.6-27BApr 20262027.4

Resources

HFQwen/Qwen3.6-35B-A3B GHQwenLM/Qwen3.6 DOCShttps://huggingface.co/Qwen/Qwen3.6-35B-A3B APIAvailable via Alibaba API

API price weights · each benchmark row carries its own source badge (see methodology)