Qwen3.6-35B-A3B

Qwen Team (Alibaba Cloud)

Parameters

35.0B total / 3.0B active

MoE: total / active

Architecture

Hybrid Gated DeltaNet + Gated Attention MoE

Released

15.04.2026

License

Apache License 2.0

Open Weights Commercial Use Multimodal BF16 Qwen3.6 en zh multi

Input Modalities

text image video

Output Modalities

text

Context (native)

262,144 tokens

Context (extended)

1,010,000 tokens

Openness Index Score 100.0/100

About

Qwen3.6-35B-A3B (Qwen/Qwen3.6-35B-A3B) is the mid-generation sparse Mixture-of-Experts multimodal model of Alibaba's Qwen3.6 line - 35B total parameters with 3B activated per token, handling text, image and video input. Its 40-layer hybrid stack repeats 10x (3x (Gated DeltaNet -> MoE) + 1x (Gated Attention -> MoE)): 30 linear-attention layers (Gated DeltaNet, 32 V / 16 QK heads, head dim 128) and 10 full-attention layers (Gated Attention, 16 Q / 2 KV heads, head dim 256, RoPE dim 64), with 256 experts (8 routed + 1 shared per token, expert dim 512), hidden size 2048, 248K vocabulary, Multi-Token Prediction (MTP) trained with multi-steps, and a 262,144-token native context extensible to 1,010,000.

Qwen3.6 delivers substantial upgrades in agentic coding - frontend workflows and repository-level reasoning with greater fluency and precision - and introduces thinking preservation, a new option to retain reasoning context from historical messages that streamlines iterative development and reduces overhead. Released April 15, 2026 under Apache 2.0.

Training Data Pre-training & Post-training. Trained with multi-token prediction (MTP) steps.

Benchmark Scores

Benchmark Score Date Standing
SWE-bench Verified
coding_agent
83.72%
16.04.2026
#33/60
better than 46%
SWE-bench Multilingual
coding_agent
73.61%
16.04.2026
#31/48
better than 36%
SWE-bench Pro
coding_agent
55.18%
16.04.2026
#50/77
better than 36%
Terminal-Bench 2.0
coding_agent
53.55%
16.04.2026
#12/23
better than 50%
Claw-Eval Avg
coding_agent
82.81%
16.04.2026
#11/40
better than 74%
SkillsBench Avg5
coding_agent
39.00%
16.04.2026
#12/16
better than 27%
QwenClawBench
coding_agent
98.36%
16.04.2026
#2/17
better than 94%
NL2Repo
coding_agent
29.38%
16.04.2026
#33/41
better than 20%
QwenWebBench
coding_agent
75.09%
16.04.2026
#3/8
better than 71%
TAU3-Bench
general_agent
97.49%
16.04.2026
#5/21
better than 80%

Showing 10 of 85 results

What is a good SWE-bench Verified score? What is a good SWE-bench Multilingual score? What is a good SWE-bench Pro score?

Hardware Requirements

Estimated VRAM per quantization at 4,096 tokens context — MoE: all experts must fit in VRAM, only 3.0B are active per token

Quantization Weights KV Cache Total VRAM On your hardware
FP16 — 16-bit 70.0 GB 0.5 GB 74.0 GB Which GPUs can run it? Calculate on your GPU
INT8 — 8-bit 35.0 GB 0.2 GB 37.0 GB Which GPUs can run it? Calculate on your GPU
INT4 — 4-bit 17.5 GB 0.1 GB 18.5 GB Which GPUs can run it? Calculate on your GPU
FP4 — 4-bit 17.5 GB 0.1 GB 18.5 GB Which GPUs can run it? Calculate on your GPU

Only datacenter GPUs — FP16 needs 74.0 GB. For consumer hardware, choose the INT4 build with a multi-GPU setup.

Consumer GPUs that fit this model with a single card (INT4):

GeForce RTX 3090 · 24GB GeForce RTX 4090 D · 24GB AMD Radeon RX 7900 XTX · 24GB GeForce RTX 3090 Ti · 24GB GeForce RTX 4090 · 24GB Apple M4 Pro · 24GB Apple M1 Max · 32GB Apple M2 Max · 32GB GeForce RTX 5090 · 32GB Apple M3 Max · 36GB Apple M4 Max · 36GB Apple M1 Ultra · 64GB

Architecture

Decoder Block ×40 input Embedding vocab 248K · d 2048 Linear / Recurrent Hybrid 16:2 · dₕ 256 ×30 Full Attention Hybrid 16:2 · dₕ 256 ×10 MoE FFN 256 experts · top-8 · dᴻ 512 Final Norm LM Head vocab 248K output
Attention
Hybrid Attention (16:2)
MoE
256 experts · top-8 per token
Layers
40
Hidden size
2048
Context
262K tokens
Parameters
35000M
Active params
3000M

Source: Hugging Face config.json · Qwen3_5MoeForConditionalGeneration · exact layer pattern · model repo

Type: Hybrid MoE Causal LM with Vision Encoder
Attention: Hybrid Gated DeltaNet + Gated Attention (10x 3:1 layout)
Decoder: 40-layer hybrid decoder
MoE: yes (256 experts)
Routing: Top-k routing 8 routed + 1 shared expert

Training Pipeline

  1. 1
    pretraining

    Pre-training

    Pre-training of the Qwen3.6-35B-A3B base model with multi-token prediction (MTP) steps.

  2. 2
    sft

    Post-training (SFT + RL)

    Post-training stage including SFT and reinforcement learning to produce the instruct/chat variant with thinking mode and agentic coding capabilities.

Training & Evaluation Datasets

NameRoleSizeModalitiesCollection
Qwen3.6 pre-training corpus (multimodal) pretraining — —

Linked Resources

blog

Qwen3.6-35B-A3B: Agentic Coding Power, Now Open to All

https://qwen.ai/blog?id=qwen3.6-35b-a3b

 github

QwenLM/Qwen-Agent

https://github.com/QwenLM/Qwen-Agent

 github

QwenLM/qwen-code

https://github.com/QwenLM/qwen-code

 demo

Qwen Chat

https://chat.qwen.ai/

 collection

Qwen3.6 Collection

https://huggingface.co/collections/Qwen/qwen36

 citation

Citation for Qwen3.6-35B-A3B

https://qwen.ai/blog?id=qwen3.6-35b-a3b

Comments (0)

No comments yet

Be the first to share your thoughts!

What now?

You have read the full card for Qwen3.6-35B-A3B — here is where to go next.

Compare with another model Test on your own GPU Related models Data as JSON