Skip to content
Llama Builds

Alibaba Qwen · Released April 2026

Qwen3.6 35B-A3B

What it takes to run Qwen3.6 35B-A3B locally: memory by quantization, the GPUs that handle it, and how fast it runs.

Updated

Parameters
35B (3B active)
Architecture
Mixture of experts
Context
256K tokens
License
Apache 2.0 license
Memory at Q4_K_M
~24GB

What do you need to run Qwen3.6 35B-A3B?

Qwen3.6 35B-A3B needs about 24GB of VRAM at Q4_K_M with 8K context (22GB for the weights plus KV cache and overhead). The smallest GPU that can hold it is a 20GB card like the RTX 4000 Ada at Q3_K_M, with little memory to spare. For a good balance of price and speed we recommend the RTX 3090 (~160 tokens/sec). Because it is a mixture-of-experts model with 3B active parameters, it generates text much faster than a dense model of the same size, and it tolerates partial CPU offload better.

See the full hardware guide for Qwen3.6 35B-A3B

What to buy

GPUs for Qwen3.6 35B-A3B

Cheapest that runs it

RTX 4000 Ada

20GB · $1,250 MSRP

Quant
Q3_K_M
Speed (est.)
~88 tok/s
Fits, tight
Check price at Amazon

Best value

RTX 3090

24GB · ~$1,050 used

Quant
Q3_K_M
Speed (est.)
~160 tok/s
Runs well
Check price at Amazon

Fastest

RTX 5090

32GB · ~$4,300

Quant
Q5_K_M
Speed (est.)
~195 tok/s
Runs well
Check price at Amazon

Speeds are estimates at the quantization shown with 8K context. We may earn a commission when you buy through these links. Learn more.

Memory

Qwen3.6 35B-A3B VRAM requirements by quantization

Lower-bit quantizations shrink the model at some cost to quality. Q4_K_M is the usual sweet spot. Totals include the KV cache for the context length shown plus runtime overhead.

Memory needed for Qwen3.6 35B-A3B at each quantization level
QuantizationQualityWeightsTotal @ 8K ctxTotal @ 32K ctx
Q2_KRough14 GB15 GB15 GB
Q3_K_MUsable18 GB19 GB20 GB
Q4_K_MRecommendedDefault22 GB24 GB24 GB
Q5_K_MVery good26 GB28 GB28 GB
Q6_KNear-lossless30 GB32 GB32 GB
Q8_0Practically lossless39 GB41 GB41 GB
BF16Full precision74 GB76 GB77 GB

KV cache note: Hybrid layout: 30 of 40 layers use Gated DeltaNet linear attention with a fixed-size state, and only the 10 full-attention layers keep a growing cache: 2 x 10 x 2 KV heads x 256 dims x 2 bytes = 20,480 bytes per token. llama.cpp reports 5,120 MiB at the full 262,144 tokens, which matches this figure.

Performance

How fast Qwen3.6 35B-A3B runs

Single-user generation speed on popular GPUs that fit the model entirely in memory, at the best quantization each can hold.

  1. RTX 5090

    Q5_K_M · 32GB

    ~195tok/s

  2. RTX 4090

    Q3_K_M · 24GB

    ~175tok/s

  3. RTX 3090

    Q3_K_M · 24GB

    ~160tok/s

  4. RTX PRO 6000

    Q8_0 · 96GB

    ~155tok/s

  5. RTX PRO 5000 72GB

    Q8_0 · 72GB

    ~130tok/s

  6. RX 7900 XTX

    Q3_K_M · 24GB

    ~125tok/s

  7. V100 32GB

    Q5_K_M · 32GB

    ~105tok/s

  8. RTX 6000 Ada

    Q8_0 · 48GB

    ~100tok/s

  9. M5 Ultra 512GB

    Q8_0 · 512GB

    ~83tok/s

  10. M5 Ultra 256GB

    Q8_0 · 256GB

    ~83tok/s

Estimated (method)

Measured

Qwen3.6 35B-A3B benchmarks

Overview

About Qwen3.6 35B-A3B

Qwen3.6 35B-A3B is the April 2026 refresh of Qwen's fast mixture-of-experts model, focused on agentic coding and on preserving its reasoning across multi-turn tool use. With only 3B active parameters it generates text very quickly, and its hybrid linear attention keeps the KV cache tiny: the full 262K context needs about 5GB. At Q4_K_M it needs about 24GB with 8K context, which is tight on a 24GB card and comfortable on 32GB GPUs and Macs with 36GB or more; 12GB to 16GB GPUs run it well with experts offloaded to system RAM. It replaces Qwen3 30B-A3B and Qwen3 Coder 30B-A3B for most local users.

Strengths

  • +Very fast: 3B active parameters per token
  • +Tiny KV cache: 20 KiB per token, about 5GB at the full 262K context
  • +Strong agentic coding and tool use, with image input
  • +Apache 2.0 license

Weaknesses

  • −The 22GB Q4_K_M file is tight on 24GB GPUs once context is added
  • −Less world knowledge than dense 27B to 31B models
  • −Hybrid attention needs recent llama.cpp, MLX, or vLLM builds
CodingAgents & toolsReasoningChatVisionLong contextImage inputToggleable reasoning

Setup

How to run Qwen3.6 35B-A3B

Ollama
$ ollama run qwen3.6:35b
llama.cpp
$ llama-server -hf unsloth/Qwen3.6-35B-A3B-GGUF:Q4_K_M -c 8192 -ngl 99

Weights: Qwen/Qwen3.6-35B-A3B · GGUF: unsloth/Qwen3.6-35B-A3B-GGUF

Alternatives

Models like Qwen3.6 35B-A3B

Alibaba Qwen

Qwen3.8 27B

27B · 256K ctx

~18GBat Q4_K_M
CodingAgents & toolsReasoning

Google

Gemma 4 26B A4B

26B (3.8B active) · 256K ctx

~17GBat Q4_K_M
ChatVisionReasoning

Alibaba Qwen

Qwen3 30B-A3B (2507)

30.5B (3.3B active) · 256K ctx

~21GBat Q4_K_M
ChatReasoningCoding

OpenAI

gpt-oss-20b

20.9B (3.6B active) · 128K ctx

~15GBat MXFP4
ReasoningCodingAgents & tools

Frequently asked questions

How much VRAM do you need to run Qwen3.6 35B-A3B?

About 24GB at Q4_K_M with 8K tokens of context: 22.3GB of weights, 0.2GB of KV cache, and 1.3GB of runtime overhead. Higher-precision quants and longer contexts need more; see the table above.

What is the cheapest GPU that can run Qwen3.6 35B-A3B?

The smallest card in our database that can hold it is the RTX 4000 Ada (20GB), at Q3_K_M with little memory to spare, generating ~88 tokens/sec.

Is Qwen3.6 35B-A3B free for commercial use?

Qwen3.6 35B-A3B is released under the Apache 2.0 license. That is a permissive license, so commercial use is allowed.

How long a context does Qwen3.6 35B-A3B support?

Up to 262,144 tokens (256K). Long contexts need a lot of extra memory for the KV cache, so most local setups run 8K to 32K.

Sources: Qwen3.6 35B-A3B announcement, Unsloth Qwen3.6 35B-A3B GGUF files, Artificial Analysis: Qwen3.6 35B A3B. Capability score basis: Estimated between Qwen3.5 35B-A3B (37) and Qwen3.6 27B (46) on the Artificial Analysis Intelligence Index v4 (April 2026); no exact reading was confirmed.