Skip to content
Local AI Wizard
Your machine
Theme

Qwen3.5 35B A3B

Alibaba · Qwen/Qwen3.5-35B-A3B

Text & chatVisionVideo

Q4 file 22.0 GB (Q4_K_M) · runs well from 32 GB · Weights license: Apache 2.0

LMArena · text1,394 ± 4Rank 172 · release 2026-10-22 · its own scale
Memory needed~27 GBQ4 weights + working room, before the machine's reserve
Parameters35 Blanguage component · 3B active
Context262,144 tokensPublished for this model

Quality

LMArena · text · rating
1,394 ± 4Confidence interval 1,390–1,398 · 30,443 votes · rank 172 of the releaseLMArena 2026-10-22 · CC BY 4.0

No result in LiveBench 2026-06-25 Not in LiveBench 2026-06-25

Reported by the publisher

79 results · Alibaba's model card · read 2026-10-25
MMLU-Pro
85.3Language
MMLU-Redux
93.3Language
C-Eval
90.2Language
SuperGPQA
63.4Language
IFEval
91.9Language
IFBench
70.2Language
MultiChallenge
60.0Language
AA-LCR
58.5Language
Every reported result79
LongBench v2
59.0Language
HLE w/ CoT
22.4Language
GPQA Diamond
84.2Language
HMMT Feb 25
89.0Language
HMMT Nov 25
89.2Language
SWE-bench Verified
69.2Language
Terminal Bench 2
40.5Language
LiveCodeBench v6
74.6Language
CodeForces
2,028Language
OJBench
36.0Language
FullStackBench en
58.1Language
FullStackBench zh
55.0Language
BFCL-V4
67.3Language
TAU2-Bench
81.2Language
VITA-Bench
31.9Language
DeepPlanning
22.8Language
HLE w/ tool
47.4Language
Browsecomp
61.0Language
Browsecomp-zh
69.5Language
WideSearch
57.1Language
Seal-0
41.4Language
MMMLU
85.2Language
MMLU-ProX
81.0Language
NOVA-63
57.1Language
INCLUDE
79.7Language
Global PIQA
86.6Language
PolyMATH
64.4Language
WMT24++
76.3Language
MAXIFE
86.6Language
MMMU
81.4Vision Language
MMMU-Pro
75.1Vision Language
MathVision
83.9Vision Language
Mathvista (mini)
86.2Vision Language
DynaMath
85.0Vision Language
ZEROBench
8Vision Language
ZEROBench_sub
34.1Vision Language
VlmsAreBlind
97.0Vision Language
RealWorldQA
84.1Vision Language
MMStar
81.9Vision Language
MMBench EN-DEV-v1.1
91.5Vision Language
SimpleVQA
58.3Vision Language
HallusionBench
67.9Vision Language
OmniDocBench1.5
89.3Vision Language
CharXiv (RQ)
77.5Vision Language
MMLongBench-Doc
59.5Vision Language
CC-OCR
80.7Vision Language
AI2D_TEST
92.6Vision Language
OCRBench
91.0Vision Language
ERQA
64.8Vision Language
CountBench
97.8Vision Language
RefCOCO (avg)
89.2Vision Language
ODInW13
42.6Vision Language
EmbSpatialBench
83.1Vision Language
RefSpatialBench
63.5Vision Language
LingoQA
79.2Vision Language
Hypersim
13.1Vision Language
SUNRGBD
33.4Vision Language
Nuscene
14.6Vision Language
VideoMME (w sub.)
86.6Vision Language
VideoMME (w/o sub.)
82.5Vision Language
VideoMMMU
80.4Vision Language
MLVU
85.6Vision Language
MVBench
74.8Vision Language
LVBench
71.4Vision Language
MMVU
72.3Vision Language
ScreenSpot Pro
68.6Vision Language
OSWorld-Verified
54.5Vision Language
AndroidWorld
71.1Vision Language
SLAKE
78.7Vision Language
PMC-VQA
62.0Vision Language
MedXpertQA-MM
61.4Vision Language
Across Qwen3.5 sizes66 shared results · same publisher and settings
Qwen3.5: results the publisher reports for each size under the same settings
Benchmark (publisher settings)9B35B A3B
MMLU-Pro82.585.3
MMLU-Redux91.193.3
C-Eval88.290.2
SuperGPQA58.263.4
IFEval91.591.9
IFBench64.570.2
MultiChallenge54.560.0
AA-LCR63.058.5
LongBench v255.259.0
GPQA Diamond81.784.2
HMMT Feb 2583.289.0
HMMT Nov 2582.989.2
LiveCodeBench v665.674.6
OJBench29.236.0
BFCL-V466.167.3
TAU2-Bench79.181.2
VITA-Bench29.831.9
DeepPlanning18.022.8
MMMLU81.285.2
MMLU-ProX76.381.0
NOVA-6355.957.1
INCLUDE75.679.7
Global PIQA83.286.6
PolyMATH57.364.4
WMT24++72.676.3
MAXIFE83.486.6
MMMU78.481.4
MMMU-Pro70.175.1
MathVision78.983.9
Mathvista (mini)85.786.2
DynaMath83.685.0
ZEROBench3.08
ZEROBench_sub31.134.1
VlmsAreBlind93.797.0
RealWorldQA80.384.1
MMStar79.781.9
MMBench EN-DEV-v1.190.191.5
SimpleVQA51.258.3
HallusionBench69.367.9
OmniDocBench1.587.789.3
CharXiv (RQ)73.077.5
MMLongBench-Doc57.759.5
CC-OCR79.380.7
AI2D_TEST90.292.6
OCRBench89.291.0
ERQA55.564.8
CountBench97.297.8
RefCOCO (avg)89.789.2
EmbSpatialBench83.083.1
RefSpatialBench58.563.5
LingoQA80.479.2
Hypersim13.513.1
Nuscene11.814.6
VideoMME (w sub.)84.586.6
VideoMME (w/o sub.)78.482.5
VideoMMMU78.980.4
MLVU84.485.6
MVBench74.474.8
LVBench70.071.4
MMVU67.872.3
ScreenSpot Pro65.268.6
OSWorld-Verified41.854.5
AndroidWorld57.871.1
SLAKE79.078.7
PMC-VQA57.962.0
MedXpertQA-MM49.961.4

Values as printed on the card, with the publisher's own settings (card section: Benchmark Results › Language; Benchmark Results › Vision Language); column “Qwen3.5-35B-A3B”. Not comparable across publishers and not used in any ranking. Model card

Run it

Run it

File Q4_K_M · 22 GB · unsloth/Qwen3.5-35B-A3B-GGUF · What is Q4?

Measured 63.1 tok/s · MXFP4 · 1 machine · 2 sources · all rows

ollama run hf.co/unsloth/Qwen3.5-35B-A3B-GGUF:Q4_K_M
  1. Install Ollama
  2. Run the command above
  3. Ask: Explain what a context window is, in two sentences.

Runs on

Fast memory holds it 33

The planning estimate fits the memory the model can use quickly: graphics card VRAM, or unified memory within the macOS GPU share.

Show all 33

Within memory, outside the fast part 50

Within the machine's memory, outside its fast part: on a graphics card the layers that do not fit run on the CPU from PC RAM (the listed RAM, or 32 GB assumed); on a Mac, above the macOS default GPU share. Not a speed figure.

Show all 50

One memory tier short 9

The same machines with the next memory size up would hold it in fast memory.

Show all 9

Qwen3.5 35B A3B is a text model from Alibaba with 35B parameters, 3B of them active per token (mixture of experts). It was released on February 24, 2026 and reads up to 262,144 tokens of context. Its Q4 weights plus working room come to about 27 GB before the machine's reserve; the smallest reference machines whose fast memory holds that are Mac mini M5 Pro · 64 GB unified (~35 GB with its reserve) and RTX 5090 · 32 GB VRAM (~29 GB with its reserve). We collected 27 downloadable GGUF and MLX files from 2 publisher sources; the most downloaded GGUF source is unsloth's (5.5M downloads on the Hub). It has no result in LiveBench 2026-06-25. Its LMArena text rating is 1,394 ± 4 (release 2026-10-22), a separate human-preference scale.

Hardware profiles grouped by the Q4 planning estimate on each machine (dedicated GPUs reserve 2 GB; unified and system memory 25%, 2 to 8 GB): the fast memory holds it, it is within memory but outside the fast part (layers offloaded to PC RAM, or above the macOS GPU share), or the machine is one memory tier short. Each card opens its hardware profile. Not a measured fit or speed.

On the reference machines

Qwen3.5 35B A3B on the reference machines
Where Q4 sits
Mac mini M2 · 8 GB unifiedLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column~29 GBLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column
Mac mini M4 · 16 GB unifiedLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column~31 GBLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column
Mac mini M4 · 24 GB unifiedLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column~33 GBLarger than its memoryNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column
Mac mini M4 · 32 GB unifiedLarger than its memoryUD-IQ4_XSGGUF · 17.5 GB · needs ~30 GB131,072 tokens~35 GBLarger than its memoryUD-IQ4_XSGGUF · 17.5 GB · needs ~30 GB131,072 tokens
Mac mini M5 Pro · 64 GB unifiedFast memory holds itQ8_0GGUF · 36.9 GB · needs ~53 GB131,072 tokens~35 GBFast memory holds itQ8_0GGUF · 36.9 GB · needs ~53 GB131,072 tokens
Mac Studio M5 Max · 128 GB unifiedFast memory holds itQ8_0GGUF · 36.9 GB · needs ~53 GB262,144 tokens~35 GBFast memory holds itQ8_0GGUF · 36.9 GB · needs ~53 GB262,144 tokens
Mac Studio · M5 Ultra · 36-core CPU / 80-core GPU · 512 GB unifiedFast memory holds itQ8_0GGUF · 36.9 GB · needs ~53 GB262,144 tokens~35 GBFast memory holds itQ8_0GGUF · 36.9 GB · needs ~53 GB262,144 tokens
RTX 4060 · 8 GB VRAMIn memory, outside the fast partNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column~29 GBIn memory, outside the fast partNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column
RTX 3060 · 12 GB VRAMIn memory, outside the fast partNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column~29 GBIn memory, outside the fast partNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column
RTX 5080 · 16 GB VRAMIn memory, outside the fast partNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column~29 GBIn memory, outside the fast partNo file on the site's quant ladder fits the fast memory at 8,192 tokensNo file in the previous column
RTX 4090 · 24 GB VRAMIn memory, outside the fast partUD-IQ4_XSGGUF · 17.5 GB · needs ~24 GB32,768 tokens~29 GBIn memory, outside the fast partUD-IQ4_XSGGUF · 17.5 GB · needs ~24 GB32,768 tokens
RTX 5090 · 32 GB VRAMFast memory holds itQ4_K_MGGUF · 22.0 GB · needs ~29 GB131,072 tokens~29 GBFast memory holds itQ4_K_MGGUF · 22.0 GB · needs ~29 GB131,072 tokens
DGX H100 · 8 × 80 GBFast memory holds itQ8_0GGUF · 36.9 GB · needs ~61 GB262,144 tokens~43 GBFast memory holds itQ8_0GGUF · 36.9 GB · needs ~61 GB262,144 tokens

The file column uses the site's quant ladder (Q8_0 to Q3_K_M, MLX 8 to 4-bit on a Mac): the largest collected file whose size × 1.2 plus the context cache and the machine's reserve fits the fast memory. Planning estimates, not measured runs.

Plan another quantization or context

Consider it for

Compare it with the dense Qwen 27B checkpoint for coding; the published MoE design activates about 3B language parameters per token.

Check first

Active parameters are not the weights to store. The full 35B weight count matters to the generic memory plan.

Versions to download

Files you can download for this checkpoint. The section opens on the most downloaded collected GGUF source, unsloth (5.5M Hub downloads), at Q4_K_M. On a Mac it opens on the most downloaded MLX source. Sizes are the Hub's published bytes. What is Q4?

unsloth · 5.5M downloads

unsloth/Qwen3.5-35B-A3B-GGUF · Qwen3.5-35B-A3B-Q4_K_M.gguf · 22.0 GB

Run this file

ollama run hf.co/unsloth/Qwen3.5-35B-A3B-GGUF:Q4_K_M
llama-server -hf unsloth/Qwen3.5-35B-A3B-GGUF:Q4_K_M
lms get https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF
  • mlx-lm: mlx-lm runs MLX repositories; pick the MLX version on a Mac

Quality scores on this page belong to the original checkpoint; these files are not scored or ranked. File sizes read 2026-10-23 · download counts read 2026-10-17 · commands: Hugging Face Hub docs: Use Ollama with any GGUF model · llama.cpp server README (-hf <user>/<model>[:quant]) · LM Studio docs: lms get with a Hugging Face URL · mlx-lm README (mlx_lm.generate --model)

Fine-tunes, adapters and other versions186 fine-tunes
Quantized versions
310
Fine-tunes
186
Adapters · merges
48 · 8

Browse them on the Hugging Face model tree · read 2026-10-23

Uncensored and abliterated versions114

Some community uploads remove the model's safety refusals. Quality may drop, the original publisher does not support them, claims about them are the uploader's, and the original license still applies. This site never recommends or ranks them.

114 derivative repositories have “uncensored” or “abliterated” in their name (read 2026-10-23).

Measured speed

Tokens per second by machine

Output tok/s · 2 machines · every row in the table · What is a token?

DGX Spark · MXFP4, 63.13 · MXFP4 (MoE; 35B total / 3B active) · llama.cpp CUDA (SM121a, source build) · llama.cpp b8922 (commit 13d36cf89); cmake -DGGML_CUDA=ON -DGGML_CUDA_FA_ALL_QUANTS=ON -DCMAKE_CUDA_ARCHITECTURES=121 -DGGML_CUDA_F16=ON -DGGML_NATIVE=ON; llama-bench -p 512 -n 128 -ngl 99 -fa 1 -mmp 0 -r 5; NVIDIA DGX OS (Ubuntu 24.04.3), kernel 6.14.0-1015-nvidia, CUDA 13.063.1
MacBook Air 13-inch M5 32GB · 2-bit oQ2, 55.0 · 2-bit oQ2 (group size 64), MLX safetensors · oMLX (MLX) with native/lightning MTP speculative decoding, TurboQuant KV cache enabled, max 4 concurrent requests · oMLX v0.5.2.dev1; MLX and macOS versions not stated55.0
MacBook Air 13-inch M5 32GB · 3-bit oQ3, 55.0 · 3-bit oQ3 (group size 64), MLX safetensors · oMLX (MLX) with native/lightning MTP speculative decoding, TurboQuant KV cache enabled, max 4 concurrent requests · oMLX v0.5.2.dev1; MLX and macOS versions not stated55.0
Measured speed
DGX SparkDGX Spark · 1× NVIDIA GB10 · 128 GB unified63.13MXFP4 (MoE; 35B total / 3B active)0 / 128 · Single llama-bench process; server concurrency not measuredllama.cpp CUDA (SM121a, source build)llama.cpp b8922 (commit 13d36cf89); cmake -DGGML_CUDA=ON -DGGML_CUDA_FA_ALL_QUANTS=ON -DCMAKE_CUDA_ARCHITECTURES=121 -DGGML_CUDA_F16=ON -DGGML_NATIVE=ON; llama-bench -p 512 -n 128 -ngl 99 -fa 1 -mmp 0 -r 5; NVIDIA DGX OS (Ubuntu 24.04.3), kernel 6.14.0-1015-nvidia, CUDA 13.0Flash Attention enabledQwen3.5-35B-A3B MXFP4 GGUF (the source names neither the repository nor the file)nabe2030/dense-27b-31b-dgx-spark (GitHub) · community benchmark · 2026-10-09
MacBook Air 13-inch M5 32GBMacBook Air · Apple M5 · 32 GB unified (GPU cores not stated)55.02-bit oQ2 (group size 64), MLX safetensors1024 / 128 · Single request (oMLX configured for up to 4 concurrent requests); batched throughput is lower per requestoMLX (MLX) with native/lightning MTP speculative decoding, TurboQuant KV cache enabled, max 4 concurrent requestsoMLX v0.5.2.dev1; MLX and macOS versions not statedmlx-works/Qwen3.5-35B-A3B-oQ2-mtp (third-party 2-bit MLX conversion with embedded MTP)mlx-works model card (Hugging Face) · third-party quantization · 2026-10-09
MacBook Air 13-inch M5 32GBMacBook Air · Apple M5 · 32 GB unified (GPU cores not stated)55.03-bit oQ3 (group size 64), MLX safetensors1024 / 128 · Single request (oMLX configured for up to 4 concurrent requests); batched throughput is lower per requestoMLX (MLX) with native/lightning MTP speculative decoding, TurboQuant KV cache enabled, max 4 concurrent requestsoMLX v0.5.2.dev1; MLX and macOS versions not statedmlx-works/Qwen3.5-35B-A3B-oQ3-mtp (third-party 3-bit MLX conversion with embedded MTP)mlx-works model card (Hugging Face) · third-party quantization · 2026-10-09

Sources: nabe2030/dense-27b-31b-dgx-spark (GitHub) · community benchmark · mlx-works model card (Hugging Face) · third-party quantization · mlx-works model card (Hugging Face) · third-party quantization · not estimated

How to run

~27 GB of Q4 weights plus working room, before the machine's reserve. Formula: parameters × 0.625 bytes, plus 20% working allowance; each machine then adds its system reserve: 25% of unified or system memory, 2 to 8 GB, or 2 GB per GPU on dedicated cards. This is neither measured VRAM nor checkpoint download size.

Q4 planning estimate only. Check checkpoint format, runtime support, KV cache, context, batch size and offload before choosing hardware.

Documented ways to run it

Copied as printed in the linked source (read 2026-10-18); not run by this site.

vLLM · model card · vLLM
vllm serve Qwen/Qwen3.5-35B-A3B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 
SGLang · model card · SGLang
python -m sglang.launch_server --model-path Qwen/Qwen3.5-35B-A3B --port 8000 --tp-size 8 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3
Transformers · model card · Hugging Face Transformers
transformers serve --force-model Qwen/Qwen3.5-35B-A3B --port 8000 --continuous-batching

Evidence level: Generic estimate only; no execution reproduced by this site.

Publisher sampling defaults
temperature 1 · top_p 0.95 · top_k 20generation_config.json · read 2026-10-23source
Chat template markers
tool calls · thinkingmarkers found in the template, not a test of the feature · read 2026-10-23source
Before downloading
  1. Check the exact checkpoint and license linked here.
  2. Confirm a compatible runtime and precision or quantization for your platform.
  3. Allow for context, batch size and any extra media components beyond the weights.

Technical details

What the source and the catalog store for this checkpoint. Missing fields are left out, never zero.

Checkpoint / version
Qwen/Qwen3.5-35B-A3B
Task and modality
language and reasoning · text + image + video
Count and basis
35B · language component · 3B active
Context
262,144 tokens published for this model
Architecture
MoE
Released
QwenLM Qwen GitHub README (News) · read 2026-10-13source
Runtime and precision
No local execution verified.
Languages
201 languages Publisher's statement · read 2026-10-18source
Commercial use
Allowed Apache License 2.0 text (SPDX) · read 2026-10-18license text
Every stored field12 fields
Revision checked
59d61f3ce65a6d9863b86d2e96597125219dc754
Architecture
Qwen3_5MoeForConditionalGeneration
Library
transformers
Hub task tag
image-text-to-text
Declared base models
Qwen/Qwen3.5-35B-A3B-Base
Formats
transformers, safetensors
Access
Open download
Publisher kind
Official
Repository created
2026-02-24creation is not the model's release
Repository updated
2026-04-24a commit is not a release date
Hub downloads (30 days) · likes
1,476,096 · 1,546this repository only · read 2026-10-25
Catalog state
Hub config or weights found; no inventory benchmark link

The MoE activates about 3B language parameters per token, but all 35B weights still matter for storage and the generic memory estimate.

Sources

Identity and published specifications come from the checkpoint page. Benchmark scores belong to the named release and do not measure speed on your computer.

Catalog reviewed: 2026-09-30. Benchmark release: 2026-06-25; snapshot collected: 2026-09-26.