SAVRN
Search Contact SAVRN

Open-weight model · Text generation

Qwen3.5-397B-A17B-VQ-2.2bpw

by Noah Zelezny TheDrainFlorist/Qwen3.5-397B-A17B-VQ-2.2bpw

Qwen3.5-397B-A17B-VQ-2.2bpw is an open-weight model for text generation from Noah Zelezny, released under Apache License 2.0. It has 62.6B parameters and a 262,144-token context. At 16-bit it needs about 150.2 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index. It draws 1.6k downloads a month.

88.7 GiB text weights — the accessibility build, the roomiest fit on a 128 GB Mac. Also ships the bf16 vision tower (+0.85 GiB) and an optional MTP draft head (+5.4 GiB); full download 95.0 GiB. (v2, mixed geometry.) v2 — updated 2026-08-22.

Parameters62.6B
Context262,144
Weights102.0 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads1.6k

Runs On

What it takes to serve Qwen3.5-397B-A17B-VQ-2.2bpw (62.6B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

PrecisionWeightsMemory neededCheapest setupPer hourAlso fits
16-bit 125.1 GB 150.2 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x MI325X $2.00 · 1x MI355X $2.59
8-bit 62.6 GB 75.1 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x H100 $1.99 · 1x MI325X $2.00
4-bit 31.3 GB 37.5 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x H100 $1.99 · 1x MI325X $2.00

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 7, 2026.

Qwen3.5-397B-A17B-VQ-2.2bpw on every accelerator the SAVRN Index prices, at every precision

Model Card

By Noah Zelezny, published under apache-2.0, revision f10167305589.

88.7 GiB text weights — the accessibility build, the roomiest fit on a 128 GB Mac. Also ships the bf16 vision tower (+0.85 GiB) and an optional MTP draft head (+5.4 GiB); full download 95.0 GiB. (v2, mixed geometry.)

v2 — updated 2026-08-22. This repository now serves a rebuilt artifact at the same size and the same bits per weight, with a different codebook geometry that measures better on both perplexity corpora. v1's numbers are kept below rather than quietly overwritten, and v1's bytes remain downloadable by pinning the previous revision:

snapshot_download("TheDrainFlorist/Qwen3.5-397B-A17B-VQ-2.2bpw",
                  revision="4554635165011f67e8166fd94d4bcc8cbf91401c")  # v1

A vector-quantized build of Qwen3.5-397B-A17B built to answer one question: how small can a 397B get and still be worth running? 88.7 GiB text weights — it runs on a single 128 GB Apple Silicon machine with ≈7 GiB more headroom than our VQ-2.4bpw build, no cluster, no patches, stock mlx-lm.

Changelog

2026-10-01 — vq-skipzero: dead rows dropped

Read the full model card (2,914 words)

Configuration

Architecture
Qwen3_5MoeForConditionalGeneration
Context length (tokens)
262,144
Layers
60
Hidden size
4,096
Attention heads
32
Key/value heads
2
Head dimension
256
Vocabulary size
248,320
Experts
512
Experts active per token
10
Model type
qwen3_5_moe

Identity and Version

Repository
TheDrainFlorist/Qwen3.5-397B-A17B-VQ-2.2bpw
Publisher
Noah Zelezny
Task
Text generation
Modality
Text
Library
mlx
Parameters
62.6B parameters
Languages
en
Revision
f101673055892f18fd0645c3f9af8e1c5f7093a2
First published
2026-08-16
Last updated
2026-10-02

Files and Weights

44 files, 107.8 GB in total. The weights are 29 files totalling 102.0 GB in safetensors.

Weights29 files · 102.0 GB
Configuration8 files · 714.9 KB
Tokenizer2 files · 20.0 MB
Documentation1 file · 20.4 KB
Other3 files · 5.8 GB
Repository1 file · 1.6 KB
Every file
FileTypeSizeSHA-256
model-00001-of-00027.safetensorsWeights2.2 GB 4c9d468f14fd
model-00002-of-00027.safetensorsWeights2.3 GB 53a0a50986d8
model-00003-of-00027.safetensorsWeights2.9 GB 7931585f0181
model-00004-of-00027.safetensorsWeights3.2 GB 830f8012d9a8
model-00005-of-00027.safetensorsWeights3.0 GB f41620a614e8
model-00006-of-00027.safetensorsWeights3.1 GB aef6a858297c
model-00007-of-00027.safetensorsWeights3.5 GB d22ab12bda77
model-00008-of-00027.safetensorsWeights3.4 GB 4dbb6ccfcd44
model-00009-of-00027.safetensorsWeights3.5 GB e9d20b1fd94c
model-00010-of-00027.safetensorsWeights3.7 GB 90445e50e95c
model-00011-of-00027.safetensorsWeights3.7 GB d5a030ac02e8
model-00012-of-00027.safetensorsWeights3.7 GB 003895d9f2a3
model-00013-of-00027.safetensorsWeights4.0 GB 24671c33a20d
model-00014-of-00027.safetensorsWeights3.9 GB 9871794b6b74
model-00015-of-00027.safetensorsWeights3.9 GB aec2476a3a9b
model-00016-of-00027.safetensorsWeights4.0 GB d2b1f4f57771
model-00017-of-00027.safetensorsWeights4.0 GB 473f1c4c4b77
model-00018-of-00027.safetensorsWeights4.0 GB de0b845182b3
model-00019-of-00027.safetensorsWeights4.2 GB cf5298e7cfb2
model-00020-of-00027.safetensorsWeights4.0 GB f7653046f408
model-00021-of-00027.safetensorsWeights4.1 GB cc39e393a478
model-00022-of-00027.safetensorsWeights4.1 GB 4b3ec939a385
model-00023-of-00027.safetensorsWeights3.7 GB d1aa71049342
model-00024-of-00027.safetensorsWeights3.7 GB 344a4a809cf4
model-00025-of-00027.safetensorsWeights3.5 GB 3cca0405ecc8
model-00026-of-00027.safetensorsWeights3.7 GB 4b92275e0669
model-00027-of-00027.safetensorsWeights2.2 GB ae7440e308ff
model-vision-graft.safetensorsWeights912.1 MB b47e83150554
mtp-head-q6.safetensorsWeights5.8 GB 1563d297c7bc
config.jsonConfiguration148.7 KB —
generation_config.jsonConfiguration244 B —
model.pyConfiguration262.3 KB —
model.safetensors.index.jsonConfiguration288.6 KB —
preprocessor_config.jsonConfiguration390 B —
skipzero_load.pyConfiguration6.3 KB —
video_preprocessor_config.jsonConfiguration385 B —
vqlab_provenance.jsonConfiguration8.1 KB —
README.mdDocumentation20.4 KB —
chat_template.jinjaOther7.8 KB —
mtp-head-q6.safetensors.fp32norms-bakOther5.8 GB a5cad76f094e
vqlab_provenance.history.jsonlOther52.8 KB —
.gitattributesRepository1.6 KB —
tokenizer.jsonTokenizer20.0 MB 06b9509352d2
tokenizer_config.jsonTokenizer1.2 KB —

License and Download

License
apache-2.0
Access
Open weights, no gate
Download size
102.0 GB
Download from Noah Zelezny

Released by Noah Zelezny through its official repository on Hugging Face. Read the license.

Built From

  • Derived from Qwen/Qwen3.5-397B-A17B
  • Quantized from Qwen/Qwen3.5-397B-A17B

Memory Requirements

PrecisionWeights in memory
As published102.0 GB
16-bit125.1 GB
8-bit62.6 GB
4-bit31.3 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About Qwen3.5-397B-A17B-VQ-2.2bpw

How much GPU memory does Qwen3.5-397B-A17B-VQ-2.2bpw need?

About 150.2 GB at 16-bit and 37.5 GB at 4-bit: the weights (62.6B parameters) plus a working margin. A long context needs more.

What is the cheapest GPU to run Qwen3.5-397B-A17B-VQ-2.2bpw on?

At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use Qwen3.5-397B-A17B-VQ-2.2bpw commercially?

Yes. Qwen3.5-397B-A17B-VQ-2.2bpw is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

What is Qwen3.5-397B-A17B-VQ-2.2bpw's context length?

262,144 tokens, from the maximum position embeddings in its published configuration.

Similar Models

This checkpoint was structurally pruned with the released Less-is-MoE mean-absolute-gradient method. It removes exactly 50% of routed-expert FFN neurons using 64 calibration samples from the gpqamain configuration of Idavidrein/gpqa revision 633f5ee89ab8ad4522a9f850766b73f62147ffdd. The released loader settings are preserved: train, Question plus shuffled choices, Explanation, selectionseed=1234, BF16, and no optimizer step. The samples are full length: no tokenizer maxlength, truncation, or padding. The longest input for this tokenizer is 1,632 tokens. The source checkpoint was loaded and pruned in BF16. The source-row selection hash is…

Open weights apache-2.0 64.1B parameters 262,144 tokens
View model

This checkpoint was structurally pruned with the released Less-is-MoE mean-absolute-gradient method. It removes exactly 50% of routed-expert FFN neurons using 64 calibration samples from the gpqamain configuration of Idavidrein/gpqa revision 633f5ee89ab8ad4522a9f850766b73f62147ffdd. The released loader settings are preserved: train, Question plus shuffled choices, Explanation, selectionseed=1234, BF16, and no optimizer step. The samples are full length: no tokenizer maxlength, truncation, or padding. The longest input for this tokenizer is 1,632 tokens. The source checkpoint was loaded and pruned in BF16. The source-row selection hash is…

Open weights apache-2.0 64.1B parameters 262,144 tokens
View model

This checkpoint was structurally pruned with the released Less-is-MoE mean-absolute-gradient method. It removes exactly 50% of routed-expert FFN neurons using 64 calibration samples from the gpqamain configuration of Idavidrein/gpqa revision 633f5ee89ab8ad4522a9f850766b73f62147ffdd. The released loader settings are preserved: train, Question plus shuffled choices, Explanation, selectionseed=1234, BF16, and no optimizer step. The samples are full length: no tokenizer maxlength, truncation, or padding. The longest input for this tokenizer is 1,632 tokens. The source checkpoint was loaded and pruned in BF16. The source-row selection hash is…

Open weights apache-2.0 64.1B parameters 262,144 tokens
View model

Model · Text generation

Qwen3.5-397B-A17B-VQ-3.1bpw

Noah Zelezny

125.3 GiB text weights — the quality build. Also ships the bf16 vision tower (+0.85 GiB) and an optional MTP draft head (+5.4 GiB); full download 131.6 GiB. A vector-quantized build of Qwen3.5-397B-A17B for machines with memory to spend: the strongest quantization we know how to make of this model at this size, on stock mlx-lm, no patches. 12.43% of the 64-element weight groups in the Qwen3.5-397B teacher sit in output rows whose weights are about 1e-29. The vq-skipzero format drops those rows' codes and scales on disk and in memory; they output exact zeros, and the live rows are byte-identical to the previous revision (vqlab sz-check, every module). Text weights go from 141.71 to 125.31…

Open weights apache-2.0 64.5B parameters 262,144 tokens mlx
View model

Model · Text generation

Qwen3.5-122B-A10B-NVFP4

NVIDIA

The NVIDIA Qwen3.5-122B-A10B-NVFP4 model is the quantized version of Alibaba's Qwen3.5-122B-A10B model, which is an auto-regressive language model that uses an optimized transformer architecture. For more information, please check here. The NVIDIA Qwen3.5-122B-A10B NVFP4 model is quantized with Model Optimizer. This model is ready for commercial/non-commercial use. This model is not owned or developed by NVIDIA. This model has been developed and built to a third-party’s requirements for this application and use case; see link to Non-NVIDIA (Qwen3.5-122B-A10B) Model Card from Alibaba. Global Developers looking to take off-the-shelf, pre-quantized models for deployment in AI Agent systems…

Open weights apache-2.0 64.6B parameters 262,144 tokens Model Optimizer
View model

This checkpoint was structurally pruned with the released Less-is-MoE mean-absolute-gradient method. It removes exactly 50% of routed-expert FFN neurons using 64 calibration samples from the gpqamain configuration of Idavidrein/gpqa revision 633f5ee89ab8ad4522a9f850766b73f62147ffdd. The released loader settings are preserved: train, Question plus shuffled choices, Explanation, selectionseed=1234, BF16, and no optimizer step. The samples are full length: no tokenizer maxlength, truncation, or padding. The longest input for this tokenizer is 1,511 tokens. The source MXFP4 checkpoint was explicitly dequantized to BF16 before scoring and pruning. The source-row selection hash is…

Open weights apache-2.0 59.5B parameters 131,072 tokens
View model