Skip to page content
SAVRN
Search Plan a project

Open-weight model · Image and text to text

Qwen3.8-27B-PARO-MXFP4-v2

by Ms MisterSnrub/Qwen3.8-27B-PARO-MXFP4-v2

Qwen3.8-27B-PARO-MXFP4-v2 is an open-weight model for image and text to text from Ms, released under Apache License 2.0. It has 27.8B parameters and a 262,144-token context. At 16-bit it needs about 66.8 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index.

Qwen3.8-27B in MXFP4 for one AMD Radeon AI PRO R9700 (32 GB). Needs GGZ14's vLLM-Radiance fork plus our patches (quantmethod: paroquantmxfp4; stock vLLM and transformers can't load it).

Parameters27.8B
Context262,144
Weights17.2 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads—

Runs On

What it takes to serve Qwen3.8-27B-PARO-MXFP4-v2 (27.8B 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 55.6 GB 66.8 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x H100 $1.99 · 1x MI325X $2.00
8-bit 27.8 GB 33.4 GB 1x MI300X (192 GB)
Vultr		 $1.85 1x H100 $1.99 · 1x MI325X $2.00
4-bit 13.9 GB 16.7 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 16, 2026.

Qwen3.8-27B-PARO-MXFP4-v2 on every accelerator the SAVRN Index prices, at every precision

Model Card

By Ms, published under apache-2.0, revision 34e954e00fca.

Qwen3.8-27B in MXFP4 for one AMD Radeon AI PRO R9700 (32 GB). Needs GGZ14's vLLM-Radiance fork plus our patches (quantmethod: paroquantmxfp4; stock vLLM and transformers can't load it). Setup: github.com/mtstanfield/vllm-mxfp4, branch r9700-tp1, R9700-HANDOFF.md. MTP speculative decoding with 4 draft tokens. Full BetterBench report and setup: R9700-HANDOFF.md on the branch. pairwise rotations per 128-channel group, retrained for the MXFP4 grid from z-lab's int4 rotations. - GPTQ onto the MXFP4 grid in the rotated basis, with per-block scale search. - lmhead fp8 per-channel, embedtokens fp8 per-row, MTP head fp8. extras/qwen38-paro-v2-mtp-gptq-a05.pt is the MTP head requantized to MXFP4 with…

Read Ms's full model card

Qwen3.8-27B in MXFP4 for one AMD Radeon AI PRO R9700 (32 GB). Needs GGZ14's vLLM-Radiance fork plus our patches (quant_method: paroquant_mxfp4; stock vLLM and transformers can't load it). Setup: github.com/mtstanfield/vllm-mxfp4, branch r9700-tp1, R9700-HANDOFF.md.

Measured on one R9700 (TP=1, 300 W)
Decode, 1 stream (BetterBench 0.6.0, weighted) 102.3 tok/s (median over all tasks 108.1)
Prefill 3,357 tok/s at 5.9k tokens, 3,000 at 47k
4 concurrent streams 278 tok/s total, 85 per stream
8 concurrent streams (64k context variant) 426 tok/s total
KV cache 375,633 tokens fp8, max context 262,144
Served perplexity (fp8 KV) wikitext-2 5.970, code 2.304
KL vs bf16 on agent transcripts, vs plain round-to-nearest MXFP4 64-68% lower

MTP speculative decoding with 4 draft tokens. Full BetterBench report and setup: R9700-HANDOFF.md on the branch.

Recipe

  • Decoder linears: MXFP4 (e2m1, e8m0 scale per 32) in a rotated basis, using ParoQuant pairwise rotations per 128-channel group, retrained for the MXFP4 grid from z-lab's int4 rotations.
  • GPTQ onto the MXFP4 grid in the rotated basis, with per-block scale search.
  • Calibration: 70% coding-agent transcripts (private) + 30% wikitext-2 train.
  • lm_head fp8 per-channel, embed_tokens fp8 per-row, MTP head fp8. extras/qwen38-paro-v2-mtp-gptq-a05.pt is the MTP head requantized to MXFP4 with GPTQ (goes in $CACHE/mtp_gptq/).
  • Served as W4A8: per-token fp8 activations.

Credits: Qwen (base model), z-lab (ParoQuant), GGZ14 (vLLM-Radiance), StillDeadcode (libr4d).

Configuration

Architecture
Qwen3_5ForConditionalGeneration
Context length (tokens)
262,144
Layers
64
Hidden size
5,120
Feed-forward size
17,408
Attention heads
24
Key/value heads
4
Head dimension
256
Vocabulary size
248,320
Model type
qwen3_5
Quantization
paroquant_mxfp4

Identity and Version

Repository
MisterSnrub/Qwen3.8-27B-PARO-MXFP4-v2
Publisher
Ms
Task
Image and text to text
Modality
Image and text
Library
Not stated by the source
Parameters
27.8B parameters
Languages
Not stated by the source
Revision
34e954e00fcaa32267f13b861a10d33b1c744089
First published
2026-09-29
Last updated
2026-09-29

Files and Weights

35 files, 17.2 GB in total. The weights are 22 files totalling 17.2 GB in pt, safetensors.

Weights22 files · 17.2 GB
Configuration5 files · 273.7 KB
Tokenizer4 files · 22.9 MB
Documentation1 file · 1.9 KB
Other2 files · 9.2 KB
Repository1 file · 1.6 KB
Every file
FileTypeSizeSHA-256
extras/qwen38-paro-v2-mtp-gptq-a05.ptWeights225.6 MB 943d18921d53
model-00001-of-00018.safetensorsWeights1.7 GB 1acb36f13d36
model-00002-of-00018.safetensorsWeights814.7 MB 5596bc734bdd
model-00003-of-00018.safetensorsWeights40 B 60d95b10b6e1
model-00004-of-00018.safetensorsWeights1.1 GB 77b0ecb9c1df
model-00005-of-00018.safetensorsWeights562.7 MB 2212c219eef3
model-00006-of-00018.safetensorsWeights1.1 GB f8611dc2ebdd
model-00007-of-00018.safetensorsWeights564.3 MB 3cbf99406c87
model-00008-of-00018.safetensorsWeights1.1 GB 80a1e15865d9
model-00009-of-00018.safetensorsWeights564.3 MB 071dca16a1c1
model-00010-of-00018.safetensorsWeights1.1 GB 8e55ce7392f6
model-00011-of-00018.safetensorsWeights564.3 MB 5894695dd3ba
model-00012-of-00018.safetensorsWeights1.1 GB 4f6c777bca6c
model-00013-of-00018.safetensorsWeights564.3 MB 3b3b6943847e
model-00014-of-00018.safetensorsWeights1.1 GB a8193e33e3ec
model-00015-of-00018.safetensorsWeights564.3 MB de527f63a44e
model-00016-of-00018.safetensorsWeights1.1 GB 2c8b922213e0
model-00017-of-00018.safetensorsWeights564.3 MB eb0d8f7b6399
model-00018-of-00018.safetensorsWeights40 B 60d95b10b6e1
model-embed-fp8.safetensorsWeights1.3 GB 66ae83d6c398
model-lmhead-fp8.safetensorsWeights1.3 GB 49601b615f67
model-mtp.safetensorsWeights425.0 MB 2b1166af2436
config.jsonConfiguration4.0 KB —
generation_config.jsonConfiguration202 B —
model.safetensors.index.jsonConfiguration268.7 KB —
preprocessor_config.jsonConfiguration390 B —
video_preprocessor_config.jsonConfiguration385 B —
README.mdDocumentation1.9 KB —
chat_template.jinjaOther9.0 KB —
crc32.txtOther238 B —
.gitattributesRepository1.6 KB —
merges.txtTokenizer3.4 MB —
tokenizer.jsonTokenizer12.8 MB 0997f410c57a
tokenizer_config.jsonTokenizer17.9 KB —
vocab.jsonTokenizer6.7 MB —

License and Download

License
apache-2.0
Access
Open weights, no gate
Download size
17.2 GB
Download from Ms

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

Built From

Memory Requirements

PrecisionWeights in memory
As published17.2 GB
16-bit55.6 GB
8-bit27.8 GB
4-bit13.9 GB

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

Questions About Qwen3.8-27B-PARO-MXFP4-v2

How much GPU memory does Qwen3.8-27B-PARO-MXFP4-v2 need?

About 66.8 GB at 16-bit and 16.7 GB at 4-bit: the weights (27.8B parameters) plus a working margin. A long context needs more.

What is the cheapest GPU to run Qwen3.8-27B-PARO-MXFP4-v2 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 Qwen3.8-27B-PARO-MXFP4-v2 commercially?

Yes. Qwen3.8-27B-PARO-MXFP4-v2 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 Qwen3.8-27B-PARO-MXFP4-v2's context length?

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

Similar Models

Model · Image and text to text

Qwen3.6-27B-FP8

Qwen

Following the February release of the Qwen3.5 series, we're pleased to share the first open-weight variant of Qwen3.6. Built on direct feedback from the community, Qwen3.6 prioritizes stability and real-world utility, offering developers a more intuitive, responsive, and genuinely productive coding experience. This release delivers substantial upgrades, particularly in For more details, please refer to our blog post Qwen3.6-27B. Empty cells (--) indicate scores not yet available or not applicable. For streamlined integration, we recommend using Qwen3.6 via APIs. Below is a guide to use Qwen3.6 via OpenAI-compatible API. Qwen3.6 can be served via APIs with popular inference frameworks. In…

Open weights apache-2.0 27.8B parameters 262,144 tokens transformers
View model

Model · Image and text to text

Qwen3.8-27B

Qwen

Following the widespread community adoption of the Qwen3.5 and Qwen3.6 series, we are pleased to introduce Qwen3.8, the most capable generation in the Qwen open-model family to date. Built on the architectural foundation of Qwen3.5, Qwen3.8 delivers substantial gains across coding, professional work, research, and long-horizon agentic tasks. Qwen3.8-27B brings these advances to a compact, deployment-friendly dense model: a native vision-language model that understands images and videos, with flexible thinking control, designed to carry complex, multi-step tasks through to completion with greater reliability. For streamlined integration, we recommend using Qwen3.8 via APIs. Qwen3.8 can be…

Open weights apache-2.0 27.8B parameters 262,144 tokens transformers
View model

Model · Image and text to text

Qwen3.8-27B-FP8

Qwen

Following the widespread community adoption of the Qwen3.5 and Qwen3.6 series, we are pleased to introduce Qwen3.8, the most capable generation in the Qwen open-model family to date. Built on the architectural foundation of Qwen3.5, Qwen3.8 delivers substantial gains across coding, professional work, research, and long-horizon agentic tasks. Qwen3.8-27B brings these advances to a compact, deployment-friendly dense model: a native vision-language model that understands images and videos, with flexible thinking control, designed to carry complex, multi-step tasks through to completion with greater reliability. For streamlined integration, we recommend using Qwen3.8 via APIs. Qwen3.8 can be…

Open weights apache-2.0 27.8B parameters 262,144 tokens transformers
View model

Model · Image and text to text

Qwen3.6-27B

Qwen

Following the February release of the Qwen3.5 series, we're pleased to share the first open-weight variant of Qwen3.6. Built on direct feedback from the community, Qwen3.6 prioritizes stability and real-world utility, offering developers a more intuitive, responsive, and genuinely productive coding experience. This release delivers substantial upgrades, particularly in For more details, please refer to our blog post Qwen3.6-27B. Empty cells (--) indicate scores not yet available or not applicable. For streamlined integration, we recommend using Qwen3.6 via APIs. Below is a guide to use Qwen3.6 via OpenAI-compatible API. Qwen3.6 can be served via APIs with popular inference frameworks. In…

Open weights apache-2.0 27.8B parameters 262,144 tokens transformers
View model

Model · Image and text to text

Qwen3.5-27B

Qwen

Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance. Qwen3.5 represents a significant leap forward, integrating breakthroughs in multimodal learning, architectural efficiency, reinforcement learning scale, and global accessibility to empower developers and enterprises with unprecedented capability and efficiency. For more details, please refer to our blog post Qwen3.5. WMT24++: a harder subset of WMT24 after difficulty labeling and rebalancing; we report the averaged scores on 55 languages using XCOMET-XXL. Empty cells (--) indicate scores not yet available or not applicable. Empty cells (--) indicate scores not…

Open weights apache-2.0 27.8B parameters 262,144 tokens transformers
View model

Model · Image and text to text

JEV-27B-VL

AutoTrust AI Lab

The benchmark. ERQA (Embodied Reasoning Question Answering) is the benchmark Google DeepMind released with Gemini Robotics (March 2025) to measure whether a vision-language model understands the physical world a robot acts in. It has 400 multiple-choice questions about real robot and everyday scenes, with 1–16 images per question (287 questions have one image). 382 questions have 4 options and the rest 2–3, so chance is about 26%. ERQA is one of the standard embodied-reasoning benchmarks; Gemini Robotics-ER, PhysBrain and Hy-Embodied report it. With 400 questions the 95% interval of a score is about ±5 points. How we test. Each question is one choice decision through POST /v1/decide: the…

Open weights apache-2.0 27.8B parameters 262,144 tokens transformers
View model