Sulat.com
AI models
$10 off the fastest DeepSeek V4.1 Flash, Kimi K3 and GLM 5.3 from Synthetic
AMD logo

Model details

Qwen3.8 27B

Qwen3.8-27B is a compact, dense member of the Qwen open-model family, released by the Qwen organization with post-trained weights published on Hugging Face for use with Transformers, vLLM, SGLang, and TokenSpeed. It is built on the architectural foundation of Qwen3.5 and is described as a native vision-language model that can understand images and videos, giving it broader perceptual reach than text-only checkpoints of similar scale. The open-weights release is positioned as the most capable Qwen generation to date, with improvements aimed at coding, professional work, research workflows, and long-horizon agentic tasks where reliable multi-step planning and execution matter.

Practically, Qwen3.8-27B is aimed at developers who want a 27B-scale dense model that can carry complex, multi-step jobs through to completion while remaining straightforward to deploy. The weights release supports flexible thinking control and stronger autonomous planning with better handling of environment feedback, which suits agent-style pipelines and tool-augmented applications. Alongside the open artifacts, Qwen has announced a hosted Qwen Cloud variant coming soon with a 1M-token default context and official built-in tools, signaling that the same weights are intended to span both local deployment and managed production use.

AMDQwen3.8-27Bqwen

Quick Info

Powered by
Provider
AMD
Model key
Qwen3.8-27B
Release date
Aug 14, 2026
Last updated
Aug 14, 2026
Input modalities
Output modalities
Capabilities

Cost

A provider subscription or plan supersedes token-based pricing for this model.

Limits

Output tokens
32,768 tokens
Context window
131,072 tokens

Latest news about Qwen3.8 27B

AMD

Official sourceAnnouncement

AMD announced Day 0 support for Qwen3.8 27B on August 14, 2026, allowing developers to run the dense Qwen-family model locally on AMD Ryzen AI Max+ processor-based PCs and a single AMD Radeon AI PRO R9700 32 GB graphics card via llama.cpp, with additional support on AMD hardware carrying more than 24 GB of Variable Gra Early testing reported by AMD reached up to 24.5 tokens per second on an AMD Ryzen AI Max+ 395 and up to 51.8 tokens per second on a single AMD Radeon AI PRO R9700, measured on Windows through llama.cpp with the Vulkan backend using MTP=4 on Ryzen AI Max+ 395 and MTP=2 on Radeon AI PRO R9700 across three or more runs.

AMD

CoverageRelease Notes

DataNorth confirms that Alibaba released Qwen3.8-27B on 14 August 2026 as an open-weights dense model accepting text, images, and video under the Apache 2.0 licence. The article reports a 262,144-token native context window that Alibaba says extends to 1,000,000 tokens, and positions Qwen3.8-27B as the smaller companio The article clarifies that Qwen3.8-27B is a native vision-language model rather than a text model with an adapter, noting the published configuration file includes a full vision encoder with the language-model-only flag set to false — correcting earlier coverage that described the open weights as text-only. Alibaba pos

AMD

CoverageRelease Notes

Alibaba's Qwen team officially unveiled Qwen3.8-27B as a new open-weight addition to the Qwen3.8 series, released under the Apache 2.0 license. The blog states the model is a native multimodal dense checkpoint designed to balance high performance with cost-efficiency, with Alibaba noting it can run on consumer-grade ha The page highlights that Qwen3.8-27B is a native vision-language model handling images and videos while offering flexible thinking control, and ships with a 262K-token native context window that extends to 1 million tokens. Alibaba reports that within two days of release the model ranked among the top 5 most-liked mode

AMD

CoverageAnalysis

Local AI Zone published a comprehensive technical analysis of Qwen3.8-27B on August 15, 2026, detailing a 27,781,427,952-parameter dense multimodal architecture from Alibaba's Tongyi Lab released under Apache 2.0 on August 14, 2026. The analysis specifies 64 Transformer blocks with a 5,120 hidden dimension, 24 query he The article documents a hybrid attention design with a 3:1 ratio of Gated DeltaNet linear attention (48 layers, 75%) to gated full attention (16 layers, 25%), which the author credits with reducing computational complexity from O(n²) to O(n) for the majority of layers. It frames the release as a step that uses architec

AMD

CoverageBenchmark

Alibaba's Qwen team released Qwen3.8-27B at 15:00 UTC on August 14, 2026, as a 27.78-billion-parameter dense multimodal checkpoint shipping under Apache 2.0 with text, image, and video inputs and a native 262,144-token context window. The launch-day review reports Qwen-stated gains over Qwen3.6-27B on agentic coding, c The article applies critical scrutiny to the launch numbers, flagging that every score is Qwen-reported, that several benchmarks are in-house or modified, and that the SWE-bench Pro comparison imports Anthropic's Opus result rather than rerunning it. It also warns that the most dramatic local-hardware claims ignore KV

Videos about Qwen3.8 27B

More models around Qwen3.8 27B