Skip to content
AIBRAINWORX / OpenSourceCoder.comLOCAL-FIRST. BUILT IN THE OPEN.
Text & Coding LLMs / Multimodal & Vision

Qwen Releases Apache 2.0 Qwen3.8-27B Dense Multimodal Coding Model

Qwen has published open weights for Qwen3.8-27B under the Apache 2.0 license, pairing a hybrid Gated DeltaNet and Gated Attention architecture with native vision-language and agentic coding support.

Listen to this articleBrowser Voice
Ready
Official company logo or product image for qwen-releases-apache-2-0-qwen3-8-27b-dense-multimodal-coding-model
Official image or organization logo from huggingface.co; identification only, no endorsement.
Event / context: 2026-08-05. Source check: October 28, 2026.

Qwen has released open weights for Qwen/Qwen3.8-27B on Hugging Face under the Apache 2.0 license, introducing a 27-billion-parameter dense model aimed at agentic coding, research, and multimodal workflows. Built as a native vision-language system rather than a text-only checkpoint, the post-trained release combines image and video comprehension with configurable reasoning controls for open-source inference stacks.

Core Announcement and Operational Scope

Positioned as the next step after the Qwen3.5 and Qwen3.6 model families, Qwen3.8-27B brings Qwen's latest generation to a compact, dense parameter footprint designed for self-hosted and cloud deployments alike. The repository provides weights and configuration files in Hugging Face Transformers Safetensors format for the post-trained model, targeting both conversational interfaces and multi-step autonomous execution. According to the model card, Qwen3.8-27B is designed to process text alongside visual inputs, spanning STEM diagrams, structured documents, and hour-scale videos. Developers can also govern how the model allocates compute to internal reasoning: thinking mode is enabled by default, can be toggled off per request, and supports tunable depth via a reasoning_effort parameter alongside a preserve_thinking option to retain reasoning context across multi-turn histories.

Hybrid Architecture and Reported Benchmarks

Under the hood, Qwen3.8-27B uses a Causal Language Model paired with a Vision Encoder and retains the qwen3_5 architectural lineage. The language backbone spans 64 layers with a hidden dimension of 5,120, a padded vocabulary embedding and output size of 248,320 tokens, and a feed-forward network intermediate dimension of 17,408. Instead of uniform quadratic attention across every layer, Qwen arranges the 64 layers into 16 repeating blocks, each consisting of three Gated DeltaNet plus FFN layers followed by one Gated Attention plus FFN layer. Within Gated DeltaNet, the model uses 48 linear attention heads for values and 16 for queries and keys with a head dimension of 128. The Gated Attention layers use 24 query heads, 4 key-value heads, a 256 head dimension, and a 64-dimension Rotary Position Embedding. The model also incorporates multi-step Multi-Token Prediction (MTP) and supports a 262,144-token native context window, extensible to 1,000,000 tokens.

On text and coding evaluations published by Qwen on the model card, Qwen3.8-27B posts substantial gains over its predecessor Qwen3.6-27B. Using the Claude Code evaluation harness at temperature 1.0, top_p 0.95, and a 256K context window, Qwen reports 61.7 on SWE-bench Pro (compared with 53.5 for Qwen3.6-27B and 57.6 for Qwen3.7-Plus) and 42.2 on DeepSWE 1.1 (up from 13.3). On NL2Repo-Bench—evaluated with repository-access commands such as git clone, pip install, and pip download disabled to curb reward hacking—Qwen reports 42.3. Additional developer-reported figures include 73.0 on Terminal Bench 2.1 (Terminus), 90.3 on LiveCodeBench v6, 79.0 on Qwen's in-house QwenSWEBench, 70.7 on CoWorkBench, 33.4 on JobBench, 79.5 on IFBench, 89.2 on GPQA Diamond, 30.8 on HLE, and 20.4 Pass@1 (42.9 Score) on Agents' Last Exam.

Developer Workflow and Serving Integration

For open-source engineering teams, the release arrives with day-one compatibility across mainstream Python and container serving runtimes. In Hugging Face Transformers, developers can invoke the image-text-to-text pipeline directly or load the checkpoint with AutoProcessor and AutoModelForMultimodalLM alongside device_map auto-placement. High-throughput serving is documented for both vLLM and SGLang, each exposing OpenAI-compatible chat completion endpoints that accept mixed text and image URL payloads. Local and containerized workflows are further supported via Docker Model Runner, TokenSpeed, and community-oriented quantization paths targeting llama.cpp, Ollama, and LM Studio. For teams preferring managed infrastructure, Qwen notes that Qwen Cloud will offer a hosted Qwen3.8-27B service featuring a default 1-million-token context window and official built-in tools, though that hosted tier is listed as coming soon.

Licensing, Evaluation, and Infrastructure Caveats

While the Apache 2.0 license on the Hugging Face weights gives developers permissive commercial and modification rights over the open checkpoint, several practical caveats apply. First, all performance numbers currently originate from Qwen's own model card rather than independent third-party audits, and the repository's referenced vision-language benchmark results were not yet detailed in the primary text excerpt. Furthermore, Qwen notes that QwenSWEBench is an in-house benchmark, and that SWE-bench Pro baselines were re-evaluated after correcting problematic tasks—except for Opus4.6 Max, whose score reflects an officially reported figure rather than an identical harness run. Hardware demands also remain nontrivial despite the compact 27B parameter count: Qwen's SGLang Docker guidance assumes multi-GPU execution with 32 GB of shared memory, meaning local workstations will likely rely on quantized builds or shortened context windows until memory-optimized runtimes mature. Finally, features reserved for the upcoming Qwen Cloud endpoint, such as built-in tools and out-of-the-box 1M-token defaults, remain separate from the self-hosted Apache 2.0 release.

Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-20.

Follow your curiosity.

Pass the signal.