Sunday, Aug 23 | --:--
Back to home

Alibaba Releases Qwen3.8-27B Open Weights Under Apache 2.0

On August 14, 2026, Alibaba’s Qwen team posted Qwen3.8-27B on Hugging Face—a 27-billion-parameter native vision-language dense model with 262K context (extensible to 1M), thinking mode by default, and Apache 2.0 weights—the local companion to July’s 2.4T Qwen3.8-Max preview.

Tech Insights Reporter 5 min read Hangzhou / Shanghai
Cover illustration for Alibaba Releases Qwen3.8-27B Open Weights Under Apache 2.0

TLDR

Alibaba’s Qwen team released Qwen3.8-27B weights at Qwen/Qwen3.8-27B on August 14, 2026 (ModelScope metadata 15:00 UTC). The checkpoint is a 27B dense native vision-language model (images and video), Apache 2.0, 262,144 native context (YaRN to 1,000,000). Thinking mode is on by default, with reasoning_effort at xhigh / medium / low. This is the single-GPU companion promised when Qwen3.8-Max-Preview (2.4T) launched July 19—not a dual-file of that flagship.

What the model card says

Item Official card
Type Causal LM + vision encoder; 64 layers; hybrid Gated DeltaNet + Gated Attention
Params 27B (card); HF filesize line also says 28B params in BF16
Context 262,144 native; 1M via YaRN
License Apache 2.0
Vendor SWE-bench Pro 61.7 (Claude Code harness; Qwen notes Opus 4.6 Max official 53.4 on their re-eval table)
Vendor DeepSWE 1.1 42.2
Vendor OSWorld-Verified 84.3
Hosted Qwen Cloud version “coming soon,” 1M context + built-in tools

Qwen’s table is lab-run. Several rows use in-house benches (QwenSWEBench, CoWorkBench, RecreationBench). Treat 61.7 SWE-bench Pro as vendor-claimed until independent boards replicate.

Independent rankings

Source Snapshot Standing
Hugging Face card / ScaleAI SWE-bench Pro eval widget Launch card Displays 61.7 (same as vendor table; asterisked)
Artificial Analysis / LMArena Research window Aug 17 No dated Index or Elo cited here for 3.8-27B — not yet independently ranked in those boards at write time

Do not invent an AA Intelligence Index number for 27B.

Why this story matters

The 2.4T Max preview was a datacenter artifact. 27B Apache multimodal is what actually runs on a workstation next to Muse Glimmer 30B (Aug 10). If even a fraction of the vendor SWE-Pro / OSWorld numbers hold, Qwen just dropped the default local agent backbone for the rest of 2026. Watch: independent DeepSWE and Arena votes, and whether Max weights follow or stay API-only.

Sources

Prior Coverage

Earlier Times of AI reporting on this thread.

Scroll to continue reading