FULL STORY

Qwen3.8-27B: From Leak to Open-Source Release

Alibaba's Qwen3.8-27B went from Reddit leaks and official previews to a full open-source release under Apache 2.0 on August 14, with the company claiming it comprehensively outperforms Qwen3.7-Plus.

2026-08-11 ~ 2026-08-15 · 3 episodes · 61 posts

Episode 1 · Qwen3.8-27B Model Confirmed for September 3 Release (2026-08-11, 2 posts)

Alibaba's Qwen team has confirmed the release of the new Qwen3.8-27B model. According to ModelScope, the model is scheduled to launch on September 3 at 6:00 AM (UTC+8).

Episode 2 · Alibaba Teases Qwen3.8-27B with Unmatched Intelligence Density (2026-08-12, 11 posts)

Alibaba's Qwen team teases the upcoming Qwen3.8-27B model with a Hugging Face pre-release page, promising 'unmatched intelligence density'. The official account confirms improved agentic coding capabilities and announces a livestream for Friday at 10 AM UTC+8 on YouTube and X. Community interest is high, with over 3,300 people on the waitlist, and HuggingFace's official account has joined the hype.

Confirmed

  • The Qwen team confirms the upcoming Qwen3.8-27B model with improved agentic coding capabilities.
  • A livestream is scheduled for Friday at 10 AM UTC+8 on YouTube and X to share behind-the-scenes stories and discuss multimodal tech from demo to production.
  • A Hugging Face pre-release page is live, describing the model as 'an update to the beloved Qwen model, offering unmatched intelligence density'.
  • The waitlist has exceeded 3,300 people; HuggingFace's official account jokingly compares the hype to the popular Corgi cafe.

Unconfirmed

  • A parody page claims a release date of August 14, 2026; the actual release date and performance await official announcement and livestream.

Why it matters

  • This follows the open-sourcing of the first Qwen-Max-level model, marking another significant step. Improved coding capabilities could further drive open-source agentic applications.

Episode 3 · Alibaba Open-Sources Native Multimodal Qwen3.8-27B Model (2026-08-13, 48 posts)

Alibaba's Qwen team has open-sourced the Qwen3.8-27B, a native multimodal dense model with 27B parameters under the Apache 2.0 license. The team claims it comprehensively outperforms Qwen3.7-Plus, excelling in coding and office workflows, supports 262K native context extendable to 1M, and quickly topped the Hugging Face trending charts. The community has released various quantization options for local deployment, and third-party frameworks like SGLang and Ollama have rapidly integrated support.

Confirmed

  • Model Architecture and Performance: Qwen3.8-27B is a 27B parameter dense hybrid attention model (48/64 layers use linear attention), natively supporting image-to-text tasks. It supports 262K native context, extendable to 1M via YaRN, includes a built-in MTP draft head, and is supported by vLLM 0.17.0.
  • Deployment and Quantization: The official FP8 quantized version is available on Hugging Face. Unsloth released GGUF and NVFP4 quantized weights, with the NVFP4 version also trending on HF; Ollama announced integration, supporting tools like Claude Code and OpenCode with optimizations for Apple Silicon.
  • Inference Support: SGLang achieved 200+ tok/s decoding speeds on RTX 5090 and RTX Pro 6000 using NVFP4 quantization, and it is also deployable on DGX Spark.
  • Community Feedback: The initial model card lacked benchmark results, but @Tccybo subsequently compiled the scores into comparison charts.

Unconfirmed

  • Performance Comparisons: Community opinions suggest it rivals Claude Opus 4.6. @kimmonismus stated it surpasses Claude Opus 4.6 Max in agentic coding and computer use benchmarks; @de4dee relayed user views comparing it to Opus 4.6, though officials have not made this claim directly.
  • Hardware Expectations: @OddTumbleweed574 noted it as the first sub-30B parameter model runnable on 3090 or M4 Pro, predicting a shift of more workflows to local runs by the end of the year.

Timeline

  • Aug 13: Users spotted the model on ModelScope.
  • Aug 14: Qwen officially released the FP8 quantized version on Hugging Face; the model card was published but benchmarks were pending.
  • Aug 15: Benchmark scores were released, community summary charts were posted, and Ollama and SGLang announced integration or support.

28 more related posts →