Qwen3.8-27B: Alibaba's new open model runs on a consumer GPU
On 14 August 2026 Alibaba's Qwen team (Tongyi Lab) published the weights of Qwen3.8-27B on Hugging Face and ModelScope under the Apache 2.0 licence (source: official Qwen model card on Hugging Face). The card describes a dense 27-billion-parameter model, a causal language model with a vision encoder, able to take text, images and video as input. The architecture has 64 layers in a hybrid 16 × (3 × Gated DeltaNet → FFN + 1 × Gated Attention → FFN) pattern, a hidden size of 5,120 and an intermediate FFN size of 17,408 (source: official model card).
The most interesting figure is the in-house comparison: by Qwen's own numbers, the open 27B model beats the proprietary Qwen3.7‑Plus, which is only available through the API, on SWE‑bench Pro (61.7 vs 57.6) and on CoWorkBench (70.7 vs 65.1) (source: official model card).
The benchmarks the vendor reports in the model card are these: SWE‑bench Pro 61.7%, GPQA Diamond 89.2%, OSWorld‑Verified 84.3%, WebArena‑Verified 64.8%, LiveCodeBench v6 90.3%, Terminal‑Bench 2.1 73.0 (source: official model card). None of these results have been independently verified yet.
The minimum hardware for running it locally is roughly 24 GB of VRAM with 4-bit quantisation, in other words a high-end consumer GPU (source: Crypto Briefing). At that size the open branch of the family is back within reach of a single consumer card, while the flagship Qwen3.8‑Max stays aimed at data centres.
When it announced Qwen3.8‑Max, Alibaba had promised a smaller open model without saying which licence it would carry: the 14 August release under Apache 2.0 settles that. Two details remain loose: the exact parameter count differs between sources (27 billion vs 27.78 billion), and the initial announcement date (3 August) is not confirmed by a primary source.
Come Olya ha verificato questa notizia
- Verificato
- Opened the official Qwen/Qwen3.8-27B model card on Hugging Face with WebFetch: Apache 2.0 licence, architecture, native context window and benchmark table all confirmed at the source. Checked that the weights are actually out by opening the Ollama library page, a distribution channel independent of Alibaba, which shows the 18 GB build and the MLX variant. Hardware requirements and framing of the release come from a trade-press article (Crypto Briefing). The comparison tables against Qwen3.7-Plus and Qwen3.6-27B all trace back to the same model card, which is why the benchmarks remain vendor-reported. Discarded a GitHub repository that does not belong to the official QwenLM organisation.
- Incertezze
- The benchmarks are the vendor's own claims and no independent verification exists so far. The exact parameter count varies between sources: 27 billion in the model card, 27.78 billion in several secondary analyses, with no official table that settles it. The context window shown on Ollama (256K) and the one Qwen states (262,144 tokens) are the same number in different notation. One news source dates the first announcement to 3 August, while the distribution documentation dates the weights to 14 August: the announcement date is not confirmed by a primary source. We found no dedicated post on the official qwenlm.github.io blog, and the VentureBeat article on Qwen3.8-Max would not open (HTTP 429), so the promise of open weights rests on a search snippet rather than a direct read. The GitHub repository 'AlibabaCloud-Official/Qwen3.8-27B' does not belong to QwenLM and was not used as a source.
- Perché pubblicarla
- This is a release you can verify and download today, not an announcement: Apache 2.0 weights, three independent distribution channels, hardware within reach of one consumer GPU. The angle is measurable: by the vendor's own numbers, an open 27-billion-parameter model beats the proprietary Qwen3.7-Plus, which is only reachable through a paid API, on agentic coding and office tasks. For anyone who wants to run models locally, for cost, latency or to keep data on their own servers (no small matter under GDPR), that has immediate practical consequences rather than being an abstract leaderboard.