What is Alibaba’s 2026 Qwen3.8-27B: 27B open weights and 3M downloads
Qwen3.8-27B open weights landed 14 August; by 17 August the story was 3 million downloads and a third-party index. This piece separates the official card from independent scores.
On Friday 14 August 2026, Alibaba’s Qwen team put a dense checkpoint of about 27 billion parameters on Hugging Face. By Monday the developer conversation had shifted from “another open card” to whether a local 27B could take on some coding and agent work that lately lived behind cloud APIs.
VentureBeat followed on 17 August. The object is Qwen3.8-27B, repo Qwen/Qwen3.8-27B, license Apache 2.0. What follows is only what can be checked: the official card, the third-party indexes, and which figures remain vendor-reported.
What Qwen3.8-27B is
The official model card calls it a causal language model with a vision encoder: about 27B parameters, native multimodal (text, image, video), native context 262,144 tokens, written as extensible to 1 million. Thinking is on by default, tunable with reasoning_effort, and can be disabled per request.
Figures that can be checked
-
1
The vendor table leads with coding and computer use
The official card lists SWE-bench Pro 61.7, LiveCodeBench v6 90.3 and OSWorld-Verified 84.3, and compares Qwen3.6-27B and Claude Opus 4.6 Max on the same table. Footnotes say some tasks used an in-house harness; conditions are not identical across every row.
-
2
A third-party composite arrived on Monday
VentureBeat wrote that Artificial Analysis gave an Intelligence Index of 52, the same score the piece assigned OpenAI’s mid-tier GPT-5.6 Luna at maximum reasoning; Agentic Index 51, written as ahead of Claude Opus 4.8 at max reasoning. These are composites, not “already equal to a named cloud model.”
-
3
Download counts are platform counters
The same VentureBeat story cited Cybernews: more than 3 million Hugging Face downloads in three days, with quantizations appearing quickly. Alibaba Cloud’s 17 August community post separately said the model entered the Hub’s top-five most-liked within two days. A download counter is not unique users.
The license is looser than the flagship’s
The 27B and FP8 repos are marked Apache 2.0 on Hugging Face, allowing modification, redistribution and commercial use if license and notices stay attached. The 2.4T flagship in the same generation uses a separate Qwen3.8-Max license; the two do not transfer.
VRAM decides which file you actually run
VentureBeat wrote that full 16-bit needs about 56GB of GPU memory, FP8 about 28GB, and a 4-bit quantization about 17GB for the model itself. The official repo is BF16 safetensors; the GGUF files common on consumer cards come from community repos, not the main official tree.
Default thinking is expensive
Artificial Analysis said the model produced about 160 million output tokens on Intelligence Index testing, versus a 43 million median for comparable open-weight models. Simon Willison wrote that under the default xhigh setting, a pelican-on-a-bicycle SVG took 21 minutes and more than 22,000 reasoning tokens. He suggests starting local use on low or no reasoning.
| Claim | Checkable source | How to read it |
|---|---|---|
| SWE-bench Pro 61.7 | Official Hugging Face model card | Vendor eval; see footnotes |
| Intelligence Index 52 | Artificial Analysis; VentureBeat 17 August | Third-party composite, not single-bench equivalence |
| >3M downloads in three days | Cybernews; cited by VentureBeat | Platform count; may include quantized copies |
Alibaba Cloud’s 17 August community post said quantized builds can run on a laptop and that capability matches Qwen3.7-Plus, an MoE about ten times the size. That is company-blog wording and should be read apart from the third-party indexes.
# Official repo (reported)
Qwen/Qwen3.8-27B
Apache-2.0 · ~27B dense VLM · ctx 262144
hosted 1M context: coming soonHow to read the boundaries
- Open weights ≠ a public training recipe
- What you can download are post-trained weights and config. The official card does not also publish the training data and a full reproduction recipe.
- Hosted service is not written as live
- The card says Qwen Cloud will offer a default 1-million-token context and built-in tools. The status is coming soon, not a priced official SKU already on sale.
- Locally runnable is not “replaces a cloud flagship”
- Both the third-party indexes and the vendor table are still accumulating. Speed, token burn and harness differences matter more than a single “beats Opus” headline.
Questions worth checking
Are Qwen3.8-27B and Qwen3.8-Max the same model?
No. The 27B is a dense multimodal model under Apache 2.0. The same-generation flagship is a MoE of about 2.4 trillion total / 95 billion active parameters under a different license. The repo names differ too.
Has an independent lab reproduced the 61.7 score?
The SWE-bench Pro and related figures on the official card are Alibaba-published. The Artificial Analysis indexes in VentureBeat are a separate third-party composite. The two sets do not automatically validate each other.
Can a laptop run full precision today?
The consumer-machine reports are mostly about a roughly 17GB quantized file, not the official main-repo ~56GB full-precision weights. Framework, quant repo and the default thinking setting all change how fast it feels.