Qwen3.8-27B Open-Weights: Alibaba's Free Apache-2.0 Coding Model You Can Run Locally — Full Guide 2026

On the evening of August 14, 2026, Alibaba's Qwen team released Qwen3.8-27B with fully open weights under an Apache 2.0 license, including an FP8-quantized variant that keeps performance nearly identical while shrinking file sizes. This is not a routine drop: it is the "worth running locally" model of the current generation, and it completes the Qwen 3.8 story — days after the giant Qwen 3.8-Max launch, where the open weights were listed as "coming soon." They have now arrived.

Source: @alibaba_qwen official X account

What exactly was released?

The model is available on Hugging Face as Qwen/Qwen3.8-27B and the compressed Qwen/Qwen3.8-27B-FP8 — fine-grained FP8 with block size 128, which the model card says performs "nearly identically" to the original BF16. Apache 2.0 means full commercial use with essentially no strings attached, a decisive edge over "open-ish" licenses that impose revenue caps or publication clauses.

The model is genuinely multimodal: it accepts text, images, and long (hour-scale) video as input. Its architecture is hybrid — Gated DeltaNet layers interleaved with Gated Attention in a 16-block layout across 64 layers, with Multi-Token Prediction trained over multiple steps. That is the recipe behind the latest generation of strong long-horizon reasoners.

Source: Unsplash

A million-token context — locally

The native context window is 262,144 tokens, extensible to 1,000,000 via YaRN. In practice: an entire code repository, hundreds of pages of documentation, or an hour of video with its transcript can live inside a single conversation — without shipping your data to anyone's cloud.

On the reasoning-control side, Qwen added what used to be closed-model territory: enable_thinking toggles deep thinking per request, reasoning_effort offers three depths (xhigh default, medium, low), and preserve_thinking keeps the reasoning chain across turns, improving KV-cache reuse — meaning faster responses in long conversations.

The numbers: where does 27B stand?

From the official model card:

On vision-language tasks: 84.3 on OSWorld-Verified, 81.9 on AndroidWorld, and 64.8 on WebArena-Verified — a documented ability to operate a computer, a phone, and a browser visually. These are numbers that compete with far larger closed models Alibaba itself lists as comparisons in the card (Qwen3.7-Plus, Muse Glimmer-30B, Opus 4.6 Max).

To be fair: this is a 27-billion-parameter model being compared against heavier, pricier rivals. Nobody claims it replaces frontier models like the 2.4-trillion-parameter Qwen 3.8-Max — but for everyday coding and agent work, it gets strikingly close at zero license cost.

How to run it from the Gulf and the MENA region

• Locally via vLLM or SGLang: both publish official recipes for the model (vLLM Recipes and the SGLang Cookbook). The FP8 build targets modern RTX 40/50-series GPUs with sufficient VRAM; Docker Model Runner and Transformers are supported too.
• Community GGUF weights: Unsloth published GGUF conversions the same day, opening the door to llama.cpp/Ollama on weaker hardware.
• Qwen Cloud, soon: the official page says hosted serving on qwencloud.com is "coming soon" with 1M context by default and official built-in tools — practical for teams without strong hardware, though no pricing is listed yet.
• Agent-harness support: the card explicitly mentions broader harness compatibility, including the Claude Code harness — so developers in Riyadh, Dubai, or Cairo can point their existing agent tooling at a local model.

What does this mean for you?

1. Full privacy for sensitive code: regional banks, fintech startups, and government digital entities whose compliance rules forbid sending source code to foreign APIs now have a first-tier coding model that runs inside their own network, under a license that allows commercial use.
2. Zero license cost: instead of per-seat subscriptions, you pay once for hardware. For startups across the region, that materially changes the cost structure.
3. A bridge from Max to local: want the absolute best? Use Qwen 3.8-Max via API. Want sovereignty over your data? Run 27B locally — same family, same well-known Qwen handling of Arabic.
4. Research and teaching without negotiations: universities and research centers can build derivatives and products on top of the model without licensing deals.

Quick comparison

Honest limitations

• The figures above are Alibaba's own evaluations; independent reviews have not yet confirmed every item (the HN thread passed 1,190 points and 700+ comments, and the debate is still live).
• Full BF16 inference needs server-grade hardware; FP8 eases the requirement but is still not "any laptop," and smaller GGUF quantizations trade away accuracy.
• The official Qwen Cloud hosting was "coming soon" at release, with no pricing announced.
• The computer-use scores (OSWorld et al.) are benchmark environments — production desktop-agent deployments remain early days.

Sampling settings every developer needs

The official card recommends precise sampling settings worth documenting because they change behavior dramatically: for deep-thinking mode use temperature=1.0 with top_p=0.95 and top_k=20; for direct-answer mode (thinking off) use temperature=0.7 with top_p=0.80, top_k=20, and presence_penalty=1.5. The practical difference: thinking mode is higher-quality on complex tasks but slower and more token-hungry; direct mode is faster for routine work. For anyone building production agents on top of the model, preserve_thinking matters more than anything else — keeping the reasoning chain across turns reduces recomputation and exploits KV-cache reuse far better in long conversations.

The early-adoption signal also deserves a read: the FP8 card alone logged over 123,000 downloads in its first month, and the open-source ecosystem (Unsloth, official vLLM and SGLang recipes) was ready on launch day. That kind of surrounding environment is the difference between a release that gets forgotten and one that gets built upon.

FAQ

Is Qwen3.8-27B really free?
Yes. Apache 2.0 allows download, commercial use, and modification for free; your only cost is hardware or any cloud hosting you choose.

How is it different from Qwen 3.8-Max?
Max is the closed flagship at 2.4 trillion parameters, API-only. The 27B is a smaller open-weights model you run locally, sufficient for most coding and agent tasks.

What hardware do I need to run it locally?
The FP8 build targets recent high-VRAM GPUs (upper RTX 40/50 series). Community GGUF builds run on weaker machines at lower precision. No official minimum is published.

Does it support Arabic?
The Qwen family has a strong multilingual reputation including Arabic, but the 27B card does not publish dedicated Arabic language evaluations — test it on your real workloads.

Can I use it with agent tools like Claude Code?
Yes. The official card lists broader harness compatibility including the Claude Code harness, alongside vLLM, SGLang, and Transformers.

Bottom line

With Qwen3.8-27B, a first-tier coding model under a truly open license became reality in August 2026 — not a promise. The frontier fight rages among closed labs, but the battle that matters to most developers is decided in the open layer: who delivers the best performance per gigabyte of your own memory. This release puts Alibaba at the front of that layer.

Sources

• Official Qwen3.8-27B-FP8 model card on Hugging Face
• Official announcement post on X
• Our earlier coverage: Qwen 3.8-Max by Alibaba
• AI tools on Truescho

Read next: GLM-5.3 by Z.ai: frontier coding and cyber capabilities and Meta Muse Glimmer: a 30B open model that runs on your machine