BanWave
Back to blog

Qwen 3.8 27B: What It Is and How to Try It (2026)

Qwen 3.8 27B is Alibaba’s open-weight 27B vision-language model: Apache 2.0, 262K context, runs on one consumer GPU. Specs, access options, and what is still unverified.

Oct 22, 2026BanWave Team

Qwen 3.8 27B is a 27-billion-parameter open-weight vision-language model from Alibaba's Qwen team, published on Hugging Face on August 14, 2026 under the Apache 2.0 license — downloadable, self-hostable, and small enough to run on a single consumer GPU. (kie.ai deep dive)

What Qwen 3.8 27B actually is

Qwen 3.8 27B is the compact, self-hostable member of the Qwen 3.8 generation. It is a dense model with a native vision encoder — it accepts images and video alongside text without an adapter — rather than a sparse mixture-of-experts. (kie.ai)

The confirmed core specifications:

  • 27B parameters, dense, with a 5120 hidden dimension across 64 layers in a hybrid Gated DeltaNet plus Gated Attention layout. (kie.ai)
  • 262,144 tokens of native context, extendable to 1,000,000 tokens via YaRN. (kie.ai)
  • Reasoning and function calling built in, aimed at agentic workflows. (Cloudflare Workers AI changelog)
  • Apache 2.0 license — commercial self-hosting is permitted. (kie.ai)

Qwen 3.8 27B vs Qwen 3.8 Max: not the same model

The generation name covers two very different models, and coverage often conflates them:

| | Qwen 3.8 27B | Qwen 3.8 Max | | -------------- | --------------------------------------- | ------------------------------------------- | | Architecture | Dense, 27B | Sparse MoE, 2.4T total / ~95B active | | Access | Open weights, Apache 2.0 | API only, no public weights | | Context | 262K native, 1M via YaRN | 1M advertised | | Official price | No per-token hosted price published yet | $2 / 1M input, $6 / 1M output on Qwen Cloud |

(kie.ai)

If you want to run something on your own hardware, the 27B is the only half of this generation that makes that possible.

How to try Qwen 3.8 27B

Hosted, no setup: Cloudflare added the model to Workers AI as @cf/qwen/qwen3.8-27b, callable through the Workers AI binding or the REST API. (Cloudflare changelog)

Local: community quant builds (Unsloth Dynamic GGUFs) bring the memory requirement down to roughly 16–17 GB of RAM/VRAM, which fits a single consumer GPU such as a 3090 or 4090; the model is also available through Ollama and LM Studio. (kie.ai)

One practical warning from independent hands-on testing: the model's default reasoning effort (xhigh) over-thinks routine prompts — one tester watched it burn 22,276 reasoning tokens on a simple SVG request that took 137 seconds with reasoning off. Set reasoning_effort to medium or low for everyday use, and raise the context limit in your tooling, since an 8,192-token default can be exhausted by the model's own thinking trace before it answers. (kie.ai)

What is verified vs. what is vendor claim

The headline around this release — "27B model beats Claude Opus 4.6 on coding" — traces to Qwen's own benchmark card (SWE-bench Pro 61.7, LiveCodeBench v6 90.3). No independent benchmark run has confirmed those numbers; every headline figure currently originates from the lab that built the model. Treat the Opus-class coding claim as a vendor claim until third-party evals land. (kie.ai)

Also still open: no official per-token price for a managed 27B hosted service, and real-world throughput figures vary widely across rigs. (kie.ai)

Qwen 3.8 FAQ

Is Qwen 3.8 27B free to use? The weights are free under Apache 2.0, including for commercial self-hosting. Hosted access is priced by whichever provider serves it; no official Qwen Cloud per-token price for the 27B had been published as of mid-October 2026. (kie.ai)

Can Qwen 3.8 27B see images? Yes. It is a native vision-language model: it accepts image and text inputs and generates text, with function calling and a thinking mode for multi-step problems. (Cloudflare changelog)

What hardware do I need to run it locally? Community GGUF quants run it in about 16–17 GB of RAM/VRAM — a single RTX 3090/4090-class GPU is enough. (kie.ai)

Is Qwen 3.8 27B the same as Qwen 3.8 Max? No. The 27B is a dense open-weight model; the Max is a 2.4-trillion-parameter MoE available only as an API. (kie.ai)

Does Qwen 3.8 27B really beat Claude Opus 4.6 at coding? That claim comes from Qwen's own benchmark card and had not been independently confirmed at publication time. Benchmark it on your own tasks before relying on the number. (kie.ai)

Sources

Published 2026-10-22. Benchmark claims and hosted pricing for the Qwen 3.8 family are still settling; check the sources above for updates.