Alibaba released the open-weight Qwen3.8-27B model in August 2026 alongside its 2.4 trillion-parameter Qwen3.8-Max flagship. The release puts coding, reasoning and multimodal capabilities on consumer hardware while allowing local processing of sensitive code, financial data and internal documents.

The model surpassed 1 million downloads on Hugging Face within days, while downloads through LM Studio exceeded 1.6 million. Its weights are published under the Apache 2.0 license, which permits commercial use, on Hugging Face and ModelScope.

Architecture and benchmark results

Qwen3.8-27B is a dense model with 27.78 billion parameters, 64 layers and a hidden size of 5,120. Unlike mixture-of-experts systems, it uses all parameters to generate each token, simplifying deployment and fine-tuning but increasing memory-bandwidth requirements.

The model accepts text, images and video and produces text. Its native context window is 262,144 tokens and can be extended to 1 million, although Alibaba warns that scaling may reduce quality on short prompts. The company has not disclosed its training corpus, training recipe or knowledge cutoff, meaning the weights and inference code are open but the creation process is not.

Alibaba’s official results give Qwen3.8-27B a 61.7% score on SWE-bench Pro. Compared with Qwen3.6-27B, its DeepSWE 1.1 result rose from 13.3 to 42.2, its OSWorld-Verified score increased from 63.9 to 84.3, and its result on Alibaba’s internal QwenSWEBench advanced from 49.3 to 79.0.

Those figures were produced by Alibaba, include internal tests and do not always use results obtained under identical conditions. The independent BenchLM aggregator ranked the model 80th among 216 models with 54.3 points out of 100. It placed seventh among 26 models for reasoning, 20th among 135 for agentic tasks and 88th among 179 for general knowledge.

In the Vals AI index, Qwen3.8-27B scored 48.5 points, including 86.0% on SWE-bench, 84.0% on LiveCodeBench and 67.8% on CyberBench v1.1. Runpod separately cited an average accuracy of 67% across 250 tasks in the Harvey and Engram legal-research benchmark.

Hardware and reasoning settings

A 4-bit Q4_K_M build occupies about 17-18 GB. It fits on a 24 GB RTX 3090 or RTX 4090, or a 32 GB RTX 5090, with an expected generation rate of 15-30 tokens per second. Apple Silicon Macs require at least 32 GB of unified memory. Systems with 12-16 GB of VRAM and at least 32 GB of RAM can offload layers to system memory, preserving the 4-bit build’s quality but reducing speed.

The model enables its xhigh reasoning level by default. Users can select medium or low reasoning, or disable it with enable_thinking: false. Developer Simon Willison reported that an SVG task on an M5 Max MacBook Pro with 128 GB of memory took about 21 minutes at xhigh, using 22,276 reasoning tokens and 3,223 output tokens. With reasoning disabled, the task took 137 seconds and produced a simpler result.

Local test with Ollama

Incrypted tested the model through Ollama on an RTX 3060 with 12 GB of VRAM and 64 GB of DDR4-2400 RAM. With Ollama’s full 262,144-token context, the model’s memory footprint reached 28 GB, 73% of computation ran on the CPU, loading took 4 minutes and 14 seconds, and generation reached 1.22 tokens per second.

Creating a model variant with num_ctx 16384 reduced loading time to 13 seconds, raised prompt processing from 0.68 to 15.5 tokens per second and increased generation to 6.7 tokens per second, a 5.5-fold gain.

Blender test exposed coding limits

With reasoning disabled, the model generated a 9,021-token, 866-line Blender script for an animated nighttime lighthouse scene in 31 minutes and 38 seconds. The script initially failed because it attempted to modify a read-only Blender field. After 13 lines, or about 1.5% of the script, were disabled without adding code, it created 31 objects and a working 240-frame lighthouse-beam animation.

The output still contained major spatial errors: the rock was too small, the lighthouse floated above it, and the animated camera pointed away from the lighthouse. The test showed that the model could produce substantial structured code locally, but application-specific APIs and scene geometry required verification and further iterations.

Source: Incrypted