Alibaba's Qwen research lab released Qwen 3.8 27B on Friday: an Apache 2 licensed, 27‑billion‑parameter multimodal LLM with vision capability. The model ships as roughly a 17 GB file and, according to Qwen's own benchmarks, improves over Qwen 3.6 27B and the closed‑weight Qwen 3.7‑Plus.
Test environments and builds used
The reviewer ran the model on two machines: a 128 GB M5 Max MacBook Pro and an NVIDIA DGX Spark. On both systems they used LM Studio and the 17 GB Q4_K_M quantized GGUF build; the Spark was also used with llama-server directly.
The default "xhigh" reasoning setting causes extreme over‑thinking
Qwen 3.8 documents a reasoning_effort setting with defaults of xhigh (for complex, thorough analysis), medium (balance of accuracy and speed) and low (efficient, speed/cost optimized). The reviewer found the xhigh default to be a poor choice for consumer hardware: the model spends an excessive number of reasoning tokens and produces long internal thought traces, making runs very slow.
Initially the reviewer hit LM Studio's default 8,192 token context limit and observed the model consuming the entire context with internal reasoning even for simple tasks. Increasing the context to the maximum 262,144 tokens removed that particular problem.
Example: pelican riding a bicycle (SVG)
With the larger context enabled, a first attempt to generate an SVG of a pelican on a bicycle took 21 minutes; the run used 22,276 reasoning tokens to produce 3,223 output tokens. The reviewer called it the best pelican SVG they had generated locally: correct bicycle frame, legs on either side, clear pouch, wings touching the handlebars, motion lines behind, and a tasteful background. Still, the 21‑minute generation time was not worth it for the author.
Running the same prompt with reasoning disabled produced 3,715 tokens and took 137 seconds (about two minutes), a much faster result.
The reviewer also ran the prompt through the much larger Qwen 3.8 2.4T‑A95B (released last week) via OpenRouter and obtained an animated SVG more quickly.
A simple circle prompt and excessive creativity
For a minimal prompt ("draw an svg of a circle"), the Qwen xhigh reasoning trace expanded the task into an elaborate, multi‑element, animated "geometric study" with gradients, rings, animation, and a chosen palette — producing a beautiful but unsolicited result after several minutes. The recommendation: ignore the xhigh default and start with low or no reasoning.
Vision capability: bounding boxes and labeling tools
As a vision test, the reviewer asked the model to return 0–1000 scaled JSON bounding boxes for pelicans in a photo. The model returned two bbox entries:
- [195, 290, 370, 780]
- [445, 320, 675, 850]
These aligned well with the objects in the photo, and the reviewer rendered the boxes on the image.
From a single prompt the model also generated a complete HTML tool that loads an image URL, reads a JSON textarea with 0–1000 scaled bboxes, measures the image dimensions, scales coordinates to the actual size, and renders labeled boxes. Because the reviewer had not lowered reasoning, the generated interface was over‑engineered and even included a sample/demo scene (a generated placeholder image with pelican‑like silhouettes) that the model invented so the demo would be self‑contained.
When reasoning was turned off, the model produced a nearly working tool, but the boxes were placed incorrectly — illustrating that reasoning can help reach a correct single‑round solution.
Driving coding agents
A key question for local models is whether they can power coding agent loops that require long context, reliable code generation, and tool calls. The reviewer tested Pi as an agent framework (chosen for its shorter system prompt) configured to use Qwen 3.8 27B running in LM Studio on the Spark via tailscale serve and with reasoning=true in the model definition.
Using Pi, a prompt such as "how does auth work?" produced a solid reply after multiple reasoning steps and tool calls. The reviewer then asked the agent to write Python to convert a Pi JSONL transcript to Markdown; the model produced and tested a pi_jsonl_to_md.py script that performed the task.
The speed problem and community optimizations
Qwen 3.8 27B is promising in functionality, but performance is the main limitation. The reviewer measured about 15–30 tokens/second from LM Studio — usable but significantly slower than hosted APIs. The article references "Artificial Analysis" token speed tracking that shows OpenAI 5.6 Sol at 74 tokens/s and 5.6 Luna at 184 tokens/s for comparison.
The community has already been exploring speedups. One important optimization supported by the model is Multi‑Token Prediction (MTP), where a cheaper mechanism predicts several tokens ahead and the main model quickly verifies them, improving inference throughput. Following a tip from Georgi Gerganov (creator of llama.cpp), the reviewer ran llama serve with --spec-type draft-mtp on the Spark and observed a substantial speedup: a comparative benchmark showed the MTP server outperformed the LM Studio default GGUF by about 72%.
More community serving and inference innovations are likely over the coming weeks.
Observations and takeaways
That a 17 GB file can provide multimodal vision, long context, tool calling and capable code generation on consumer and prosumer machines is striking. Qwen 3.8 27B demonstrates that an open‑weights general purpose model with these capabilities can be compact and locally runnable.
The principal downside remains latency: dense (non‑MoE) models require considerable memory bandwidth, and on the hardware tested (M5 Max Mac and DGX Spark) the model felt slow. Nevertheless, the combination of capability and compact size suggests locally hosted models can increasingly compete with hosted offerings if inference speed and default reasoning settings are tuned.
Final verdict
Qwen 3.8 27B is a very capable, smallish multimodal LLM that runs locally and advances what small models can do. Users should beware the xhigh default reasoning setting — start with lower reasoning or apply community optimizations like MTP to get practical, responsive performance.
