1. Summary

Qwen3.8-27B is a dense vision-language model with 27B parameters based on the Qwen3.5 architecture. According to the Hugging Face Model Card, it achieves significant performance improvements over the previous generation in coding, research, and agent tasks.

It natively supports image and video understanding and features thought control.

Regarding quantized build performance, it has been reported that the 4-bit Q4_K_M quantization can maintain performance equivalent to the full model in specific benchmarks. On the other hand, performance degrades significantly at 1-bit.

Within the community, discussions have occurred regarding its practicality on devices such as the Mac Studio, as well as its output redundancy and inference speed. Mention has also been made of the impact that default thought control settings have on token consumption.

2. Bunrin Bench (BUNRIN LABO Original Test)

Results of the "Bunrin Bench" automated scoring benchmark (covering 7 categories including output contracts, fact retention, instruction following, and code traps) developed by we, conducted on local hardware. For the methodology, refer to About Bunrin Bench.

Target Test Version Total Contract Number Constraint Hearsay Code Non-Fabrication Title Speed Test Date
qwen3.8-27b-mtp (MacBook Pro (M5 Max, 128GB)) ijiwaru-v1 35/41 (85%) 2/3 9/9 3/6 3/3 11/11 5/5 2/4 49.1 tok/s 2026-09-03

Testing Environment:

  • qwen3.8-27b-mtp — Model: root4k/Huihui-Qwen3.8-27B-abliterated-oQ6e-mtp (MLX version). Runtime: oMLX 0.6.4 (mlx 0.32.2, mlx-lm 0.31.3). Sampling: temperature 1 / top_p 0.95 / top_k 20 / min_p 0.08 / presence_penalty 0

3. Various Benchmarks

The values listed on the Hugging Face Model Card are as follows:

Category Benchmark Qwen3.8-27B Qwen3.6-27B Qwen3.7-Plus Muse Glimmer-30B Opus4.6 Max
Coding Terminal Bench 2.1 73.0 63.4 64.0 51.7 78.2
Coding SWE-bench Pro 61.7 53.5 57.6 51.2 53.4
Coding NL2Repo-Bench 42.3 36.2 41.1 -- 47.6
Coding DeepSWE 1.1 42.2 13.3 14.2 -- --
Coding QwenSWEBench 79.0 49.3 59.2 -- 63.8

Additionally, the evaluation results from Artificial Analysis are as follows:

Metric Value
Intelligence Index 52

4. Official announcements

  • According to the Hugging Face Model Card, Qwen3.8 is stated to be the most capable generation within the Qwen family.
  • The same card indicates that a hosted version via Qwen Cloud is "coming soon."

5. Real-world performance (Community reception)

Artificial Analysis reported that the model scored 52 on the Intelligence Index. On the other hand, it was pointed out that the output is very redundant and the inference speed is slow at 47 tokens/s.

According to reports from Hacker News, the 17GB Q4_K_M quantized build maintains performance equivalent to the full model in Terminal-Bench 2.1. This allows the model to operate in environments with 24GB of VRAM, such as the RTX 4090, while retaining a context of approximately 64k tokens. However, performance drops significantly at 1-bit.

Simon Willison stated that after testing the Q4_K_M version in LM Studio and other tools, the default reasoning_effort (xhigh) setting consumes a vast number of tokens even for trivial problems.

According to Terminal Bytes, smooth operation is possible on Mac Studio through memory residency. The site provided examples of the model being utilized for everyday tasks, such as summarizing RSS feeds and organizing PDF files.

Users on Reddit mentioned an improvement in autonomous coding capabilities using screenshots. Additionally, a user on Reddit reported a case of creating a Minecraft clone using Q4 quantized build.

The llama.cpp release reported optimizations regarding KV cache restoration. This significantly reduces the time required to restore non-contiguous cells. Furthermore, the argument --tensor-read-lazy has been changed to --lazy-mode.

A user on Reddit pointed out that reducing thinking time in certain fine-tuned versions led to a decline in performance. Meanwhile, a user on Reddit mentioned the speed in an environment using two R9700s.

Additionally, users on Reddit are discussing the repetition_penalty=1.0 setting.

6. Recommended parameters

  • The settings listed on the Hugging Face Model Card are as follows:
    • thinking mode: On by default. Can be disabled per request.
    • reasoning_effort: Adjusts the depth of inference.
    • preserve_thinking: Retains inference context from history messages.

7. Sources