A team writing under the name mncai says it has built Hunmin-CUA, a computer-use model based on a 397B-parameter mixture-of-experts model, using a selective low-rank transfer from Qwen-CUA. In the team’s evaluations, the approach improved computer-use performance without carrying over the source model’s measured weaknesses in high-resolution screen grounding and Korean benchmarks.
Watch Desk analysis
What happened
The team’s community article on Hugging Face describes Hunmin-397B-A17B-CUA, with 17 billion active parameters. It says the project ran on one node with eight B200 GPUs. Rather than apply extensive supervised fine-tuning to its base model, the team transferred a compressed portion of the weight difference between that model and Qwen-CUA, then continued post-training.
The method kept different parts of the model at different ranks: 448 for attention and shared-expert paths, and 64 for routed experts. In the team’s reported evaluation, Hunmin scored 77.12 on OSWorld, compared with 48.16 for its base model and 70.45 for Qwen-CUA. On ScreenSpot-Pro, it scored 75.61, against 72.74 for the base and 62.20 for Qwen-CUA. Its KMMLU-Redux score was 81.43, close to the base model’s 82.59 and above Qwen-CUA’s 78.07. Read the team’s account.
Key findings
- A sizeable gain on OSWorld
Hunmin scored 77.12 against the base model’s 48.16; the team says OSWorld results came from single runs. - Grounding held up better than in the source model
Hunmin scored 75.61 on ScreenSpot-Pro, above Qwen-CUA’s 62.20 and the base model’s 72.74. - Korean benchmark performance stayed close to the base
Hunmin scored 81.43 on KMMLU-Redux, compared with 82.59 for the base and 78.07 for Qwen-CUA.
Why it matters
Computer-use tuning can improve an agent’s ability to operate a desktop while weakening other capabilities that matter in general use. The team’s results suggest that transferring selected parts of an already capable model may offer a way to improve tool use without simply inheriting every measured shortcoming of the source.
That is a useful technical result, not a guarantee that the recipe generalises. The team reports three runs for ScreenSpot-Pro and KMMLU-Redux, but only one for OSWorld; it also notes that its evaluations do not establish how the approach performs across other models, tasks or settings.
Our read
The interesting idea is not “bigger model wins”. It is that a carefully chosen slice of a model’s learned changes may be more useful than copying the whole thing. The results make a credible case for testing that idea further, while the single-run OSWorld score deserves a pencil mark in the margin rather than a victory parade.
What to watch
- Whether the team publishes the full training recipe and evaluation artefacts.
- Whether repeated OSWorld runs confirm the reported improvement.
- Whether the selective-transfer method works on other model families and computer-use tasks.
Discussion spark: If a computer-use model improves substantially while keeping other benchmark scores near its base model, should developers favour selective transfer over more fine-tuning?
Sources and evidence
- Hunmin-CUA: Post-Training a 397B MoE for Computer Use (8 October 2026, 00:49 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.