The main efficiency gain comes from removing model capacity rather than changing what each token computes: the checkpoint retains top-10 routing and shared experts, but trims each layer's routed-expert pool from 512 to 384. That reduces the stated weight footprint from 95.12 GiB to 79.78 GiB, a useful trade-off for serving a large multimodal model on constrained hardware.
Key Capabilities
- Neural Architecture Search selected the smaller expert configuration; attention, the vision tower, tokenizer, and n-gram embedding table remain unchanged. This makes the size reduction easier to attribute to expert pruning.
- The checkpoint is a complete merged model, not a LoRA adapter that must be loaded separately. It retains the base FP4-oriented weight layout, with selected Gated DeltaNet projections stored as BF16 after training updates were folded in.
- Three GRPO-style LoRA updates target short-budget CVSS completion and structured security findings. The reported benchmark results are mixed: some scores exceed the base reference, while GPQA-Diamond and MMMU are lower; the base figures were not rerun in the same evaluation.
- Serving was validated with vLLM on one B200, but the repository includes a custom patch for its weight layout, so compatibility should not be assumed across serving stacks.
Who It Suits—and the Trade-off
A good fit for practitioners testing lower-memory MoE inference or studying expert-count reduction and verifiable-reward tuning. Look elsewhere if you need an officially supported Qwen release or dependable CVSS scoring: this is an unofficial derivative, and its own evaluation reports only about 4% exact CVSS agreement despite better task completion and structured-output rewards.
