A model can reach “1 bit per parameter” without actually storing every retained weight at one bit: this release gets there by combining quantization with removing half the routed experts. That distinction matters when comparing model-size claims—the reported 1.05 bits per original transformer parameter includes experts that were deleted, while the parameters that remain average 2.06 bits.
Key Capabilities
- Compression is split across two techniques. RCO selects 192 of 384 experts per layer for removal; the surviving weights retain the per-tensor quantization scheme of the IQ2_XXS build. This reduces transformer storage to 72.1 GB, rather than achieving the reduction through extreme quantization alone.
- The pruning is targeted, not general-purpose. Expert selection is optimized against calibration data to preserve code, tool use, long-context, and vision capabilities. The calibration distribution therefore helps determine what the compressed model can still do well.
- Coding scores stay relatively close to the base. At reasoning effort 50, the model scores 77.00 on SWE-bench Verified and 88.57 on LiveCodeBench v6, retaining 94.8% and 98.7% of the corresponding base-model scores. The lower retention on multi-step repository work is a reminder that errors can accumulate across agentic turns.
- The headline size does not include the Engram tables. The complete set of listed model files totals 182.72 GB, including 110.60 GB of lookup tables. Those tables are shared with the unpruned release and are neither pruned nor included in the transformer’s 1.05-bit calculation.
Who It Suits—and the Trade-off
A good fit for local-inference users prioritizing coding and agentic tasks who can accommodate a large overall model footprint and want to experiment with expert pruning. Look elsewhere if you need balanced general-purpose performance: removing half the experts inevitably reduces capacity, and capabilities outside the calibration targets may degrade. The release is experimental, so its results should be treated as task-specific rather than evidence of a universally smaller equivalent.
