Capability-targeted compression of Qwen3.8-Flash-Next: half the experts are removed and remaining weights quantized to 3.5 bpw, producing a 58.4 GB GGUF (29.6 GB resident) that preserves coding and multimodal ability while trading off other domains.
Performs a byte-level transplant of 144 tensors in an already-quantized GSQ-RCO Qwen3.8-Flash-Next to ablate the model's refusal direction while preserving GSQ-learned scales and the upstream per-tensor type assignment; multimodal, 262K context. Intended for local inference, red-teaming and quantization research; no retraining or built-in safety.
Provides non-uniform GGUF quantizations of DeepSeek-V4.1-Flash-1017, using GSQ and RCO to allocate precision per tensor under a size budget. The IQ2_XXS build matches the base model on the three reported benchmarks.
Reduces refusals by removing a refusal direction from selected dense residual writers in a quantized DeepSeek-V4.1-Flash-1017. It preserves the upstream GSQ-RCO routed experts and offers three low-bit size options.
Compresses the model by pruning half its experts and quantizing the retained weights, targeting code, agentic workflows, long context, and vision. It retains 94.8%–98.7% of the base model’s scores on two coding benchmarks, at the cost of broader capabilities.