PRISM
Projected Refusal Isolation via Subspace Modification identifies and modifies directions associated with over-refusal, bias, and propaganda.
Weights · model cards · model-specific evaluationsResearch.
Model intervention, quantization, compression, and inference.
Current methods.
PRISM-DQ
Precision assignment by tensor class and measured model sensitivity for a specified runtime and hardware target.
Package size · runtime · hardware · retained featuresInference and compression
EAGLE-3 speculative decoding, native multi-token prediction, model merging, and expert pruning.
Latency · throughput · acceptance rate · artifact sizePublications and reports.
JIT LoRA: Real-Time Conversational Knowledge Injection on Apple Silicon via MLX
Background LoRA training updates a running language model. On M4 Max, the paper reports 61/105 pooled recall, 60/60 general-knowledge preservation, and 69.6 seconds for 180 steps.
Qwen3.6 27B EAGLE-3 speculative decoding
Full and compressed EAGLE-3 drafter variants. The model card reports 1.97× single-stream decode on the PRISM-PRO target in SGLang.
MiniMax SLURPY: per-tensor empirical SLERP
A 48,239-tensor merge of MiniMax M2.5 and M2.7 with per-tensor interpolation derived from measured parent deltas and native FP8 re-quantization.
Kimi K2.5 PRISM + REAP expert pruning
A 50% expert-pruned Kimi K2.5 build with reusable saliency data, INT4 packaging, and explicit provenance for the external REAP method.
Research collaboration.
Send the model, method, evaluation, hardware target, and intended artifact.
contact@tensorbend.ai