Tensorbend turns model intervention, quantization, compression, and inference research into an inspectable paper, method records, code, weights, model cards, measured reports, and deployment recipes.
Intervention class, quantization method, serving technique, model family, and access terms remain separate. PRISM-PRO is a model class; PRISM Pro membership is a commercial access tier.
Projected Refusal Isolation via Subspace Modification (PRISM) identifies and modifies model-behavior directions for scoped over-refusal, bias, and propaganda removal. PRISM-LITE and PRISM-PRO name release classes; quantization, format, license, and price remain separate.
Published as model weights, cards, and model-specific evaluations.PRISM-DQ is a deployment method, not a behavioral tier. Precision and compression choices are made against model sensitivity and documented for each target artifact.
Published with package size, runtime target, and retained model features.EAGLE-3 drafting, native multi-token prediction, model merging, and expert pruning connect model research to measurable serving behavior.
External methods such as REAP are credited to their original authors.Results below are attributed to their paper or published model card. Links open the underlying artifact, not a summary detached from the work.
A system for modifying a running language model through background LoRA training. The paper reports 61/105 pooled recall, 60/60 general-knowledge preservation, and 69.6 seconds for 180 steps on an M4 Max.
A long-horizon EAGLE-3 drafter with full and compressed variants. The published card reports 1.97× single-stream decode on the PRISM-PRO target in SGLang.
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.
A 50% expert-pruned Kimi K2.5 build with reusable saliency data, INT4 packaging, and explicit provenance for the external REAP method.
Tensorbend’s publication standard keeps authorship, implementation, packaging, access, and evidence legible—even when a release combines first-party work with an external research method.
Identify base weights, model family, upstream license, architecture, and the exact derived artifact.
Distinguish Tensorbend interventions from quantization, formats, runtimes, and systems techniques; clearly credit externally authored methods.
Attribute reported measurements to the paper or model card that produced them; do not turn a model-specific result into a universal claim.
Where terms permit, release the paper, code, weights, model card, evaluation, or deployment recipe needed to inspect the work.
The Kimi K2.5 release is a Tensorbend implementation of PRISM with REAP expert pruning. REAP is credited to Cerebras Research and linked to the original paper; the implementation report and upstream method are not presented as the same authorship.
Name the model, method, evaluation, target hardware, and the artifact you want to publish or operate.
eric@tensorbend.ai