Prism capability extraction: specialists carved from Qwen 27B at <20% memory match full model
tokenbender · x · 2026-11-11
tokenbender shared progress on Prism: Unlocking Language Model Capability Extraction — isolating single capabilities from a trained LLM for cheap, targeted deployment:
- Early reproducible preprint released; the core question: can Pareto-driven capabilities be isolated and deployed with huge cost savings while preserving performance?
- Latest Oct 13 results show successful capability isolation with high performance at low memory cost: agentic specialists extracted from Qwen 3.8 27B run with <20% of full-precision weight memory.
- Benchmark table: function calling specialist 66.5% vs 66.2% big-model (100.4%), Terminal-Bench 2.1 77.8% vs 74.7% (104.1%), AIME 2025 96.67% vs 99.17% (97.5%), RedactionBench 102.5%.
- All results come from mask extraction via inference-level attribution, requiring no post-training; code to be released soon.
More from Infra
- OpenRouter officially acquired by Stripe, serving 10M+ developers — deedydas · 2026-11-11
- Anthropic's amended S-1 reveals $34.9B Lambda GPU deal is now terminable after missed facility deadline — rohanpaul_ai · 2026-11-11
- Nvidia has wiped out $1.82 trillion in market cap since its October 6 all-time high — Polymarket · 2026-11-11
- Fully Local Voice-to-Voice AI on a Samsung Phone with Liquid AI's LFM2.5-Audio — alexcovo_eth · 2026-11-11
- SpaceX locks 100+ GW of solar cell equipment from SINGULUS for Starmind AI satellites — ns123abc · 2026-11-11
- CoreWeave upgrades to CNCF Platinum member, runs its GPU cloud entirely on Kubernetes — altryne · 2026-11-11