Weight-Space Programming replicates: hand-edited 18GB 'Nanseek' nears frontier scores at 1/25th the size
teortaxesTex · x · 2026-11-09
rikiyatakehi's 'Weight-Space Programming' — GPT-6 Astra directly editing 0.016% of Qwen3.5-0.8B's weights, no gradients needed, boosting math 6x and instruction following 3x — has been replicated: Deepseek-V4.1-1017 hand-edited Nanbeige4.3 weights into 'Nanseek' over 17 hours and $7 of compute. Benchmarks show Nanseek v0.2 (18GB, Q4) scoring 0.137 on GF-CR versus an estimated 453GB for frontier models at that level — 25x smaller — trailing only Opus-5.5 (0.337) and Deepseek-V4.1-Flash-1017 (0.148). Frontier LLMs can now natively read and write neural network weights.
More from Models
- User says Terra Max's batch browser use is the main reason for their ChatGPT subscription — amplifiedamp · 2026-11-09
- DeepSeek V4.1 Flash 1017 hits 1,000 tok/s on Vercel while a home 4x 3090 rig does 65 — AccBalanced · 2026-11-09
- Fastino open-sources GLiDE decision models: 27B scores 72.49 on Decision Index at 112ms — RobertTLange · 2026-11-09
- Grok usage limits reset within hours after Musk-led engineering review — XFreeze · 2026-11-09
- Qwen4-27B leaks questioned: 480 tps benchmark may be Qwen4-Max-preview on low batches — teortaxesTex · 2026-11-09
- The 'Jensen loves Chinese models' joke: not the LLMs everyone assumed — haider1 · 2026-11-09