Writing 100 facts into Qwen's n-gram Engram table with zero weight changes: 84% recall
Electronic_Put4530 · reddit · 2026-09-21
A Reddit user demonstrated ENGRAFT: injecting knowledge into Qwen3.8-Flash-Next's 320M-row n-gram lookup table (DeepSeek's Engram design) by training only table rows—no weights touched, yielding a 9 MB removable overlay. On an 841-sentence held-out test of 100 fictional facts, exact-answer accuracy hit 0.841 vs 0.005 base, 0.873 on a second seed, with only a 9-point spread across question shapes. Cross-language grafts scored 0.841 (Italian), 0.797 (English), 0.676 (Chinese). Training 14,032 rows took 2.4 hours on an AMD Ryzen AI MAX+ 395 mini PC. The whole project was built by one human driving Claude Code sessions, with design, implementation, adversarial review and verification split across separate model instances.
More from Infra
- Mozilla AI runs a local 30B model end-to-end to open a real bugfix PR, fully offline — mozilla-ai · 2026-09-21
- Cohere Labs launches Local AI community program for local inference and hardware tuning — Cohere_Labs · 2026-09-21
- Gewell: custom Gemma 4 inference engine cuts KV cache VRAM to 0.625x, losslessly — stoppableDissolution · 2026-09-21
- DeepSeek-V4.1-Flash redesigns the Transformer for agents, cutting KV cache to 890 bytes/token — AndLukyane · 2026-09-21
- Jev Engineering gives agents a decision brain, 193x faster and 444x cheaper in tests — agihouse_org · 2026-09-21
- UK's £225m Isambard-AI supercomputer cost about the same as one road bridge — charlieharris01 · 2026-09-21