AI/ML●●●●Gem
KV-Cache Grafting – Boosting frozen 12B LLMs to 93.3% AIME accuracy
Frozen 12B model hits 93.3% AIME by grafting verified KV states, not retraining.
WizardryBig BrainZero to One
Corbenic
3011d ago
Back to browse
A 6M-token movable window on a single 46GB GPU
by Wetime·Jul 28, 2026·3 points·1 comment
AI Analysis
●●●●GemWizardryZero to OneBig Brain
6M-token context window on one GPU when vLLM caps at 30K tokens.
Strengths
- •Execution-bound capability fully decoupled from parameter scaling with 180/180 accuracy
- •1.4 microsecond memory selection with bit-exact deterministic outputs forever
- •Public testbench with SHA-256 provenance manifest for full reproducibility
Weaknesses
- •Only works on verified problem families, not general reasoning tasks
- •Frontier models still dominate raw from-scratch reasoning benchmarks
Category
Target Audience
ML researchers and engineers working on model efficiency
Similar To
Retrieval-augmented generation systems · Neural caching approaches
Similar Projects
AI/ML●●●Banger
Virena, a minimal vision-language-action robot model you can read
Frozen CLIP plus tiny trained head runs VLA training on Mac — no GPU required.
Big BrainNiche GemWizardry
georgia_bucea
2018d ago
AI/ML●●Solid
Aurora Subtitles RealTime speech translation running locally on Windows
Local Whisper + NLLB translation with 300ms latency overlay for Discord and games.
Niche GemSolve My Problem
bartuff
204mo ago
Developer Tools●●Solid
In-browser real LLM token counter and cost estimation
Local tokenizer execution beats sending text to external counters.
CozySolve My Problem
angelmm
311mo ago
Developer Tools●●Solid
AIST – 950-token protocol for preserving AI session state
60x compression for AI context, but handoff format viability depends on LLM adoption.
Big BrainNiche Gem
robigewii
115mo ago
AI/ML●●Solid
I fine-tuned Qwen 3.5 (0.8B–4B) on a Mac for text-to-SQL – 2B beats 12B
Unified memory trick lets a 2B model beat 12B; trains on MacBook with zero cloud costs.
Ship ItNiche GemBig Brain
sciences44
714mo ago