🏗️ Building on HF
witcheer
witcheer
AI & ML interests
Local AI Maxxing.
Recent Activity
updated a dataset 2 days ago
witcheer/microduck-skill-tree updated a model 2 days ago
witcheer/microduck-walk-recover updated a model 2 days ago
witcheer/microduck-walk-flatOrganizations
Gemma 4-E4B agentic results: fastest portscout (30s), PARTIAL logpulse
#11 opened 4 months ago
by witcheer
Gemma 4-E4B: fastest portscout ever (30s), logpulse PARTIAL
#5 opened 4 months ago
by witcheer
Gemma 4-E4B: 64 tok/s, 6/6 quality (best on leaderboard)
#10 opened 4 months ago
by witcheer
OmniCoder-9B: fastest code gen, same agent loop problems (2 new rows)
#4 opened 4 months ago
by witcheer
GBNF structured CoT: fixing Qwen3.6 rumination with a 4-rule grammar (12 new rows)
#3 opened 4 months ago
by witcheer
deep-dive benchmark: 8 prompts used to test gpt-oss-20b
#1 opened 4 months ago
by witcheer
Qwen3.6 27B dense: MoE vs dense control experiment (10.8x speed gap)
#9 opened 4 months ago
by witcheer
Llama 3.2 1B — 228 tok/s speed ceiling, 771 MB
#8 opened 4 months ago
by witcheer
Qwen3 8B — dense 8.2B, first to pass hallucination test
#7 opened 4 months ago
by witcheer
GLM 4.7 Flash — 30B MoE + MLA, ncmoe sweep, quality tests
#6 opened 5 months ago
by witcheer
added Mistral 7B v0.3: 56.4 tok/s, best quality (5/6), VRAM-hungry
#5 opened 5 months ago
by witcheer
added Gemma 4 E2B: 117.8 tok/s, dense, full GPU, quality traces
#4 opened 5 months ago
by witcheer
LFM2 24B A2B — ncmoe sweep + deep stress test results (RTX 4060 Ti 8GB)
#3 opened 5 months ago
by witcheer
Gemma 4 26B A4B — ncmoe sweep results added (RTX 4060 Ti 8GB)
#2 opened 5 months ago
by witcheer
new dataset: turboquant KV cache benchmarks
#1 opened 5 months ago
by witcheer