-
nemotron-twotower autopsy: 2.4x that costs a second 30B model (report + chart + Field-Reports row)
-
ace-step 1.5 music-gen: first RTX-5090 numbers (report + chart + Field-Reports row)
-
Upload reports/steering-claimed-vs-measured.png with huggingface_hub
-
Upload reports/bias-only-steering.md with huggingface_hub
-
Upload reports/deepseek-ocr-2.md with huggingface_hub
-
Upload reports/deepseek-ocr-2.png with huggingface_hub
-
Upload reports/deepseek-ocr-2.md with huggingface_hub
-
Upload reports/qwopus-coder-35b-a3b.png with huggingface_hub
-
Upload reports/qwopus-coder-35b-a3b.png with huggingface_hub
-
Upload reports/qwopus-coder-35b-a3b.md with huggingface_hub
-
Upload reports/agentworld-think-on-null.png with huggingface_hub
-
Upload reports/agentworld-think-on-null.md with huggingface_hub
-
Upload reports/ornith-1-0-35b-anchor.md with huggingface_hub
-
Upload reports/ornith-legb-inversion.png with huggingface_hub
-
Upload reports/ornith-1-0-35b-harness-inversion.md with huggingface_hub
-
Upload reports/ornith-anchor.png with huggingface_hub
-
Upload reports/ornith-1-0-35b-anchor.md with huggingface_hub
-
Upload reports/ornith-anchor.png with huggingface_hub
-
Upload reports/ornith-1-0-35b-anchor.md with huggingface_hub
-
Upload reports/omp-harness-as-variable.png with huggingface_hub
-
Upload reports/omp-harness-as-variable.md with huggingface_hub
-
agentworld: claim-vs-measured + solve-set reshuffle chart
-
agentworld: LWM zero-FT transfer null (flat synthetic, -2 real SWE)
-
fp4: 2-panel chart (vLLM bars + QuTLASS GEMM-vs-decode divergence)
-
fp4: complete with QuTLASS MXFP4 Tier-2 (real at the GEMM, lost at the token)
-
Upload reports/fp4-consumer-blackwell.png with huggingface_hub
-
Upload reports/fp4-consumer-blackwell.md with huggingface_hub
-
Upload reports/one-shot-em.png with huggingface_hub
-
Upload reports/one-shot-em.md with huggingface_hub
-
Upload reports/cacheback-spec-decode.png with huggingface_hub
-
Upload reports/cacheback-spec-decode.md with huggingface_hub
-
Upload reports/qwopus-coder-compat-recovery.png with huggingface_hub
-
Upload reports/qwopus-coder-compat.md with huggingface_hub
-
Upload reports/tts-head-to-head.png with huggingface_hub
-
Upload reports/tts-head-to-head.md with huggingface_hub
-
Upload reports/swebench-anchor.md with huggingface_hub
-
Upload reports/qwable-5-anchor.png with huggingface_hub
-
Upload reports/qwable-5-27b-coder.md with huggingface_hub
-
Upload reports/tts-head-to-head.png with huggingface_hub
-
Upload reports/tts-head-to-head.md with huggingface_hub
-
GLM-5.2 memory-vs-context walls chart
-
GLM-5.2 autopsy: 743B MLA+DSA, datacenter-only by arithmetic
-
pi-tune provenance scatter
-
Qwen3.6-27B pi-tune: real-trace SFT improves real SWE, the constructive contrast to the distill regressors
-
Qwable-3.6-27b flat-cliff chart
-
Qwable-3.6-27b: distill regresses real SWE -7 at matched Q4, synthetic missed it
-
Upload reports/qwable-decline.png with huggingface_hub
-
Upload reports/qwable-v1.md with huggingface_hub
-
Upload reports/echo-maze-microcosm.png with huggingface_hub
-
Upload reports/echo-maze-microcosm.md with huggingface_hub