RT @JoelDeTeves: Testing @DJLougen "Ornstein" 27B - Q4_K_M Harmonic (same author) already felt like the smartest Qwen3.5-27B variant I'd tested. However *THIS* model @ Q4 feels much more intelligent than it has any business being. Here is my layman's understanding of the difference (I might be completely wrong - I am not a neuroscientist): - Both draw from the same high-quality "premium" reasoning traces (exactly 799 premium examples in both cases). These traces are deep (~1,667 words on average), statistically validated, and engineered to include self-correction (100% of rows), verification, and exploration of alternatives. - However, the key difference is that Harmonic-27B uses *only* the 799 premium traces, whereas Ornstein-27B builds on those same 799 premium traces and *adds 430 curated degenerate traces* (total 1,229). In other words it deliberately includes examples of bad reasoning - loops, restating without progress, filler, superficial padding, so the model learns what effective thinking is *not*. Absolute mad science happening here. Follow @DJLougen for more! Speed: 31 tokens/second (good) Mmproj -> Unsloth/Qwen3.5-27B (image recognition tested and works great) VRAM usage: 21.6 GB Configuration: -m Ornstein-27B-Q4_K_M.gguf --mmproj mmproj-F16.gguf --n-gpu-layers 99 --ctx-size 262144 --cache-type-k turbo4 --cache-type-v turbo4 --fit on --jinja --reasoning-format auto --flash-attn on Using @spiritbuun TurboQuant fork of Llama.cpp - running @ max context to test the limits of this version and see where context rot starts to happen. Also worth follow! Next test: will it perform in…

Mehr auf Arint.info

#for #gguf #Llama #Llamacpp #max #science #the #Unsloth #arint_info

https://x.com/JoelDeTeves/status/2041761720520339545#m