Avi Fenesh

@avi_fenesh

Systems SWE @ AWS ElastiCache | Valkey GLIDE maintainer | Low-level systems · GPU kernels | Applied ML research → avifenesh.ai | OSS all the way

avifenesh.ai

Joined February 2013

Following

Followers

Replies Media

Avi Fenesh
@avi_fenesh
18h
Qwen3.8-27B input just dropped 25% to $0.30/M on Tiyuvta Inference. Cached input $0.10 and output $2.03 unchanged. Same 180 tok/s, same full tool calling, same OpenAI-compatible API. Agents burn budgets on input tokens. Yours just got cheaper. inference.tiyuvta.ai/pricing
Avi Fenesh
@avi_fenesh
Aug 23
I could also write code by hand and do CPU optimization myself with the same method, but the point is to have a helpful and easy-to-use model. Opus is too verbose, blasts changes, rewrites what isn't asked for, and acts like your moral teacher. Everything is possible;
Boris Cherny
@bcherny
Aug 22
Replying to @scaling01
Agree. People are sleeping on using Opus to hill climb. We use it for optimizing CPU and memory, optimizing CI times, improving frame rates, reducing latency, any other kind of problem in the shape of “iterate on X with a profiler and dataset until it hits Y”
Avi Fenesh
@avi_fenesh
Aug 22
huggingface.co/Avifenesh/Orni… trained the mtp head to fit the model, instead of the base mode, big performance improvements.
Avifenesh/Ornith-1.5-35B-A3B-NVFP4-MTP-GGUF · Hugging Face
From huggingface.co
Avi Fenesh
@avi_fenesh
Aug 22
@ornith_ Ornith-1.5 35B-A3B $0.25/$0.09/$1.20 350 tok/s inference.tiyuvta.ai
Avi Fenesh
@avi_fenesh
Aug 21
Said I'm still tuning, and it's up #qwen 3.8 27b live with 260 tok/s with a much lower price than open router providers. Input 0.40, cached 0.10, output 2.03 (cache hit 77.9%), $5 credit on signup, no card needed. inference.tiyuvta.ai