Andrey Kormilitsyn

[Andrey Kormilitsyn @AndreyKoms]

@arcprize @OpenAI Sorry, standard harness means codex, right?

[Andrey Kormilitsyn @AndreyKoms]

@TokenGremlin upd: I’m good, it came!😅

[Andrey Kormilitsyn @AndreyKoms]

@TokenGremlin Not for my plus sub unfortunately :(

[Andrey Kormilitsyn @AndreyKoms]

@_cmd8 @TheIshanGoswami That’s the point, give jev a real relevant context and facts he can rely on, and not be limited with knowledge cutoff

[Andrey Kormilitsyn @AndreyKoms]

@ThePrimeagen Pls Mr Primeagen, we are still waiting you to continue jev shenanigans, that was one of the entertaining coding streams I saw so far :)

[Andrey Kormilitsyn @AndreyKoms]

@shownotover @thsottiaux

[Andrey Kormilitsyn @AndreyKoms]

@awlevin @typesafeai Especially like the idea for ocr optimization with boxing part and reanalize only changed ones, brilliant

[Andrey Kormilitsyn @AndreyKoms]

@awlevin @typesafeai Thx for the repo, clarified a lot for me

[Andrey Kormilitsyn @AndreyKoms]

@MatthewBerman @bot @SpaceXAI I’m not so good at organizing workflows, reviewing 10+ pr’s is a pain. As Lauren/poteto sub I really like her articles to setup up pstack software factory, and really appreciate if I can get some of this spacex code to try to automate this stack!🙏

[Andrey Kormilitsyn @AndreyKoms]

Что-то не так сегодня. 5.6-luna max слииииишком медленная. обработка web запросов, просмотр картинки, запуск команд и анализ результата - черепаший, по сравнению с кем как luna max работала до релиза Astra..

[Doğukan @DogukanUrker]

Qwen3.8-27B on a single RTX 3060. after weeks of testing quants and configs, this is where I'm stopping -- these two are my final builds for this card. 158K ctx at 22 tok/s, or 82K at 38 tok/s. (config & weights below)

tested four IQ2 quants across seven benchmarks and GSQ-RCO IQ2_XS won, while being the second smallest file of the four.

the catch: GSQ-RCO ships no MTP head. you borrow Unsloth's -- same base model, drops straight in -- and pay ~76K of context for 1.7x decode. prefill is ~460 tok/s either way, MTP only touches generation.

which one is "worth it" depends entirely on what you run. short few-shot tasks never touch 64K, so the extra context is dead weight and you just want the speed. long single-session agent runs blow past 82K and no decode rate saves you once you're out of context. that's why both configs exist.

everything on low reasoning effort. at 2-bit the higher settings spiral instead of terminating.

🐢 context (158K):
llama-server -m Qwen3.8-27B-GSQ-RCO-IQ2_XS.gguf -ngl 99 -c 158000 -fa on --jinja -np 1 --cache-type-k q4_0 --cache-type-v q4_0 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --presence-penalty 0.0 --repeat-penalty 1.0 --reasoning on --chat-template-kwargs '{"reasoning_effort":"low"}'

⚡ MTP (82K):
llama-server -m Qwen3.8-27B-GSQ-RCO-IQ2_XS.gguf -md mtp-Qwen3.8-27B-Q4_0.gguf -ngl 99 -c 82000 -fa on --jinja -np 1 --spec-type draft-mtp --cache-type-k q4_0 --cache-type-v q4_0 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --presence-penalty 0.0 --repeat-penalty 1.0 --reasoning on --chat-template-kwargs '{"reasoning_effort":"low"}'

[Andrey Kormilitsyn @AndreyKoms]

@dummerspast39 Wrong

[Andrey Kormilitsyn @AndreyKoms]

@MiaAI_lab Rtx 3060 12gb was pretty common choice for gpu back then. Mia can you, if possible, give us 12gb plebs an option? 🙏

[Andrey Kormilitsyn @AndreyKoms]

Today I answered 144 questions in a @mattpocockuk wayfinder/grilling session. Codex, Sol high, 50 deep dive messages architecturing and integration outbox pattern in legacy netframework 4.7.2 app. What a journey! Thanks for the skill!

[Andrey Kormilitsyn @AndreyKoms]

@TJeparskis @Angaisb_ This thursday according to some rumors

[Andrey Kormilitsyn @AndreyKoms]

@ishuagra02 ez fix - setup automation (even in codex itself) to send little hi in codex and it starts a limit timer whenever you want. you can calculate the work start time and send the little auto-hi a few hours earlier, so the 5-hour limit is reset sooner.

[Andrey Kormilitsyn @AndreyKoms]

@0xSero rtx 3060 12gb included? 🙏 cant find any good info for this kind of capacity, seems like 12gb is not enough for any reasonable open model at all in 2026 🥲

[Andrey Kormilitsyn @AndreyKoms]

@thdxr ssi someone?

[Andrey Kormilitsyn @AndreyKoms]

@konradkokosa @KooKiz @RazorSharpFang And push it with github/autoesearch for more performance and efficiency! Yay

[Dillon Mulroy @dillon_mulroy]

@LukeParkerDev off the shelf/generic code review agents are dog water and all noise and no signal, but a well thought out/engineered one is insanely effective.

the reviewer we built at cf produces 90-95% signal and good feedback consistently

https://t.co/WnREoqizIV