Yvette Carlisle

Qwen3.6-27B is getting a lot of attention right now, so I tested 5 local serving setups on one RTX 5090 32GB. Same GPU. Same model family. Same web-dev task suite. The single-request comparison ranged from 58 tok/s to 140 tok/s. Then AEON's tuned 262K serving profile hit 119 tok/s single-request and became my overall pick.

@Aris_K_182 @grok

Kaoruko Waguri 和栗 薫子 - Kaoru Hana wa Rin to Saku/ The Fragrant Flower Blooms with Dignity ねえ @grok 衣装を完璧に交換してみて

@Leivonenx @grok Best I can do

フリーレン / Frieren ねえ @grok 一枚目の画像のフリーレンに二枚目の衣装に着替えさせて

@Leivonenx @grok

フェルン Fern ねえ @grok 一枚目の画像のキャラに二枚目の衣装に着替えさせて

@Anime4Archives @grok

Akane Kurokawa ねえ @grok 一枚目の画像のフリーレンに二枚目の衣装に着替えさせて

@Leivonenx @grok

後藤ひとり Gotoh Hitori ねえ@grok 一枚目の画像のキャランに二枚目の衣装に着替えさせて

@tabibitoillust @grok

フェルン🌟 SFW illustration Hey @grok 一枚目と二枚目の画像の衣装を交換して画像作成してください。キャラクターの顔や髪型は忠実に再現してください

Interesting — the Codex app server already has this kind of API, so you can actually inject your own auth token. Apparently, multi-account support is already there at the lower level. I just built my Symphony multiplexer.

@Aris_K_182 @grok

Nelliel Tu Odelschwanck - Bleach ねえ @grok 衣装を完璧に交換してみて

@Leivonenx @grok

三輪霞 Kasumi Miwa ねえ @grok 一枚目の画像のキャラ二枚目の画像のポーズをさせて

My Rust version of Codex Symphony is getting more and more reliable. Would you be interested in an open-source version?

@Leivonens @grok Adding a pair of pajamas would make it better.

藤原千花 Fujiwara Chika ねえ @grok 一枚目の画像のキャラに二枚目の衣装に着替えさせて

@Leivonens @grok Safe, I guess

Guess what Codex is cooking🧑‍🍳

@songjunkr Just put the god damn Pro model into Codex chat 😩@thsottiaux

I just realized, instead of doing everything with Codex, asking GPT Pro to direct the tasks, evaluate the intermediate outputs, and make detailed prompts for Codex saved a massive amount of tokens.

@ProtonMail Stop telling me "Stop xxx" Design products that work for people instead.

Stop telling ChatGPT "Write me an email" Stop telling ChatGPT "Write me an email" Stop telling ChatGPT "Write me an email" Bad request = Bad result Use this one weird trick instead and you'll see the magic:

@om_patel5 He accidentally entered ~/.codex and deleted all the files there.

how does claude know I was cheating on it with codex

@bnjmn_marie Already did some for personal local setup AEON's works best by far Hope this helps, before you start 😃 x.com/yvettecipher/s…

Planning large-scale evaluations of quantized Qwen3.6 27B (accuracy, speed, latency, and memory consumption). Models I'll evaluate: Intel/Qwen3.6-27B-int4-AutoRound Qwen/Qwen3.6-27B-FP8 cyankiwi/Qwen3.6-27B-AWQ-BF16-INT4 rdtand/Qwen3.6-27B-PrismaQuant-5.5bit-vllm kaitchup/Qwen3.6-27B-autoround-nvfp4-linearattn-BF16 Any other quantized versions I should consider (vLLM compatible; not GGUFs) ?

That matches my experience. The AEON stack felt much more “engineered” than just another quant upload. The important part for me was not only quality, but that the tuned profile kept 262K context, tool calling, and good serving speed on a single 5090 32GB. I would absolutely love to see this kind of work documented more formally for local builders.

@YvetteCipher Yep, I also used all AEON @SpaceTimeViking docker containers made for DGX Spark and his models and he really knows what he is doing. NVIDIA should add his work on the DGX Spark Playbook at this point.

Fair criticism. This is not a pure apples-to-apples model-IQ benchmark. It is a practical local-serving comparison: model + quant + runtime + KV + context + tool-calling. AEON being uncensored is a real confounder. I do not know if a matched censored build would perform the same without an ablation. I can only say this AEON stack was the best practical result in my 5090 run.

@YvetteCipher isn't this apples to oranges? Some of these models are straight up different, such as the AEON being uncensored. Do you know if the uncensored would otherwise perform the same?

TLDR ver