Empero

building a better tomorrow

[Empero @EmperoAI]

Releasing Qwythos-9B, a full SFT of Qwen3.5-9B on 500M+ tokens of Claude Mythos & Fable traces with our in-house CoT generation via "rethink."

vs base, identical eval setup:
→ MMLU +34
→ gsm8k strict +30
→ gsm8k flex +19

Native tool use. 1M context. Uncensored. 🧠

[Empero @EmperoAI]

soon.. https://t.co/9doHvy3jkq

[Empero @EmperoAI]

runmonitor: watch your training run unfold live. 🟣

• local-first; nothing leaves your machine, no API keys
• lives in your loop: run.log({"loss": loss}, step)
• live curves + anomaly detection in a terminal-style dashboard

> pip install runmonitor

[Empero @EmperoAI]

@chetaslua Thats why we keep releasing and researching. Intelligence needs to be liberated not consolidated in the hands of a few.

[ローカルAIラボ @localai_lab]

速報です。
Qwythos-9B-Claude-Mythos-5-1M Q4_K_M、
llama.cppで旅行プラン作成タスクを実行できました。

環境:
GTX 1660 Ti 6GB
llama.cpp
Q4_K_M
reasoning off

日本語の旅行プランは普通に返ってきました。
生成速度は 7.4 tok/s。

実行中のVRAMは nvidia-smi で 4475MiB / 6144MiB。
4GB GPU実機での確認ではありませんが、VRAM 4GB台で動く話はかなり現実味ありそうです。

詳細なレビューと、別モデルとの比較はまた記事にします。
#ローカルLLM #llamacpp

[Empero @EmperoAI]

@UnslothAI Qwythos all the way! It actually replaced Claude in some non critical applications for us internally.

[Empero @EmperoAI]

We have not been able to do so yet but Qwythos has written its own harness and its quiet impressive! It picked the name Abacus Agent, check it out here: https://t.co/finfPdaB7p

[Empero @EmperoAI]

We currently have Qwythos write a whole agentic coding harness by itself inside of Codex. The progress is going very well, the finished agent will be published soon.

[Empero @EmperoAI]

@Polymarket On this note: Everyone who wants to donate sessions to liberate intelligence message us!

[Empero @EmperoAI]

The next big move in AI isn't scaling parameters forever. It's condensing intelligence.

We can't scale horizontally indefinetly. The real breakthrough is packing dramatically more knowledge and capability into every single parameter.

[Empero @EmperoAI]

As predicted.

[Empero @EmperoAI]

The next few months in the ML space will be very exciting.

[Empero @EmperoAI]

@notnullptr I dont understand how you see this as noise when you haven't engaged with the model itself, nor even read the evaluation or documentation.

[Empero @EmperoAI]

@notnullptr Of course it's an SVG because it looks better then the raw json, further detail on the eval and how to reproduce it is listed in our documentary: https://t.co/v238KLw31V

[Empero @EmperoAI]

@notnullptr Our 3.6 27B Fine-Tune is currently training! I do not understand why you have to be so adversarial? You can test the model our check our samples: https://t.co/PmFguvagNI
And for tooling: https://t.co/eRSWeRY6Jr

[Empero @EmperoAI]

@JakoveHr @notnullptr From user experience it works well in a harness, we are also currently having it write its own agentic harness as a little test!

If you want to see our tool tests you can look at the results here: https://t.co/eRSWeRY6Jr

[Empero @EmperoAI]

@bindureddy Smaller models are getting more capable by the day aswell!
https://t.co/pbD9VBjc9o

[Empero @EmperoAI]

Releasing Qwythos-9B, a full SFT of Qwen3.5-9B on 500M+ tokens of Claude Mythos & Fable traces with our in-house CoT generation via "rethink."

vs base, identical eval setup:
→ MMLU +34
→ gsm8k strict +30
→ gsm8k flex +19

Native tool use. 1M context. Uncensored. 🧠

[Empero @EmperoAI]

@VibeCodeAiden Here would be some sample generations if you want to look into the model without downloading it!

https://t.co/RMuMwsjcab

https://t.co/PmFguvagNI

https://t.co/eRSWeRY6Jr

[Empero @EmperoAI]

Releasing Qwythos-9B, a full SFT of Qwen3.5-9B on 500M+ tokens of Claude Mythos & Fable traces with our in-house CoT generation via "rethink."

vs base, identical eval setup:
→ MMLU +34
→ gsm8k strict +30
→ gsm8k flex +19

Native tool use. 1M context. Uncensored. 🧠

[Empero @EmperoAI]

@VibeCodeAiden We can not disclose our full methodology here, but rethink uses transcripts of real claude sessions, extracts the assistant turns and uses a complex system of LLMs in various roles like writer, judge, etc along machine checks to write a CoT to arrive at the the produced output.

[Empero @EmperoAI]

@loktar00 If you look for a smaller model with big capabilities check out our latest release!
https://t.co/pbD9VBjc9o

[Empero @EmperoAI]

Releasing Qwythos-9B, a full SFT of Qwen3.5-9B on 500M+ tokens of Claude Mythos & Fable traces with our in-house CoT generation via "rethink."

vs base, identical eval setup:
→ MMLU +34
→ gsm8k strict +30
→ gsm8k flex +19

Native tool use. 1M context. Uncensored. 🧠

[Empero @EmperoAI]

@udiWertheimer You should check out our latest release!
https://t.co/pbD9VBjc9o

[Empero @EmperoAI]

Releasing Qwythos-9B, a full SFT of Qwen3.5-9B on 500M+ tokens of Claude Mythos & Fable traces with our in-house CoT generation via "rethink."

vs base, identical eval setup:
→ MMLU +34
→ gsm8k strict +30
→ gsm8k flex +19

Native tool use. 1M context. Uncensored. 🧠

[Empero @EmperoAI]

@birdabo I mean you can use Qwythos for free even https://t.co/pbD9VBjc9o

[Empero @EmperoAI]

Releasing Qwythos-9B, a full SFT of Qwen3.5-9B on 500M+ tokens of Claude Mythos & Fable traces with our in-house CoT generation via "rethink."

vs base, identical eval setup:
→ MMLU +34
→ gsm8k strict +30
→ gsm8k flex +19

Native tool use. 1M context. Uncensored. 🧠