Empero

🚀 Qwythos-27B-v1 is here! The 27B you've been waiting for. The bigger sibling to Qwythos-9B. Native MTP head intact, full vision tower, still uncensored, still 1M context. Apache-2.0. huggingface.co/empero-ai/Qwyt…

@sachindetrax @0x0SojalSec We ran 20 side by side trainings to compare both base models, 3.6 degrade a lot after SFT training while 3.5 takes training well. We have had 18/20 degraded runs on 3.6

@0x0SojalSec i think this one is better because it is trained on qwen 3.6 - huggingface.co/DavidAU/Qwen3.…

Top uncensored model right now trained on real Mythos traces. A 27B-v1 model Uncensored model smoked the base Qwen3.5 by 34 points on MMLU while thinking like Claude - Claude-style deep thinking - Trained on half a billion tokens of Claude Mythos chain-of-thought. - the terminal/tool perplexity went from 356 to 2.76. - pure reasoning fuel. - 1M context window. - Native function calling & vision - Self-correcting tool use. - Runs smooth on consumer hardware.

@pirolaoficial We have planned exactly that!

@EmperoAI Could you please bonsai it to 8gb? Thanks

🚀 Qwythos-27B-v1 is here! The 27B you've been waiting for. The bigger sibling to Qwythos-9B. Native MTP head intact, full vision tower, still uncensored, still 1M context. Apache-2.0. huggingface.co/empero-ai/Qwyt…

@WFD_ryufps Yes

@EmperoAI Hey guys just a request can do the same for 4b mode,l i know 9b exists but due to some hardware limitations its a bit slow so,can you pls make qwen 3.5 4b with either fable/mythos 5

@juanginerpu For the later release after reinforcement learning we will release a detailed benchmark

@EmperoAI Could we get some benchmarks vs the original model?

@MgkMshrmBrkfst Its a good generalist base and in our tests 3.6 took training worse then 3.5 due to its heavy RL

@EmperoAI qwen3.5 is not the best base

@sachindetrax We are training a QAT 2 bit, 1.58 bit and 1 bit in the coming weeks

@EmperoAI can we have a q3 or q2 version something that we can run on 16gb vram?

@DogukanUrker Yes! It’s the next one training

@EmperoAI Congratulations!! Any plans to drop 35b-a3b?

11 GGUF files: Q4_K_M / Q5_K_M / Q6_K / Q8_0 / BF16, each in trunk-only and MTP-enabled builds, plus mmproj for image input. SHA256SUMS included. MTP builds do self-speculative decoding via --spec-type draft-mtp. huggingface.co/empero-ai/Qwyt…

GGUFs are purpose-built, not a default dump. In every K-quant the whole Gated-DeltaNet state path, ssm_alpha, ssm_beta and ssm_out is held at Q8_0 or better. Those tensors hate low-bit. Costs ~2-4% size. Q4_K_M is 16.95 GB and runs on a 24GB card.

Running Qwythos 9B on a single RTX 3060: 500K context at ~52 tok/s. (config below) a Qwen3.5-9B finetune with 1M YaRN baked in -> weights are only 5.4GB, so the entire fight is KV cache. q4_0 on both K and V halves it -> same card went from 290K to 500K on the KV quant alone. 100% on the GPU, zero CPU offload, pinned at ~96% of the 12GB. llama-swap config: llama-server -m Qwythos-9B-Claude-Mythos-5-1M-Q4_K_M.gguf -ngl 99 -c 500000 -fa on --jinja -np 1 --cache-type-k q4_0 --cache-type-v q4_0 -b 1024 -ub 512 --temp 0.6 --top-p 0.95 --top-k 20 --repeat-penalty 1.05 --reasoning on

We are currently looking for an arXiv endorsement in the domain cs.LG or cs.CL please let us know if you would be able to endorse our paper :)

Open Source will become more important by the day

We are very busy with some very exciting advancements!

@AxcanNathan Very soon it will be ready!

@EmperoAI do qwen27b and you will find glory

Thank you all for 2 million downloads! We have just published new GGUF quantizations with a fixed chat template for de-escalated branding, if you still use Qwythos v1 consider redownloading the GGUFs here: huggingface.co/empero-ai/Qwyt…

@BachelotBernard Very glad to hear that! We are doing custom models too if required :)

@EmperoAI I’m a project manager for cultural associations; I just installed V2—I used to use Gemma 4, and I have to say, I’m really blown away by qwythos-9b v2 gguf.

🚀 Qwythos-9B-v2 is here — the new & improved Qwythos. Same deep reasoning as before, but the looping behavior is fixed: 6.7% -> 0% under greedy decoding. MTP head restored, cleaner identity, still uncensored, still 1M context. Apache-2.0. huggingface.co/empero-ai/Qwyt…

thank you all for 1000 followers on @huggingface 🤗