We just shipped Qwythos-27B-v1 the long-awaited bigger sibling to Qwythos-9B, now on a Qwen3.5-27B base. It is an open-weights bf16 reasoning model, released under Apache-2.0, with weights and purpose-built GGUFs up now. People have been asking for the 27B since the 9B landed. This is it.
Qwythos-27B-v1 is the bigger Qwythos, from the same curriculum family as the 9B and sized up on Qwen3.5-27B. The base is dense and hybrid-attention: Gated-DeltaNet linear attention, with a full-attention layer every fourth layer. There are 64 layers, a hidden size of 5120, 24 attention heads, 4 KV heads, and a 248,320-token vocabulary.
It also keeps Qwen3.5's native function-calling specification. Pass tools=[...]
and it works without a wrapper. That matters to us: a reasoning model should be able to use real tools when the answer needs checking, rather than make a performance out of remembering.
The weights are Apache-2.0. The model is developed by Empero, and the repositories are Qwythos-27B-v1 and Qwythos-27B-v1-GGUF.
Nothing was ablated
This is the spine of the release. At 27B, community fine-tunes often lose at least one of the useful parts of the base. We did not want a larger model that had quietly become less capable at the edges. Qwythos-27B keeps all three intact and active.
Native MTP. The native multi-token-prediction head is present with nextn_predict_layers=1
. Runtimes that support it can use the model for self-speculative decoding. The GGUF release includes MTP-enabled variants for exactly this reason.
Full vision. Image input works, and we ship the mmproj
. This is the full vision tower, not a text-only substitute described as multimodal after the fact.
A real 1M context configuration. Qwythos uses YaRN at 4x over the 262,144-token native window, for 1,048,576 tokens. One million tokens sounds like a line item until you are working through a long tool trajectory, a pile of documents, or a codebase that should fit in one prompt. Then it is simply room to keep the thread intact.
What the size-up buys
There is a small closed-book example that makes the difference more concrete than a capability adjective. On an organophosphate-poisoning prompt, the 27B volunteered: “avoid physostigmine (it can worsen toxicity).” It also gave correct atropine/pralidoxime dosing, without tools. In the corresponding 9B evaluation, that same fact required a web search to get right.
That is not a case for skipping verification in clinical use. It is a useful illustration of what the larger model can retain without reaching for a tool, and why the 27B is more than the 9B with a larger number on the label.
We also measured target-format acquisition on held-out terminal/tool sessions. Perplexity against the base moved from 356.6 to 2.76. This is not a standardized academic benchmark and we are not treating it like one. It is, however, directly relevant to the kind of long-form terminal and tool work we want the model to handle.
One more honest note: standardized MMLU and GSM8K benchmarks have not been run on this checkpoint. We are not borrowing the 9B's numbers, and there is no benchmark table here. The useful claims for this release are the claims we have measured.
Running it
For local use, start with the GGUF release. Q4_K_M
is 16.95 GB and runs on a 24GB card. There are 11 files in total: Q4_K_M, Q5_K_M, Q6_K, Q8_0, and BF16, each in trunk-only and MTP-enabled variants, plus mmproj-Qwythos-27B-F16.gguf
. We ship SHA256SUMS too.
| If you need | Start here |
|---|---|
| A 24GB card | Q4_K_M (16.95 GB) |
| Self-speculative decoding in llama.cpp | An MTP-enabled variant with --spec-type draft-mtp |
| Image input | The shipped mmproj-Qwythos-27B-F16.gguf |
There is one non-default choice in these quantizations that is worth being explicit about. In every K-quant, the whole Gated-DeltaNet state path — ssm_alpha
, ssm_beta
, and ssm_out
— remains at Q8_0 or better. Those state tensors are disproportionately sensitive to being pushed into Q4, Q5, or Q6; a naive conversion degrades quality there. Holding them higher costs about 2–4% in file size. We think it is the correct trade.
For long-context serving with vLLM:
VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 vllm serve empero-ai/Qwythos-27B-v1 \
--max-model-len 1010000
Qwythos is a reasoning model, so give it room. Our starting settings are temperature=0.6
for agentic or tool use, or 1.0
for open-ended and creative work; top_p=0.95
; top_k=20
; repetition_penalty=1.05
; and max_new_tokens=16384
or more.
The one throughput number we are publishing is deliberately narrow: Q4_K_M, fully offloaded to a single RTX 5090 with 32 GB, at 4k context in llama.cpp CUDA, generated 74 tok/s and processed prompts at 622 tok/s. One card, one data point, not a benchmark sweep.
Uncensored on purpose
Qwythos is intentionally uncensored for technical and research work. It engages seriously with cybersecurity, biomedical, and technical questions where over-aligned models often refuse or hedge. That is a deliberate choice for researchers, not a claim that deployment needs no thought. If you put it in front of end users, add your own review layer.
v1, before RL
This is v1, the checkpoint before reinforcement learning. We are releasing it as a complete and usable model, not as a teaser for a future one. An RL-trained v2 follows soon.
If you build with it, we want to hear what holds up and what does not. The weights and GGUFs are live now; the Empero newsletter at empero.org is where the next release notes will land.
— kodee