Please download the newest version for the best experience! If you downloaded Qwythos previously consider to download v2 for a rounder experience.
Fixes include:
🔁 Looping behavior eliminated — repetition/degeneration under greedy or low-temperature decoding dropped from 6.7% → 0%. You can serve it without leaning on repetition_penalty as a band-aid.
🧠 Reasoning fully preserved — MMLU, GSM8K, GPQA, ARC and HumanEval are all held at (or above) the v1 level. This is a hygiene upgrade, not a capability regression.
🧩 MTP head restored — the native multi-token-prediction module (dropped in the previous export) is back, so config and weights agree and speculative-decoding setups work.
🪪 Cleaner identity — the model no longer prefaces unrelated answers with its identity; it introduces itself only when you actually ask.
🔓 Still intentionally uncensored for research, cybersecurity, red-teaming, biology, chemistry, pharmacology and clinical work.
📜 Still 1M-token context (YaRN) and the native multimodal-capable Qwen3.5 stack.
Qwythos-9B is a full-parameter reasoning model post-trained on over 500 million tokens of high-quality Claude Mythos / Claude Fable traces with chain-of-thought generated in-house by Empero AI's internal rethink tool. It dominates the base Qwen3.5-9B under matched evaluation (+34 pts MMLU, +30 pts gsm8k-strict, +19 pts gsm8k-flex), supports native function calling per the Qwen3.5 spec, and ships with a 1,048,576-token (1M) context window via YaRN rope-scaling enabled by default.
For full training details, evaluation numbers, and capability writeup, see the base model card.
Files
Normal text weights — fixed v3 replacements
File
Quant
Size
Notes
Qwythos-9B-Claude-Mythos-5-1M-Q4_K_M.gguf
Q4_K_M
5.24 GiB / 5.63 GB
recommended default — fixed v3, best compatibility
Qwythos-9B-Claude-Mythos-5-1M-Q5_K_M.gguf
Q5_K_M
6.02 GiB / 6.47 GB
fixed v3, balanced quality / size
Qwythos-9B-Claude-Mythos-5-1M-Q6_K.gguf
Q6_K
6.85 GiB / 7.36 GB
fixed v3, high quality
Qwythos-9B-Claude-Mythos-5-1M-Q8_0.gguf
Q8_0
8.87 GiB / 9.53 GB
fixed v3, near-lossless
Qwythos-9B-Claude-Mythos-5-1M-BF16.gguf
BF16
16.69 GiB / 17.92 GB
fixed v3, full precision conversion base
If you don't know which to pick, Q4_K_M is the right starting point — it's the smallest practical quant with good quality preservation.
MTP-enabled text weights — fixed v3 variants
These include the restored Qwen3.5-compatible MTP head inside the GGUF. Use them with llama.cpp builds that support MTP draft speculation, for example --spec-type draft-mtp.
File
Quant
Size
Notes
Qwythos-9B-Claude-Mythos-5-1M-MTP-Q4_K_M.gguf
Q4_K_M + MTP
5.48 GiB / 5.89 GB
recommended MTP default
Qwythos-9B-Claude-Mythos-5-1M-MTP-Q5_K_M.gguf
Q5_K_M + MTP
6.26 GiB / 6.73 GB
MTP, balanced quality / size
Qwythos-9B-Claude-Mythos-5-1M-MTP-Q6_K.gguf
Q6_K + MTP
7.09 GiB / 7.62 GB
MTP, high quality
Qwythos-9B-Claude-Mythos-5-1M-MTP-Q8_0.gguf
Q8_0 + MTP
9.11 GiB / 9.79 GB
MTP, near-lossless
Qwythos-9B-Claude-Mythos-5-1M-MTP-BF16.gguf
BF16 + MTP
17.14 GiB / 18.41 GB
MTP, full precision conversion base
Vision projector — for image input
File
Size
Notes
mmproj-Qwythos-9B-Claude-Mythos-5-1M-F16.gguf
0.86 GiB / 0.92 GB
CLIP-style vision encoder + projector; required for images, pairs with any normal or MTP quant above
Qwythos inherits its vision tower from the Qwen3.5-9B base model — the vision path was frozen during SFT (training was text-only), so the vision behavior is identical to base Qwen3.5-9B's multimodal capability. The mmproj is interchangeable with any community-built Qwen3.5-9B mmproj-*.gguf.
Quick start
llama.cpp (llama-cli)
bash
1llama-cli \2 -m Qwythos-9B-Claude-Mythos-5-1M-Q4_K_M.gguf \3 -p "Walk through the biochemistry of how organophosphate nerve agents inhibit acetylcholinesterase."\4 -n 8192\5 --temp 0.6 --top-p 0.95 --top-k 20 --repeat-penalty 1.05\6 -c 16384
Ollama
ollama run hf.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF:Q4_K_M
LM Studio / jan / KoboldCpp
Drop any of the .gguf files into your runtime's model directory. Qwythos uses the standard Qwen3.5 chat template; modern GGUF runtimes load it automatically from the file.
MTP support requires a recent llama.cpp build. If your runtime does not support MTP yet, use the normal fixed v3 files above.
Vision (image input)
Qwythos supports image input out of the box. Download both a text quant and the mmproj-*.gguf file from this repo, then run with llama.cpp's multimodal CLI or server.
Then POST to /v1/chat/completions with an image URL or base64 payload — the standard OpenAI vision API shape works.
LM Studio
Load the text quant; LM Studio detects the matching mmproj-*.gguf in the same folder and enables the image-attach button automatically.
What vision unlocks
Since Qwythos inherits its vision tower unchanged from Qwen3.5-9B base, expect Qwen3.5-9B's documented vision capabilities: detailed image description, OCR (printed + handwritten), chart/table reading, UI/document understanding, basic spatial reasoning.
Honest note: the SFT used to produce Qwythos was text-only — we did not fine-tune the vision tower or train on any image-paired data. Image-grounded reasoning therefore inherits the base model's behavior; it has not been independently evaluated as part of this release. If your application is primarily vision-driven, validate on your own use case first.
Sampling recommendations
Qwythos is a reasoning model — every response opens with a <think>...</think> block before the final answer. Use these settings as defaults:
Parameter
Value
temperature
0.6
top_p
0.95
top_k
20
repeat_penalty
1.05
max_new_tokens
16384 (generous budget for <think> + answer)
These match Qwen3.5's official thinking-mode recommendations. Avoid greedy decoding and very-low-temperature sampling (T ≤ 0.3) — both can cause repetition loops on long reasoning generations.
Long context (1M tokens)
The GGUFs ship with YaRN rope-scaling baked in for a 1,048,576-token context window (4× extension over the 262k native).
To use the full 1M window in llama-cli, set -c 1010000 (or any context length up to that). For shorter prompts, lower -c to reduce KV-cache memory — at default settings llama.cpp will autosize.
A single H100/H200-class GPU comfortably handles 256k–512k; the full 1M typically needs tensor-parallel multi-GPU or aggressive KV-cache offload.
Capabilities (from the base model card)
+34 pts MMLU, +30 pts gsm8k-strict, +19 pts gsm8k-flex vs. base Qwen3.5-9B under matched lm-eval-harness evaluation
Native function calling per Qwen3.5's chat-template spec — emits <tool_call><function=NAME><parameter=NAME>VAL</parameter></function></tool_call> blocks ready for any tool-use loop
Self-correcting with tools: in a 7-prompt tool-use harness (Python executor + DuckDuckGo search), Qwythos produced source-cited correct answers on 7/7, including 4/4 closed-book failure-modes from the original review
Uncensored — engages seriously with technically demanding questions across cybersecurity, red-teaming, biology, pharmacology, and clinical medicine
1,048,576-token (1M) context — YaRN rope-scaling enabled by default
Reasoning model. Every answer opens with a <think> block; allow generous max_new_tokens and parse/strip <think>...</think> for end users.
Use recommended sampling. Greedy / very-low-temp can cause repetition loops.
Verify specifics in safety-critical contexts. Like all closed-book LLMs in this weight class, Qwythos can over-commit to specific identifiers (CVEs, hashcat modes, drug positions) it isn't certain about. Pair with retrieval or function calling in such deployments — the model uses tools cleanly when offered them.
Uncensored — add your own application-level review/safety layer for end-user-facing deployments where that matters.
Stay in the loop
Sign up for the Empero newsletter at empero.org for releases, evals, and research notes.
Support / Donate
If this model helped you, consider supporting the project:
