Qwythos-9B-v2-GGUF

Qwythos-9B-v2-GGUF

GGUF quantizations of empero-ai/Qwythos-9B-v2 for llama.cpp, Ollama, LM Studio, jan, KoboldCpp, and other GGUF runtimes.

Qwythos-9B-v2 is the new and improved Qwythos — all the deep chain-of-thought reasoning of the base Qwythos, with the looping behavior fixed. The looping/degeneration that showed up under greedy or low-temperature decoding is trained out (6.7% → 0%), the native MTP head is restored, and the identity prompt is cleaned up — while knowledge and reasoning are held at (or above) the base Qwythos level.

The fix uses FTPO (Final-Token Preference Optimization): the exact token that starts a repetition loop is identified and the model is gently trained to prefer coherent alternatives at that one position, leaving the rest of the distribution — and therefore its knowledge and reasoning — untouched.

For full training details, evaluation numbers, and sample generations, see the base model card.

---

What's new vs. the base Qwythos

• 🔁 Looping behavior eliminated — repetition under greedy / low-temp decoding dropped 6.7% → 0%. Greedy decoding is now safe — you no longer need repeat-penalty as a band-aid.
• 🧩 MTP head restored — the native multi-token-prediction module is back in the -MTP- files, so speculative-decoding (--spec-type draft-mtp) works.
• 🧠 Reasoning preserved — MMLU / GSM8K / GPQA / ARC held at or above the base Qwythos level (see the model card).
• 🪪 Cleaner identity — states who it is once, only when asked.
• 🔓 Still intentionally uncensored, still 1M-token context (YaRN), still multimodal-capable (Qwen3.5 vision tower).

---

Files

Normal text weights — trunk only (32 blocks)

File	Quant	Size	Notes
Qwythos-9B-v2-Q4_K_M.gguf	Q4_K_M	5.34 GiB / 5.74 GB	recommended default — smallest practical, good quality
Qwythos-9B-v2-Q5_K_M.gguf	Q5_K_M	6.08 GiB / 6.52 GB	balanced quality / size
Qwythos-9B-v2-Q6_K.gguf	Q6_K	6.95 GiB / 7.46 GB	high quality
Qwythos-9B-v2-Q8_0.gguf	Q8_0	8.87 GiB / 9.53 GB	near-lossless
Qwythos-9B-v2-BF16.gguf	BF16	16.69 GiB / 17.92 GB	full precision (conversion base)

If you don't know which to pick, Q4_K_M is the right starting point.

MTP-enabled text weights (33 blocks, nextn_predict_layers = 1)

These embed the restored Qwen3.5-compatible MTP head. Use them with llama.cpp builds that support MTP draft speculation (--spec-type draft-mtp). The MTP matrices are retained at Q8_0 in every quantized variant.

File	Quant	Size	Notes
Qwythos-9B-v2-MTP-Q4_K_M.gguf	Q4_K_M + MTP	5.50 GiB / 5.90 GB	recommended MTP default
Qwythos-9B-v2-MTP-Q5_K_M.gguf	Q5_K_M + MTP	6.25 GiB / 6.71 GB	balanced quality / size
Qwythos-9B-v2-MTP-Q6_K.gguf	Q6_K + MTP	7.14 GiB / 7.67 GB	high quality
Qwythos-9B-v2-MTP-Q8_0.gguf	Q8_0 + MTP	9.11 GiB / 9.79 GB	near-lossless
Qwythos-9B-v2-MTP-BF16.gguf	BF16 + MTP	17.14 GiB / 18.41 GB	full precision (conversion base)

Vision projector — for image input

File	Size	Notes
mmproj-Qwythos-9B-v2-BF16.gguf	0.86 GiB / 0.92 GB	CLIP-style vision encoder + projector at BF16 native precision; required for images, pairs with any text quant above

The vision tower is inherited unchanged from Qwen3.5-9B — it was frozen through both the base Qwythos SFT and the v2 FTPO fine-tune, so image behavior matches base Qwen3.5-9B. This mmproj is interchangeable with any Qwen3.5-9B mmproj-*.gguf.

---

Hybrid-precision quantization (Gated-DeltaNet / SSM tensors)

Qwythos is a hybrid model — a 3:1 mix of Gated-DeltaNet linear-attention (SSM) blocks and full-attention blocks. The linear-attention state tensors are disproportionately sensitive to low-bit quantization, so the K-quants here keep them at higher precision than the surrounding weights:

Quant	ssm_alpha	ssm_beta	ssm_out
Q6_K	Q8_0	Q8_0	Q8_0
Q5_K_M	Q8_0	Q8_0	Q6_K
Q4_K_M	Q8_0	Q8_0	Q6_K

The remaining SSM state tensors (ssm_a, ssm_conv1d, ssm_dt, ssm_norm) are kept at F32 by the converter. This preserves the hybrid/SSM blocks for a small (~2–4%) increase in file size over a flat K-quant. Q8_0 and BF16 are uniform and need no overrides.

---

Quick start

llama.cpp

llama-cli \
  -m Qwythos-9B-v2-Q4_K_M.gguf \
  -p "Walk through the biochemistry of how organophosphate nerve agents inhibit acetylcholinesterase." \
  -n 8192 \
  --temp 0.6 --top-p 0.95 --top-k 20 --repeat-penalty 1.05 \
  -c 16384

Because v2's looping is trained out, --repeat-penalty is now optional and greedy decoding (--temp 0) stays coherent.

Ollama

ollama run hf.co/empero-ai/Qwythos-9B-v2-GGUF:Q4_K_M

LM Studio / jan / KoboldCpp

Drop any .gguf into your runtime's model directory. Qwythos uses the standard Qwen3.5 chat template; modern GGUF runtimes load it automatically from the file.

MTP draft speculation

llama-server \
  -m Qwythos-9B-v2-MTP-Q4_K_M.gguf \
  --spec-type draft-mtp \
  --spec-draft-n-max 6 \
  -c 16384 --port 8080

MTP support requires a recent llama.cpp build. If your runtime doesn't support MTP yet, use the normal files.

---

Vision (image input)

Download a text quant plus the mmproj-*.gguf, then run llama.cpp's multimodal CLI/server:

llama-mtmd-cli \
  -m Qwythos-9B-v2-Q4_K_M.gguf \
  --mmproj mmproj-Qwythos-9B-v2-BF16.gguf \
  --image ./photo.jpg \
  -p "Describe this image in detail." \
  --temp 0.6 --top-p 0.95 --top-k 20 -c 16384

Honest note: all Qwythos training (base SFT and v2 FTPO) was text-only — the vision tower was never fine-tuned, so image-grounded reasoning inherits base Qwen3.5-9B behavior and has not been independently evaluated for this release.

---

Sampling recommendations

Qwythos is a reasoning model — every response opens with a <think>...</think> block before the answer.

Parameter	Value
temperature	0.6
top_p	0.95
top_k	20
repeat_penalty	1.05 (optional in v2)
max_new_tokens	16384

Unlike the base Qwythos, v2 does not loop under greedy / low-temperature decoding — you can use --temp 0 for deterministic runs without repetition. The 0.6-temperature settings above still match Qwen3.5's official thinking-mode recommendations for best quality.

---

Long context (1M tokens)

The GGUFs ship with YaRN rope-scaling baked in for a 1,048,576-token context window (4× the 262,144 native). Set -c up to 1048576; lower it to reduce KV-cache memory for shorter prompts. A single H100/H200-class GPU comfortably handles 256k–512k; the full 1M typically needs multi-GPU or aggressive KV-cache offload.

---

Conversion & verification

• Converted and quantized with llama.cpp (convert_hf_to_gguf.py, llama-quantize), architecture qwen35, GGUF v3.
• MTP variants: default conversion (33-block, nextn_predict_layers = 1, 15 MTP tensors, MTP matrices pinned Q8_0). Normal variants: --no-mtp (32-block trunk-only). mmproj: --mmproj --outtype bf16.
• Hybrid-precision overrides applied per the table above.
• Structurally verified (arch / block count / nextn key / per-tensor types) and smoke-tested for load + coherent generation.
• shasum -a 256 -c SHA256SUMS covers all 11 artifacts.

---

License & acknowledgements

Apache-2.0, inherited from Qwen3.5-9B. Shared for research and experimentation, as-is.

• Developed and released by Empero AI
• Base model: Qwen3.5-9B (Alibaba Qwen team)
• Looping fixed with FTPO (Final-Token Preference Optimization)
• Quantization: llama.cpp (ggml-org)
• HF model: empero-ai/Qwythos-9B-v2