Ornith-1.0-9B-MXFP4_Hybrid-Imatrix-GGUF

💎 Ornith-1.0-9B - Custom Mixed Precision GGUFs with Imatrix

Ornith-1.0, a self-improving family of open-source models specially for agentic coding tasks.
Built on top of pretrained Gemma 4 and Qwen 3.5, it achieves state-of-the-art performance among open-source models of comparable size on coding benchmarks.

This repository contains custom, highly optimized, multi-tier mixed precision GGUF weights for ornith-ai/Ornith-1.0-9B.

[!NOTE] Ornith-1.5 9B is now available!

Ornith-1.0 9B achieves state-of-the-art performance among open-source models of comparable size across a broad range of agentic coding benchmarks.

[!TIP] Highly Recommended: Always keep reasoning/thinking enabled.
Ornith thoroughly plans and reasons through code edits before execution, ensuring an efficient and clean output.
Unlike baseline Qwen models, which frequently execute blindly and backtrack after generating broken code.

<img style="width: 100%; max-width: 900px;" src="https://ornith.ai/ornith/assets/ornith_9b_eval_202606252204.png" alt="Ornith 1.0 9B Benchmark Results" title="Ornith 1.0 9B Benchmark Results">

To learn more about Ornith 1.0, read their blog post.
To learn more about how to use Ornith 1.0 9B, view the base model.
A larger variant is also available: Ornith-1.0-35B-A3B

These quants were generated using manual layer targeting to maximize quality while shrinking the massive VRAM footprint of the Mixture of Experts layers.

📄 GGUF Files

In order of quality:

📊 Importance Matrix (Imatrix)

<details> <summary>Expand to view datasets & details</summary>

The following datasets were used for the imatrix:

• I am currently remaking this imatrix with 2048 ctx, 3000 chunks at 3mb chunks, with smaller sets of text, and:

  • Ornith-generated agent skills markdown text & project documentation markdown text

• Old/Current: Custom Target Matrix (2048 ctx with 600 chunks), up to a max of 100MB of each:

  • eaddario/imatrix-calibration - tools_huge
  • Glint-Research/Fable-5-traces
  • lordx64/fable-sft-combined-v2
  • angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k
  • nohurry/Opus-4.6-Reasoning-3000x-filtered
  • eaddario/imatrix-calibration - code_huge, math_huge
  • osunlp/QUEST-RL-Data
  • osunlp/QUEST-SFT-Data-Open-ended
  • m-a-p/CodeFeedback-Filtered-Instruction
  • TuringEnterprises/Rubric-Graded-Reasoning - Rubric-CS, Rubric-DS

</details>

---

🔍 Precision Matrix & Flavor Variations

Standard global quantization presets (like stock MXFP4) compress the backbone layers uniformly, which degrades the delicate reasoning capabilities of advanced agent models.
This repository provides multiple distinct manual configuration layouts to balance precision and memory constraints:

1. The Tri-Quant Hybrid Flavor (MXFP4 + Q8_0 + F16)

Ornith-1.0-9B-MXFP4_Q8_0_F16-Imatrix.gguf - Designed for maximum quality preservation, this layout implements a strict 3-Tier Precision Matrix:

• Tier 1 (Core & Mamba Gating - F16 Precision):
  • token_embd.weight, output.weight - Protects the critical input/output vocabulary mappings. Dramatically prevents text degradation.
  • ssm_alpha, ssm_beta - Protects the integrity of the Mamba state-space calculations across long-range context tokens.
• Tier 2 (Backbone & Shared - Q8_0 Precision): ssm_out, *._shexp - Keeps the attention mechanics, and all trailing shared experts at high quality, to protect the logical research loops.
• Tier 3 (Routed Experts - MXFP4 Precision): ffn_down_exps, ffn_gate_exps, ffn_up_exps - Shrink the massive background expert parameters directly to MXFP4.

2. The Dual-Quant Hybrid Flavor (MXFP4 + Q8_0)

Ornith-1.0-9B-MXFP4_Q8_0-Imatrix.gguf - Designed for a slightly leaner memory profile, this layout utilizes 2-Tier Precision:

• Tier 1 (Backbone - Q8_0 Precision): All attention blocks, Mamba structures, vocabulary embeddings, and internal routers use the universal Q8_0 format.
• Tier 2 (Experts - MXFP4 Precision): The heavy sparse expert blocks are target-quantized directly to MXFP4.

3. Bonus Single-Quant (MXFP4)

Ornith-1.0-9B-MXFP4-Only-Imatrix.gguf - Using only MXFP4, this shrinks the model down to 4.77 GB. The quality is not the best, but it can still do decent work.

• Single Tier (All Layers - MXFP4 Precision): All layers are target-quantized directly to MXFP4, for speed and a low VRAM footprint.

---

📝 Exact Conversion Details

These files were converted via llama-quantize utilizing the following manual recipe parameters:

Convert SafeTensors to GGUF:

# Requires python3.12, with `pip install --upgrade transformers`
# --no-mtp is needed otherwise llama-quantize errors
python convert_hf_to_gguf.py "Ornith-1.0-9B/" --outtype f16 --no-mtp --outfile "Ornith-1.0-9B_F16.gguf"

Generate Tri-Quant MXFP4 + Q8_0 + F16:

llama-quantize \
  --tensor-type ".*_shexp\.weight=Q8_0" \
  --tensor-type "token_embd\.weight=F16" \
  --tensor-type "^output\.weight=F16" \
  --tensor-type "blk\..*\.(ssm_alpha|ssm_beta)\.weight=F16" \
  --tensor-type "blk\..*\.(ffn_down_exps|ffn_gate_exps|ffn_up_exps)\.weight=MXFP4" \
  --imatrix "imatrix.gguf" \
  "Ornith-1.0-9B_F16.gguf" \
  "Ornith-1.0-9B-MXFP4_Q8_0_F16-Imatrix.gguf" \
  Q8_0

Generate Dual-Quant MXFP4 + Q8_0:

llama-quantize \
  --tensor-type ".*_shexp\.weight=Q8_0" \
  --tensor-type "blk\..*\.(ffn_down_exps|ffn_gate_exps|ffn_up_exps)\.weight=MXFP4" \
  --imatrix "imatrix.gguf" \
  "Ornith-1.0-9B_F16.gguf" \
  "Ornith-1.0-9B-MXFP4_Q8_0-Imatrix.gguf" \
  Q8_0

Generate Single-Quant MXFP4:

llama-quantize \
  --tensor-type ".*_shexp\.weight=MXFP4" \
  --tensor-type "token_embd\.weight=MXFP4" \
  --tensor-type "^output\.weight=MXFP4" \
  --tensor-type "blk\..*\.(ssm_alpha|ssm_beta|ssm_out|attn_gate|attn_qkv|ffn_down|ffn_gate|ffn_up|attn_k|attn_q|attn_v|attn_output)\.weight=MXFP4" \
  --tensor-type "blk\..*\.(ffn_down_exps|ffn_gate_exps|ffn_up_exps)\.weight=MXFP4" \
  --imatrix "imatrix.gguf" \
  "Ornith-1.0-9B_F16.gguf" \
  "Ornith-1.0-9B-MXFP4-Only-Imatrix.gguf" \
  MXFP4_MOE

---

📝 Local Deployment & llama-server Configuration (config.ini)

To maintain the rock-solid reasoning loop and prevent agents from falling into repetitive tool-calling deadlocks, use the following server parameter recommendations (Similar to other Qwen 3.5+ configurations).

# --- Samplers (Dynamic & Expressive) ---
temperature = 0.55
top-k = 15
top-p = 0.90
min-p = 0.15
# --- Penalties (Prevent Syntax & Reasoner Corruption) ---
repeat-penalty = 1.08
presence-penalty = 0.00
# --- DRY Sampler (Protects Indentation & Structural Boilerplate) ---
dry-multiplier = 0.8
dry-base = 1.75
dry-allowed-length = 5
dry-penalty-last-n = 1024
dry-sequence-breaker = ["\n", ":", " ", "\t", "\"", ","]
# --- Enforced Execution Graph ---
samplers = min_p;top_k;top_p;temp;dry

Highly Recommended: Always keep reasoning/thinking enabled, for better quality results.

# --- Reasoning ---
chat-template-kwargs = { "enable_thinking":true }
reasoning = on
reasoning-budget = 4096
reasoning-format = auto

This works well with 256k context window.

[!TIP] For Maximum Quality at 100k+ Context:
Use the MXFP4_MOE + Q8_0 + F16 split-quantized version.

• Preserved at F16: token_embd.weight, output.weight, *.ssm_alpha.weight, and *.ssm_beta.weight.
• Why this matters: Keeping these critical layers at full precision prevents the model from dropping fine details during extreme "needle-in-a-haystack" retrieval tasks (large context windows).
• What to avoid: If output.weight or the embedding layers are quantized to Q8_0 or lower, logit precision rounds off, causing the model to lose accuracy and forget specific details in long-context scenarios.

---

ℹ️ Misc Details

I'm doing this as a side hobby, with my AMD 5900X, 64GB DDR4, RTX 3060 12GB & RTX 5060 Ti 16GB.

In addition to the above configuration, I also use:

slots = 1
parallel = 1
no-warmup = true

flash-attn = on
mlock = false
no-mmap = false
context-shift = false

batch-size = 2048
ubatch-size = 256

fit = on
fit-target = 768
main-gpu = 0
split-mode = layer
n-gpu-layers = 999
n-cpu-moe = 0
tensor-split = 16,12
override-tensor = (token_embd)=CUDA0,(vision|vpm|nextn)=CPU

cache-type-k = q8_0
cache-type-v = q8_0

jinja = true
chat-template = jinja
chat-template-file = chat_template.jinja

For further quality and better ssm behaviour, this configuration can help:

context-shift = false
cache-type-k = f16
cache-type-v = f16

---

🤝 Support the Journey

As a passionate developer, I'm always programming, automating, or experimenting with new ideas.
I love building open-source tools, trying out new web tech, and creating things that don't yet exist, including local AI & quantizing models.

I love sharing these creations to give back to the community.
If my projects have saved you time or helped you out, consider supporting my work below!

👉 Support me on Ko-fi

---

✨ Acknowledgments

• ornith-ai for the exceptional Ornith-1.0 base model.

📜 License

Released under MIT.

🔗 Citation

@misc{ornith_9b,
    title = {{Ornith-1.0-9B}: Agentic Coding, Open to All},
    url = {https://deep-reinforce.com/ornith_1_0.html},
    author = {{DeepReinforce Team}},
    year = {2026}
}