AtomicChat/ornith-35b-GGUF

Base model: deepreinforce-ai/Ornith-1.0-35B

Ornith 1.0 35B, self-quantized to GGUF by Atomic Chat. Built straight from DeepReinforce's original weights with a per-tensor importance matrix. Runs fully offline.

Highlights

• A self-improving open-source family for agentic coding from DeepReinforce, built for tool-calling and terminal-based coding agents.
• Sparse Mixture-of-Experts: 256 routed experts with 8 active per token plus a shared expert, across 40 layers (qwen3_5_moe).
• Post-trained on top of Gemma 4 and Qwen 3.5, the mid-size member of the Ornith 1.0 lineup.
• Strong agentic coding scores: 75.6 on SWE-bench Verified and 64.2 on Terminal-Bench 2.1 (Terminus-2).
• 262,144-token native context for long files and multi-step agent traces.
• Pure open: MIT licensed, globally accessible with no regional limits.
• Full quant ladder with an importance matrix on every quant over calibration_datav3.

These GGUFs are self-quantized from the original weights, not a repack. The importance matrix keeps low-bit quants closer to the full-precision model.

Always pass --jinja so the Ornith 1.0 35B chat template is applied. Without it the model can emit malformed turns.

Model Overview

Property	Value
Base model	deepreinforce-ai/Ornith-1.0-35B
Total parameters	~35B total (MoE; model name). Active per token not stated
Layers	40
Experts	256 routed + 1 shared, 8 active per token
Context length	262,144
Architecture	qwen3_5_moe sparse MoE, post-trained on Gemma 4 and Qwen 3.5
This repo	GGUF quants (imatrix), full ladder from the original weights

Scores are DeepReinforce's published results for the base deepreinforce-ai/Ornith-1.0-35B. These are full-precision scores; the quants here run the same model locally. Quantization preserves the large majority of this, with Q4_K_M and up sitting within a point or two of full precision.

Choosing a quant

Quant	Size	Notes
Q4_K_M	21.2 GB	Recommended default. Best balance of size, speed and quality.
UD-Q4_K_XL	21.5 GB	Dynamic. Token embeddings and output kept at Q8_0 for higher quality at a Q4 footprint.
Q5_K_M	24.7 GB	Higher quality, low loss.
Q6_K	28.5 GB	Near lossless.
Q8_0	36.9 GB	Effectively lossless, reference quality.

Pick the largest file that fits your (V)RAM with room for context. Q4_K_M or UD-Q4_K_XL is the sweet spot for most setups; Q6_K or Q8_0 for maximum fidelity. As an MoE the routed experts dominate the file size, so the quants are smaller than a dense 35B.

Get started

Run Ornith 1.0 35B locally with:

• Atomic Chat: the easiest path. Open the app, search AtomicChat/ornith-35b-GGUF, pick a quant, hit Use this model.
• llama.cpp: llama-server -hf AtomicChat/ornith-35b-GGUF:Q4_K_M --jinja -c 8192
• Ollama: ollama run hf.co/AtomicChat/ornith-35b-GGUF:Q4_K_M
• LM Studio / Jan: search the repo id, download any quant.

Best practices

Parameter	Value
temperature	0.6
top_p	0.95
top_k	20

DeepReinforce's recommended sampling parameters. The card notes that temperature=1.0 reproduces the reported benchmark setup.

Run in llama.cpp

git clone https://github.com/ggerganov/llama.cpp
cmake llama.cpp -B llama.cpp/build -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON
cmake --build llama.cpp/build --config Release -j --target llama-cli llama-server

./llama.cpp/build/bin/llama-server \
    -hf AtomicChat/ornith-35b-GGUF:UD-Q4_K_XL \
    --jinja -ngl 99 -c 8192 -fa on

How these were made

1. Download deepreinforce-ai/Ornith-1.0-35B (original weights).
2. Convert to f16 GGUF with llama.cpp; the NextN/MTP block is stripped before quantizing.
3. Build an importance matrix over calibration_datav3 with llama-imatrix.
4. Quantize the full ladder with --imatrix.
5. UD-Q4_K_XL additionally pins the token-embedding and output tensors to Q8_0.

License

Released by DeepReinforce under the MIT license, globally accessible with no regional limits. Quantized by Atomic Chat.