Skip to main content
Back

DeepReinforce's Ornith 1.0 Drops Uncensored GGUF Builds for Local Coding

A new batch of GGUF quantizations makes an uncensored, abliterated Ornith 1.0 9B coding model runnable on an 8GB laptop through llama.cpp or Ollama.

Model, Full Access
  • mradermacher released static GGUF quantizations of 3MPER0RR's abliterated Ornith 1.0 9B.
  • Q4_K_M weighs 5.7GB and runs on an 8GB laptop via llama.cpp or Ollama.
  • Base model Ornith 1.0 9B scores 69.4 on SWE-Bench Verified and 43.1 on Terminal-Bench 2.1.
  • Ornith uses RL that jointly optimizes reasoning scaffolds and solution rollouts.
  • Abliteration removes refusal direction via weight orthogonalization, no fine-tuning.
  • Full quant ladder from Q2_K (3.9GB) to f16 (18GB), MIT licensed.

Community packager mradermacher has published GGUF quantizations of 3MPER0RR’s checkpoint, a refusal-suppressed version of DeepReinforce’s Ornith 1.0 9B coding model. The files make the checkpoint directly usable with llama.cpp, Ollama, LM Studio, and other GGUF-compatible runtimes.

The Q4_K_M file occupies about 5.7GB, substantially reducing the storage and memory required for local deployment. Total memory use will be higher because the runtime also needs space for the context window, KV cache, compute buffers, and application overhead.

Three stages, one downloadable artifact

The published files sit at the end of a three-stage model chain:

  1. Base model: Ornith 1.0 9B, an agentic coding model from DeepReinforce.
  2. Edited checkpoint: 3MPER0RR’s version, which applies directional ablation to reduce refusal behavior.
  3. GGUF conversion: mradermacher’s offline quantizations for local inference runtimes.

Directional ablation estimates an activation-space vector associated with refusals, then projects selected weight matrices away from that direction. The operation changes model weights without gradient-based retraining and may affect behavior beyond refusals because internal representations are not perfectly isolated.

The conversion adds no new training or benchmark results. Reported scores for the original model therefore cannot establish the coding quality of the ablated or quantized versions.

Pick the quant by total memory

The repository provides several precision levels. Lower-bit files reduce memory use and often improve inference speed, while higher-bit files usually retain more of the source checkpoint’s output distribution.

Approximate GGUF file sizes
Quant Size Deployment trade-off
Q2_K 3.9GB Smallest option, with substantial quality loss
Q3_K_M 4.7GB Memory-first option with reduced fidelity
IQ4_XS 5.3GB Compact I-quant for recent compatible runtimes
Q4_K_M 5.7GB Balanced default for size, speed, and quality
Q5_K_M 6.6GB Higher fidelity with moderate memory growth
Q6_K 7.5GB High fidelity for systems with more headroom
Q8_0 9.6GB Large quant with limited compression loss
f16 18.0GB 16-bit reference file for comparison and evaluation

The K variants use mixed quantization blocks instead of assigning identical precision to every weight. Q4_K_M, for example, stores most weights near four-bit precision while preserving selected tensors at higher precision.

An 8GB GPU may hold Q4_K_M with a modest context window, depending on runtime settings and offloading. A computer with only 8GB of total system RAM is likely to swap or fail under load; 12GB provides limited headroom, while 16GB allows more room for the operating system and KV cache.

Two commands to start serving

Recent Ollama releases can fetch the quant directly from Hugging Face:

ollama run hf.co/mradermacher/Ornith1.0-9B-3MPER0RR-abliterated-GGUF:Q4_K_M

Recent llama.cpp builds can download the same file and expose an OpenAI-style chat endpoint:

llama-server -hf mradermacher/Ornith1.0-9B-3MPER0RR-abliterated-GGUF:Q4_K_M

The llama.cpp server normally listens on port 8080 and exposes /v1/chat/completions alongside its web interface. Runtime version, context length, GPU-layer allocation, and chat-template handling can materially change memory use and output quality.

Why Ornith differs

Ornith 1.0 includes 9B-Dense, 31B-Dense, 35B-MoE, and 397B-MoE variants. According to its model card, the family uses Qwen 3.5 and Gemma 4 backbones with reinforcement-learning post-training for agentic coding. The 9B dense release targets lower-cost, single-GPU deployment.

Its training process jointly optimizes solution rollouts and the scaffold used to produce them. A scaffold can include decomposition steps, tool calls, test execution, error inspection, and revision loops. Reinforcing that process aims to improve how the model searches for a working patch, particularly on repository-level tasks.

Scores reported for the original Ornith 1.0 9B model
Benchmark Reported score What it measures
SWE-Bench Verified 69.4 Resolution of validated issues from real GitHub repositories
Terminal-Bench 2.1 43.1 Completion of tasks through a command-line environment

Those results depend on the authors’ evaluation harness, prompts, tools, scaffold, and inference settings. They describe the original checkpoint under its reported setup, not the ablated GGUF files. Local results may also change with quant level, context size, sampling parameters, and runtime implementation.

Abliteration changes the risk profile

The edited checkpoint is designed to refuse fewer prompts, including requests involving offensive security code or other restricted material. Refusal suppression does not guarantee compliance on every prompt, and it removes a layer that may otherwise block harmful output.

The conversion repository publishes no task-level comparison between the unmodified model, the ablated checkpoint, and each quant. A controlled evaluation can separate those effects by testing three configurations:

  1. The unmodified Ornith 1.0 9B checkpoint at 16-bit precision.
  2. The ablated checkpoint at 16-bit precision.
  3. The selected GGUF quant, such as Q4_K_M.

Keeping prompts, chat templates, context limits, tool definitions, and sampling settings fixed makes the comparison more useful. Relevant measures include patch success, test pass rate, invalid edits, tool-call accuracy, latency, memory use, and refusal frequency.

Where the build fits

The release suits developers who need a local coding model in a standard GGUF toolchain, researchers measuring refusal behavior, and authorized security teams whose controlled workflows require prompts that guarded models may reject. Agent frameworks can use the llama.cpp endpoint without a custom conversion step.

An unauthenticated public endpoint is a poor deployment target because the checkpoint intentionally weakens refusal behavior. Access controls, sandboxed tools, audit logs, output validation, and strict limits on filesystem and network permissions remain necessary for agentic use.

The repository metadata declares an MIT license. Commercial deployments should also verify the upstream model terms and the separate licenses of runtimes, agent frameworks, datasets, and integrated tools.

Q4_K_M provides the most practical starting point for testing. Q3_K_M reduces memory pressure, while Q5_K_M and Q6_K trade additional memory for higher fidelity.