KAT-Ornith-Coder-35B-A3B 🦅⚡

💡 Overview & Philosophy

KAT-Ornith-Coder-35B-A3B is a specialized, post-trained hybrid Mixture-of-Experts (MoE) model built specifically for autonomous agentic software engineering (e.g., Claude Code, OpenHands, Aider, and Cline).

In agentic coding workflows, models consistently fail in one of two distinct regimes:

1. The Planning Deficit: Models with terse, direct execution often take action too quickly, failing to explore unfamiliar repositories, misdiagnosing complex bugs, or getting stuck on multi-file architectural refactors.
2. The Execution Deficit: Deep-reasoning models often produce verbose chains of thought, drift into hallucinated file states, overwrite entire 2,000-line files when fixing a one-line bug, or corrupt tool-call XML syntax in long multi-turn sessions.

KAT-Ornith-Coder-35B-A3B resolves this tradeoff through an Asymmetric Task-Vector Merge:

               ┌─────────────────────────────────────────────────────────┐
               │              KAT-Ornith-Coder-35B-A3B              │
               └────────────────────────────┬────────────────────────────┘
                                          │
                     ┌──────────────────────┴──────────────────────┐
                     ▼                                         ▼
        [ THE BRAIN: Ornith-1.5 ]                     [ THE HANDS: KAT-Coder ]
  • Autonomous RL self-improvement              • 100k KwaiClawEnv sandbox PPO
  • 79.0% SWE-bench Verified planning           • 0.28% abnormal tool label rate
  • Dominates DeltaNet Linear Attention         • Anchors Embeddings, LM Head & Routers
  • Deep hypothesis & root-cause analysis       • Surgical diff & patch generation

• The Brain (ornith-ai/Ornith-1.5-35B-A3B): Injected into the Gated DeltaNet linear attention and self-attention projections to supply deep multi-step algorithmic planning, high exploration rollouts, and SWE-bench Verified reasoning ceiling.
• The Hands (Kwaipilot/KAT-Coder-V2.5-Dev): Anchors the token embeddings, LM head, and MoE routing gates to preserve its ultra-low tool syntax error rate (0.28%), clean localized diffs, and zero-loop execution discipline.

🧠 Underlying Architecture

Both parent models share an identical, next-generation sparse MoE backbone with linear attention hybrid layers:

• Total Parameters: ~35 Billion
• Active Parameters per Token: ~3 Billion (8 routed experts + 1 shared expert)
• Layer Depth: 40 Layers
• Hybrid Attention Topology: 30 Gated DeltaNet linear attention layers interleaved with 10 full attention layers (every 4th layer)
• Expert Count: 256 fine-grained routed experts + 1 shared expert
• Context Window: Native 262,144 tokens (262k)

🔬 How It Was Created: Detailed Merge Recipe

The merge was synthesized using Asymmetric DARE-TIES (Drop And REscale with Task Information Sign-consensus). Instead of a uniform 50/50 blend (which corrupts MoE routing boundaries), the model deltas were applied differentially across functional sub-networks:

1. Functional Parameter Mapping

$$\Delta_W = W_{\text{Ornith}} - W_{\text{KAT}}$$

$$\hat{\Delta}_W = \frac{\Delta_W \odot \text{Bernoulli}(\text{density})}{\text{density}}$$

$$W_{\text{merged}} = W_{\text{KAT}} + (\text{weight} \times \hat{\Delta}_W)$$

2. Zero-Cache Shard-by-Shard Synthesis Engine

Due to the sheer size of the 35B MoE architecture (~70 GB in BF16), standard merging utilities often fail or run out of disk space. This model was synthesized using a specialized streaming merge pipeline:

• Direct HTTPS Chunk Streaming: Streamed directly into temporary buffers without creating intermediate uncompressed blobs.
• In-Memory Tensor Transformation: Transformed each tensor in RAM, explicitly cloning untransformed tensors to release memory-mapped (mmap) Linux file descriptors.
• Single-Shard Commit Pipeline: Each of the 13 shards (model-00000-of-00013.safetensors through model-00012-of-00013.safetensors) was created, validated, uploaded to Hugging Face, and purged locally to ensure zero numeric truncation.

📊 Performance Profile & Benchmark Context

The parents of this merge hold top-tier standing in the 35B-A3B open-weights class:

🚀 Deployment & Serving

1. High-Throughput Production Serving with vLLM

For local or server deployment with full XML tool-calling and hybrid attention acceleration:

vllm serve OliviaRossi/KAT-Ornith-Coder-35B-A3B \
    --tensor-parallel-size 2 \
    --max-model-len 131072 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_xml \
    --reasoning-parser qwen3 \
    --dtype bfloat16

(For the full 262k context window, allocate --tensor-parallel-size 4 on 4x A100/H100/A6000 Ada GPUs).

2. Local Inference via llama.cpp / GGUF

To run locally in quantized GGUF format:

# Convert to GGUF (from repository root)
python3 convert_hf_to_gguf.py /path/to/KAT-Ornith-Coder-35B-A3B \
    --outfile KAT-Ornith-Coder-35B-A3B-bf16.gguf \
    --outtype bf16

# Quantize to Q4_K_M (fits on ~24GB VRAM GPU or Apple Silicon unified memory)
llama-quantize KAT-Ornith-Coder-35B-A3B-bf16.gguf KAT-Ornith-Coder-35B-A3B-Q4_K_M.gguf Q4_K_M

# Run server
llama-server \
    -m KAT-Ornith-Coder-35B-A3B-Q4_K_M.gguf \
    -c 32768 \
    --n-gpu-layers 99 \
    --chat-template chatml

⚙️ Recommended Generation & Agent Settings

To achieve the best balance of reasoning and deterministic tool formatting in agent harnesses:

Compatible Agent Harnesses

• Claude Code: Configure via OpenAI-compatible endpoint with --model-name OliviaRossi/KAT-Ornith-Coder-35B-A3B.
• Aider: Run with --model openai/OliviaRossi/KAT-Ornith-Coder-35B-A3B --edit-format diff.
• OpenHands / Cline / Roo Code: Select the standard XML or OpenAI tool call parser with system prompt anchoring.

📋 Citation & Attribution

This model is a merge of two foundational research projects. If you use this model in your research or applications, please cite the parent teams:

@misc{kat_coder_2025,
  title={KAT-Coder-V2.5: Real-World Software Engineering via Multi-Teacher Distillation and Sandbox RL},
  author={KwaiPilot AI4SE Team},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/Kwaipilot/KAT-Coder-V2.5-Dev}}
}

@misc{ornith_coder_2025,
  title={Ornith-1.5-35B-A3B: Autonomous Self-Improving Reinforcement Learning for Code Intelligence},
  author={DeepReinforce / Ornith AI},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B}}
}

@misc{kat_ornith_coder_2026,
  title={KAT-Ornith-Coder-35B-A3B: Asymmetric DARE-TIES Synthesis for Agentic Coding},
  author={Olivia Rossi},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/OliviaRossi/KAT-Ornith-Coder-35B-A3B}}
}