Ornith-1.0-35B ROCmFPX (Optimized for AMD Strix Halo)

Ornith-1.0-35B-ROCmFPX is a high-performance 35B mixture-of-experts model quantized and tuned specifically for local multi-agent workloads on AMD Ryzen AI Max+ 395 / Strix Halo (Radeon 8060S) using Vulkan and ROCmFPX.

---

🐙 Code Repository & Production Vulkan Runner

Full source code, production Python launcher, systemd daemon files, and pre-compiled Strix Halo Vulkan binaries are hosted on GitHub: 👉 julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo on GitHub

---

⚠️ Important: Engine & Runtime Compatibility

Note: ROCmFPX and ROCmFP4 GGUF models use custom experimental block tensor layouts. Standard stock llama.cpp, Ollama, or LM Studio cannot load ROCmFPX GGUFs out of the box.

To run these GGUF quants:

1. Use the Bundled Launcher: Use python3 run_server.py or ./run.sh from the GitHub Repository which ships with pre-compiled ROCmFPX Vulkan binaries.
2. Build from Source: Or build llama-server from the specialized ROCmFPX Fork.

---

Model Quants Included

1. Ornith-1.0-35B-ROCmFPX-Speed-StrixHalo.gguf (Primary Production Quant)

   • Bitrate: ~4.70 bpw (Q4_0_ROCMFP4_COHERENT)
   • File Size: ~19.7 GB
   • Context Window: 262,144 tokens (256k)
   • Tool Validity: 92.5% pass rate @ 4k context ($n=40$), 100.0% @ 16k context ($n=40$), 93.8% @ 16-agent burst concurrency.
   • Performance: ~80 t/s single-stream decode, 115+ t/s aggregate decode across 16 concurrent agents.

2. Ornith-1.0-35B-ROCmFPX-Quality-StrixHalo.gguf (High-Precision Quant)

   • Bitrate: ~6.50 bpw (Q6_0_ROCMFPX)
   • File Size: ~32.2 GB
   • Context Window: 262,144 tokens (256k)
   • Tool Validity: 97.5% pass rate @ 4k context.

---

Quick Start with llama.cpp / llama-server (ROCmFPX Build)

# Start server with Vulkan backend tuned for AMD Strix Halo
llama-server \
  -m Ornith-1.0-35B-ROCmFPX-Speed-StrixHalo.gguf \
  -dev Vulkan0 \
  -ngl 99 \
  -fa on \
  -c 262144 \
  -np 8 \
  -ctk q8_0 \
  -ctv q8_0 \
  -b 4096 \
  -ub 1024 \
  -t 16 \
  --jinja \
  --mlock \
  --no-warmup \
  --no-context-shift \
  --cache-prompt \
  --cache-ram 65536 \
  --host 0.0.0.0 \
  --port 8089 \
  --api-key o4a-secret-key

OpenAI-compatible API endpoint will be live at http://127.0.0.1:8089/v1.

---

Hardware & System Requirements

• Processor: AMD Ryzen AI Max+ 395 / 390 (Strix Halo)
• GPU: Integrated AMD Radeon 8060S (gfx1151)
• Memory: 64 GB or 128 GB LPDDR5X Unified RAM
• Backend: Vulkan (Mesa RADV driver)

---

Benchmarks & Measured Performance