Instructions to use julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo
# Run inference directly in the terminal:
llama cli -hf julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo
# Run inference directly in the terminal:
llama cli -hf julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo
# Run inference directly in the terminal:
./llama-cli -hf julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo
# Run inference directly in the terminal:
./build/bin/llama-cli -hf julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo

Use Docker

docker model run hf.co/julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo

LM Studio
Jan

vLLM

How to use julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo

Ollama
How to use julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo with Ollama:
```
ollama run hf.co/julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo
```

Unsloth Studio

How to use julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo to start chatting

How to use julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo

Run Hermes

hermes

Atomic Chat new

OpenClaw new

How to use julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Docker Model Runner
How to use julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo with Docker Model Runner:
```
docker model run hf.co/julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo
```

Lemonade

How to use julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo

Run and chat with the model

lemonade run user.Ornith-1.0-35B-ROCmFPX-StrixHalo-{{QUANT_TAG}}

List all available models

lemonade list

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

This repository contains community-generated, ROCmFPX-optimized GGUF files for deepreinforce-ai/Ornith-1.0-35B.

These files are specifically compiled and quantized to run natively on AMD Strix Halo APUs (gfx1151) utilizing the experimental ROCmFPX tensor types to maximize inference speed and context length on systems with large unified memory (32GB - 128GB).

🛠️ Hardware Compatibility & Measured RAM Usage

Strix Halo utilizes a unified memory architecture (UMA). Your physical RAM is shared between the system OS and the APU's VRAM.

Because this is a 35B parameter model with native 262K context support, memory is your primary constraint. Below are real-world, measured peak unified memory footprints at different context lengths (assuming q8_0 KV cache):

Context Length	Speed (Q4, 18.3 GiB)	Quality (Q6, 30.0 GiB)	Speed Throughput (t/s)	Recommendation
512 tokens	~20 GiB	~32 GiB	1202.75 pp / 67.43 tg	Trivial, extremely fast.
4K tokens	~21 GiB	~33 GiB	1153.38 pp	Comfortable on 32GB (Speed) or 64GB (Quality).
32K tokens	~36.6 GiB (measured)	~59.8 GiB (measured)	845.44 pp / 600.43 pp (Q)	Speed: 48GB+. Quality: 64GB+.
128K tokens	~50 GiB (est.)	~82 GiB (est.)	447.89 pp	Speed: 64GB+. Quality: 96GB+.
262K tokens (Max)	~85 GiB (est. sustained)	~106 GiB (est.)	217.96 pp	Speed: 128GB. Quality: 128GB.

System RAM Compatibility Summary:

32 GB Systems: Run the Speed Candidate at contexts up to ~8K comfortably, or up to 32K with strict swap management. The Quality Candidate (30 GiB) is very tight — 4K max.

64 GB Systems: The sweet spot. Easily handles the full 128K context for the Speed Candidate, and up to 32K for the Quality Candidate.

96 GB Systems: Very comfortable. Full 128K context on Speed, up to 64K on Quality.

128 GB Systems: The ultimate Strix Halo deployment. Runs absolutely everything at max 262K context without compromise.

Note on Memory: The 32K peak RSS was rigorously measured via /usr/bin/time -v: Speed = 36.6 GiB, Quality = 59.8 GiB. The 262K prompt-fill benchmark consumed ~30.48 GiB during the prompt-processing sweep (because KV is built incrementally), but sustained generation with a full 262K KV cache resident will require significantly more. The 85 GiB estimate for 262K sustained use is extrapolated from the measured 32K baseline.

⚠️ BIOS Warning for Strix Halo: By default, your BIOS may restrict the OS-visible memory or the GPU's UMA allocation. To reach 128K or 262K contexts, you must ensure your BIOS/AGESA settings are configured to expose your full physical RAM to the unified pool.

⚠️ Critical Requirement: Custom Runtime

Because these models use AMD-specific ROCmFP4 and ROCmFPX quantizations, they will not load in standard Ollama, LM Studio, or mainline llama.cpp.

You must run them using the specific charlie12345/ROCmFPX fork of llama.cpp. We have provided an install.sh and run.sh script in this repository to automate building this runtime for you.

The Models

We offer two variants depending on your workflow needs:

File	Preset	Size	Target Use Case
`Ornith-1.0-35B-ROCmFPX-Speed-StrixHalo.gguf`	`Q4_0_ROCMFP4_COHERENT`	18.35 GiB	Recommended. Max speed (~67 t/s) and highest possible context scaling.
`Ornith-1.0-35B-ROCmFPX-Quality-StrixHalo.gguf`	`Q6_0_ROCMFPX_AGENT`	30.04 GiB	Maximum fidelity for strict coding/agent logic at the cost of context size and speed (~49 t/s).

(Both presets utilize specific "Agent" routing which protects token embeddings and attention tensors to maintain JSON and tool-calling structures, even at lower bitrates).

🚀 Quick Start Guide

1. Clone this repository

git clone https://huggingface.co/julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo
cd Ornith-1.0-35B-ROCmFPX-StrixHalo

2. Install the Runtime This compiles the custom ROCmFPX toolchain for the gfx1151 architecture. You must have ROCm 7.2.x installed.

chmod +x install.sh run.sh
./install.sh

3. Run the Model The script automatically handles the HSA_OVERRIDE and UMA environment variables. Choose your variant (speed or quality) and mode (cli or server):

# Interactive CLI — Speed model (default)
./run.sh cli speed

# Interactive CLI — Quality model
./run.sh cli quality

# OpenAI-compatible API server on port 8080
./run.sh server speed

Custom context size (via env var):

CTX_SIZE=65536 ./run.sh cli speed

Pass extra llama.cpp args (after --):

./run.sh cli speed -- --temp 0.7 --top-p 0.8

💡 Usage Tips

Context size: The default context is 8192 (CLI) or 32768 (server). Override with CTX_SIZE. Match it to your RAM — see the table above. Setting it higher than needed wastes memory; setting it too low truncates conversation history.
KV cache: Both scripts use q8_0 KV cache by default, which halves KV memory vs f16 with negligible quality loss. For maximum context on constrained RAM, try -ctk q4_0 -ctv q4_0 via the extra-args passthrough.
Sampling: For agentic/tool-calling tasks, use --temp 0.7 --top-p 0.8. For creative writing, --temp 0.9 --top-p 0.95. Pass these after --.
FlashAttention: Enabled by default (-fa on). Required for long context; disabling it will OOM at 32K+.
Which model? Start with Speed. Switch to Quality only if you notice degradation in JSON/tool-call structure or code correctness.

About the Quantizations

These are not standard llama.cpp quants. They use AMD-specific ROCmFPX tensor types that pack weights into formats native to the gfx1151 matrix multipliers:

Q4_0_ROCMFP4_COHERENT (Speed): 4-bit ROCm FP4 with "Coherent" routing — keeps attention and embedding tensors at higher precision while aggressively quantizing FFN/MoE experts. Best speed-to-quality ratio.
Q6_0_ROCMFPX_AGENT (Quality): 6-bit ROCm FPX with "Agent" routing — preserves all tensors critical for structured output (JSON, tool calls, code) at near-full precision. Use when fidelity matters more than speed.

Both will fail to load in upstream llama.cpp — you must use the ROCmFPX runtime built by install.sh.

Performance Benchmarks (Strix Halo / gfx1151)

Measured on Ubuntu 24.04, ROCm 7.2.3, FlashAttention Enabled.

Test	Speed (`Q4_COHERENT`)	Quality (`Q6_AGENT`)
Prompt Fill (pp512)	1,202.75 t/s	760.36 t/s
Generation (tg128)	67.43 t/s	49.35 t/s
Context Fill (pp4096)	1,153.38 t/s	737.51 t/s
Long Context Fill (pp32768)	845.44 t/s	600.43 t/s
Extended Context Fill (pp131072)	447.89 t/s	not tested
Max Context Fill (pp262144)	217.96 t/s	not tested

MTP (Multi-Token Prediction): The ROCmFPX build includes MTP support (build-strix-rocmfp4-mtp.sh), but llama-bench does not currently support the --spec-type flag needed to benchmark speculative decoding. MTP throughput gains have not been measured.

License & Attribution

Base model weights: MIT (DeepReinforce)
ROCmFPX Toolchain: MIT (charlie12345 / llama.cpp)

Downloads last month: 358

GGUF

Model size

35B params

Architecture

qwen35moe

Hardware compatibility

We're not able to determine the quantization variants.

View all variants

Model tree for julianmb/Ornith-1.0-35B-ROCmFPX-StrixHalo

Base model

deepreinforce-ai/Ornith-1.0-35B

Quantized

(171)

this model