| --- |
| license: mit |
| language: |
| - en |
| pipeline_tag: text-generation |
| tags: |
| - fastflowlm |
| - q4nx |
| - npu |
| - qwen3.6-moe |
| - 35b-a3b |
| - ornith |
| - quantization |
| base_model: |
| - ornith-ai/Ornith-1.0-35B |
| base_model_relation: quantized |
| quantized_by: Atomic-Germ |
| --- |
| # *IF YOU USE COMMUNITY QWEN MODELS DO NOT UPGRADE TO FLM v1.0.2+* |
|
|
| # Ornith-1.0-35B - Q4NX for FastFlowLM (AMD Ryzen AI XDNA2) |
|
|
| Ornith-1.0-35B is converted to Q4NX for hardware-accelerated inference with FastFlowLM on AMD Ryzen AI NPUs. |
|
|
| ## What is Q4NX? |
|
|
| Q4NX is FastFlowLM's native packed-quantization format - a rearranged Q4_1 layout tuned for the NPU matrix engine's tile sizes and memory access patterns. It is **not** a GGUF file and it does not run on llama.cpp or Ollama; it is meant exclusively for the [FastFlowLM](https://fastflowlm.com) engine on AMD Ryzen AI NPUs. |
| |
| ## Requirements |
| |
| - FastFlowLM >= 0.9.46 (`flm` CLI) |
| - AMD Ryzen AI processor with **XDNA2 (NPU2)** - Strix Point / Ryzen AI 300 |
| series or later |
| - Linux with the XRT NPU stack installed |
| - ~47 GB of unified system memory (Q4NX weights + activations + KV cache) |
| |
| ## Files |
| |
| | File | Purpose | |
| |---|---| |
| | model.q4nx | Quantized Q4NX text weights | |
| | config.json | FastFlowLM model configuration | |
| | tokenizer.json | Tokenizer | |
| | tokenizer_config.json | Special tokens and chat template | |
| | chat_template.jinja | Chat template | |
| | README.md | | |
| |
| ## Install and run |
| |
| This repository works with `flm-add`, a small installer that copies the model |
| into the FastFlowLM user directory and registers the tag. It never |
| modifies the system FastFlowLM install. |
| |
| `pip install flm-add` or `uv tool install flm-add` |
| |
| ```bash |
| uv tool install flm-add |
| unset FLM_CONFIG_PATH FLM_XCLBIN_PATH && flm-add Atomic-Germ/Ornith-1.0-35B-A3B-NPU2 --tag ornith1.0-moe:35b-a3b --family qwen3.6-moe |
| FLM_CONFIG_PATH="$HOME/.config/flm/model_list.json" FLM_XCLBIN_PATH="$HOME/.config/flm" flm run ornith1.0-moe:35b-a3b |
|
|
| ``` |
| |
| | Context Length | TTFT (s) | Prefill Speed (tok/s) | Decoding Speed (tok/s) |
| |---|---|---|---| |
| | 1k | 12.355 ± 0.122 | 79.07 ± 0.78 | 10.41 ± 0.01 |
| | 2k | 16.849 ± 0.142 | 115.41 ± 0.97 | 10.36 ± 0.07 |
| | 4k | 24.852 ± 0.035 | 156.09 ± 0.22 | 10.14 ± 0.10 |
| | 8k | 40.241 ± 0.385 | 192.55 ± 1.84 | 9.74 ± 0.07 |
| | 16k | 73.398 ± 1.227 | 211.03 ± 3.53 | 9.02 ± 0.07 |
| | 32k | 148.780 ± 0.630 | 208.11 ± 0.89 | 7.70 ± 0.03 |
| |
| --- |
| |
| ## Kernels |
| |
| FastFlowLM's NPU kernels (xclbins) are closed source and are not shipped in this repository. This model uses the **`qwen3.6-moe`** engine family and is shape-identical to the official **`qwen3.6-moe:35b-a3b`** model (`Qwen3.6-Moe-35BA3B-NPU2`). Point the runtime's xclbin path at the matching `xclbins` directory (or ship your own) before running. |
| |
| ## Model |
| |
| - Registry tag: `ornith1.0-moe:35b-a3b` |
| - Engine family: `qwen3.6-moe` |
| - Kernel source: official fastflowlm `qwen3.6-moe:35b-a3b` |
| - Context length: 262,144 tokens (from config) |
| - Hidden size: 2048 |
| - Layers: 40 |
| - Intermediate size: 512 |
| - Vocabulary: 248320 |
| - `model.q4nx` size: 23.2 GB |
| - Base model: [ornith-ai/Ornith-1.0-35B](https://huggingface.co/ornith-ai/Ornith-1.0-35B) |
| - License: other |
| |
| ChatCompletionChunk: {"id":"chatcmpl-81cc726ee0a82b1bf2703a65","object":"chat.completion.chunk","created":1786483208,"model":"ornith-moe:35b-a3b","system_fingerprint":"fp_7076fd14a68716c5","choices":[{"index":0,"delta":{"content":null},"finish_reason":"stop"}],"usage":{"prompt_tokens":11140,"completion_tokens":330,"total_tokens":11470,"active_kv_tokens":11470,"max_kv_token_capacity":32768,"kv_token_occupancy_rate_percentage":35.003662109375,"load_duration":1.082e-06,"prefill_duration_ttft":74.225606656,"decoding_duration":33.763416,"prefill_speed_tps":150.0829767768493,"decoding_speed_tps":9.773892546891583}} |
| |
| ## Original model card |
| |
| See the upstream model card for training details, benchmarks, and upstream |
| usage. This repository only contains the Q4NX conversion for FastFlowLM. |
| - Upstream card: [ornith-ai/Ornith-1.0-35B](https://huggingface.co/ornith-ai/Ornith-1.0-35B) |
