Ornith-1.5-35B-ROCmFP4-STRIX_LEAN

Ornith-1.5-35B — ROCmFP4-STRIX_LEAN (Strix Halo / gfx1151)

Version 1.5 — 2026-08-20

TL;DR

Ornith-1.5-35B (35B params, 3B active per token, Qwen3.5-VL-MoE family) quantized to Q4_0_ROCMFP4_STRIX_LEAN (type 106 preset, ~4.29 BPW). Tuned for AMD Strix Halo (gfx1151 / RDNA 3.5) on the ROCmFPX fork family — we serve and benchmark these files on our lab runtime (full source: pugant/strix-nebulosa; upstream: charlie12345/ROCmFPX). Runs the full vision + text multimodal model in ~17.7 GiB.

⚠️ Critical warnings — read before downloading

• Requires a ROCmFPX fork build — see Runtime. Upstream charlie12345/ROCmFPX loads these files too (any build with the custom GGML types, e.g. built via the kyuz0/amd-strix-halo-toolboxes container). The type 106 (Q4_0_ROCMFP4_STRIX_LEAN) tensor format is INVALID in stock llama.cpp — it will refuse to load. See Usage below.
• Profiled for gfx1151 only (Strix Halo / Ryzen AI Max+ 395, RDNA 3.5). Not tested on other GPUs.
• FP4 here is software on RDNA 3.5 (no FP4 silicon units): the win is bandwidth / memory, not raw compute throughput. The Strix Halo ceiling on this MoE is bandwidth-bound, which is exactly where FP4 helps.

Benchmarks

Tested on Strix Halo (AMD Ryzen AI Max+ 395, 128 GB LPDDR5X). Methodology: llama-bench -ngl 999 -fa on -p 512 -n 128, Vulkan (RADV) build of the fork, GPU-exclusive window (production service stopped). Note: the Ornith-1.0 release numbers below were measured on the ROCm (HIP) backend — different backend, not directly comparable.

Sanity check: Ornith-1.5 tracks its base-architecture sibling Qwen3.6-35B-A3B (same quant, same backend) within 1% — 82.45 vs 81.57 tg128. On this Vulkan build Ornith-1.5 decodes 23.7% faster than Ornith-1.0 on ROCm at the same footprint class.

Speculative decoding (MTP)

The source model ships a 1-layer MTP head (nextn_predict_layers=1, tensors blk.40.nextn.*), included in this GGUF and activated at runtime only if you opt in with the fork's --spec-type draft-mtp flags.

Measured on the same GPU-exclusive window, same server/prompt methodology (2 prompts × 2 runs, ctx 16k, Vulkan build of the fork):

Verdict: speculative decoding does not pay off on Ornith-1.5 — plain inference wins at every n-max (−24% prose at the best MTP setting). The measured draft acceptance explains why: position-1 acceptance is high (0.99 on deterministic tasks) but position-2 collapses to ~0.07, so the mean accepted length (1.4–1.7) never covers the draft+verify cost — the nextn layer is a full MoE layer. For comparison, on the same stack the Qwen3.6-35B-A3B base model accepts (0.87, 0.77, 0.64) and gains +37% with MTP n-max 3: the 1.5 fine-tune degraded the MTP head beyond the first drafted token. If you still want to experiment, use --spec-draft-n-max 2; above that it is pure overhead. MTP stays opt-in: with no spec flags the model runs plain inference at the headline speeds above.

System configuration at bench time

Declared for reproducibility:

• Bare metal host: Bosgame BeyondMax Series, Ubuntu 24.04.4 LTS, kernel 7.0.0-28-generic
• CPU power profile: balanced (powerprofilesctl get) — default, NOT forced to performance. Representative of an out-of-the-box setup.
• CPU scaling driver: amd-pstate-epp, scaling_governor performance (amd-pstate-epp default), EPP performance
• IOMMU / iGPU power: auto (no manual tuning)

Note: tok/s above were measured on a non-tuned system (power profile balanced). Users who set powerprofilesctl set performance may see slightly higher numbers.

Quantization details

Preset Q4_0_ROCMFP4_STRIX_LEAN (GGUF file_type 106, ~4.29 bits/weight):

• Token embeddings (token_embd.weight) → Q5_K (preserve vocab fidelity)
• Attention K/V (blk.*.attn_qkv.weight, blk.*.attn_v.weight) → q4_0_rocmfp4 (high-precision path for attention state)
• Expert FFN (blk.*.ffn_*_exps.weight) → q4_0_rocmfp4_fast (max speed path; the bulk of MoE weights)
• MTP head (blk.40.nextn.*) kept in the file (BF16/F32 as in source)
• Zero quantization fallbacks across all 753 tensors

Reference fork: charlie12345/ROCmFPX (MIT).

Serving runtime: see Runtime.

imatrix methodology

Precomputed by bartowski on 573 chunks (calibration-v6 dataset), redistributed here as imatrix-Ornith-1.5-35B-bartowski.gguf with explicit attribution. The original is at bartowski/Ornith-1.5-35B-A3B-GGUF (MIT). Unlike the Ornith-1.0 release, which used the unsloth imatrix computed on 1.0 weights.

Files

Usage

# Requires the kyuz0 Strix Halo toolbox (which builds charlie12345/ROCmFPX)
docker run --rm -p 1234:1234 --device /dev/kfd --device /dev/dri \
  -v /path/to/models:/models rocmfpx-llm-service \
  llama-server \
    -m /models/Ornith-1.5-35B-ROCmFP4-STRIX_LEAN.gguf \
    --mmproj /models/mmproj-Ornith-1.5-35B-BF16.gguf \
    -ngl 999 -fa on --jinja -c 32768 --host 0.0.0.0 --port 1234

Notes:

• MTP is opt-in at runtime. Add --spec-type draft-mtp --spec-draft-ngl all --spec-draft-p-min 0.0 --spec-draft-p-split 0.10 --spec-draft-n-max <N> to activate speculative decoding (see the MTP section above for measured n-max guidance). Plain inference (no spec flags) is the default and what the headline benchmark table reports.
• The --mmproj flag is required for the vision tower (multimodal). Without it, text-only still works.

How to replicate

1. Build the docker-llm-service-convert image from kyuz0/amd-strix-halo-toolboxes + charlie12345/ROCmFPX (must contain MODEL_ARCH.QWEN35MOE).
2. Download the BF16 GGUF from ornith-ai/Ornith-1.5-35B-A3B-GGUF.
3. Quantize with the bartowski imatrix: llama-quantize --imatrix imatrix-Ornith-1.5-35B-bartowski.gguf <bf16>.gguf <out>.gguf Q4_0_ROCMFP4_STRIX_LEAN 16.

Attribution & model tree

Qwen3.5-VL-MoE (base architecture)
    └── ornith-ai/Ornith-1.5-35B (A3B) (MIT)
            └── this GGUF (ROCmFP4-STRIX_LEAN)

• Base model: ornith-ai/Ornith-1.5-35B (MIT)
• BF16 GGUF + mmproj source: ornith-ai/Ornith-1.5-35B-A3B-GGUF (MIT)
• imatrix: bartowski/Ornith-1.5-35B-A3B-GGUF (MIT)
• Quantization fork: charlie12345/ROCmFPX (MIT)
• Container runtime: kyuz0/amd-strix-halo-toolboxes

License

MIT (inherited from ornith-ai/Ornith-1.5-35B and its GGUF release). Derivative work: original model and its license are preserved. See LICENSE and NOTICE.

Acknowledgements

Built on the shoulders of giants:

• kyuz0/amd-strix-halo-toolboxes — Strix Halo container runtime
• charlie12345/ROCmFPX — llama.cpp fork with ROCmFP4 presets (type 106)
• bartowski — precomputed imatrix for Ornith-1.5-35B
• ornith-ai / DeepReinforce Team — Ornith-1.5-35B
• llama.cpp community + Kawrakow (imatrix methodology)
• Hardware: Bosgame BeyondMax Series (Strix Halo bare metal host)

Limitations & community feedback

• Speed benchmark only. No perplexity / MMLU / quality eval is included in this release. The MoE structure is preserved from the BF16 source except for the quantized tensor formats above; quality is expected to track standard Q4_K_M-class with the ROCmFP4 attention/K-V choices, but this is not measured here.
• Profiled for gfx1151 only. Not tested on other GPUs (feel free to share yours).
• MTP head present in weights; speculative decoding is opt-in at runtime (plain inference by default). No separate drafter head exists for Ornith-1.5 as of release date.

We invite the community — especially fellow Strix Halo owners — to test and share quality results. Open a Discussion on this repo.

Citation

@misc{ornith152026,
  title  = {Ornith-1.5-35B},
  author = {DeepReinforce Team},
  year   = {2026},
  url    = {https://huggingface.co/ornith-ai/Ornith-1.5-35B}
}

Disclaimer

No affiliation with AMD, Qwen, DeepReinforce, bartowski, unsloth, kyuz0, or charlie12345. Provided as-is, without warranty. Users must comply with the base model license (MIT).

---

Runtime

Benchmark environment: one bare-metal AMD Strix Halo (Ryzen AI MAX+ 395, 128 GB) — full dated configuration and measurement policy: BARE-METAL.md

Requires a ROCmFPX fork build (custom tensor types — stock llama.cpp refuses the file). Recommended: our lab build (pugant/strix-nebulosa, main) — reasoning budget and persistent prompt cache on every model; drafter features where the model ships one: see the engine section of its README.

Honest note: the fine-tune degraded the MTP head (pos-2 acceptance ~0.07) — run it plain, skip the drafter.

Everything here is experimental and provided as-is, at your own risk.