AWQ W4A16 quantization of ornith-ai/Ornith-1.5-35B-A3B, in compressed-tensors pack-quantized format, produced for SGLang on Intel Arc Pro B-series.

The multi token prediction head is preserved at BF16 so it remains available for SGLang speculative decoding.


Size	{{ARTIFACT_SIZE_GIB}} GiB
Format	`compressed-tensors` / `pack-quantized`
Scheme	W4A16 asymmetric, group size 32
Quantized	routed MoE experts only
Vision tower	preserved, BF16
MTP head	preserved, BF16
Context	262,144 tokens (native configuration)

The MTP head

Ornith-1.5-35B-A3B ships a multi token prediction head, declared in its config as mtp_num_hidden_layers: 1. On this model the head is not small: it carries its own full 256-expert mixture of experts, 785 tensors and about 822 M parameters, 2.3 percent of the whole model.

transformers does not instantiate an MTP module for Qwen3_5MoeForConditionalGeneration. The subtree is therefore absent from the model object during quantization, and save_pretrained cannot write what is not there. The head is dropped, silently, and nothing in the process errors. This is why a quantization workflow can silently produce an artifact with zero mtp.* tensors.

Here the head is carried across from the source checkpoint byte for byte into model-mtp.safetensors. It was never loaded and never quantized, so it is bit-identical BF16.

Without it, speculative decoding has no weights to load. Nothing errors; the accept length simply pins at 1.00 and the model looks like a weak drafter rather than a broken one.

What is quantized, and what is not

Ornith-1.5-35B-A3B is a fine-grained mixture of experts: 256 experts per layer, 8 active per token, across 40 layers, with moe_intermediate_size 512. The routed experts are where the parameters live.

Quantized: the routed experts, and only those. 40 layers x 256 experts x 3 projections = 30,720 modules.

Left at BF16:

	params	share
linear attention (`linear_attn.*`, 30 layers)	1.007 B	2.81%
MTP head	0.822 B	2.30%
`embed_tokens`	0.509 B	1.42%
`lm_head`	0.509 B	1.42%
vision tower (27 blocks)	0.411 B	1.15%
full attention (`self_attn.*`, 10 layers), shared experts, router gates	0.336 B	0.94%

How much of this is actually 4 bit


stored at 4 bits	32.21 B parameters, 90.0 percent of the model
kept at BF16	3.6 B parameters, 10.0 percent

Activations are BF16 throughout; only weights are quantized, hence W4A16.

Footprint and throughput are not comparable across quantization formats. A build that packs more of the model into fewer bits will be smaller and will usually decode faster on the same hardware, because decode at batch one is bound by how many bytes move per token. Which model is better for your use is answered by measuring quality on your workload, not by comparing file sizes.

Quantization details

Source: ornith-ai/Ornith-1.5-35B-A3B BF16, not re-quantized from the FP8 or NVFP4 releases.
Method: AWQ via llm-compressor, W4A16 asymmetric, group size 32, MSE observer, int8 zero points, pack-quantized.
Calibration: 512 sequences of 1024 tokens, 60 percent code instructions from codeparrot/self-instruct-starcoder and 40 percent general instructions from HuggingFaceH4/ultrachat_200k, chat-template formatted.
Pipeline: sequential, one decoder layer at a time (sequential_targets=["Qwen3_5MoeDecoderLayer"]), so each layer is calibrated against the quantized output of the layers before it.

Group size 32 rather than 128, and this is the opposite choice from a dense model. A group is a run of weights along the input dimension sharing one scale and one zero point, so what matters is how many groups each row gets. These experts are narrow: moe_intermediate_size is 512, so at group 128 an expert's down_proj would carry four scales for the entire row and a single outlier would set the quantization step for 128 weights. A dense model's input dimensions are 8x to 34x larger and never hit this. Group 32 costs about 11 percent more bytes read per token and buys back the resolution.

Calibration size matters more here than for a dense model. With 256 experts and top-8 routing, each expert observes roughly one thirty-second of the calibration tokens, so the sample count sets how well the tail experts are seen, not merely how stable the scales are.

Quality

This release passed a deterministic coherence gate, not a benchmark, while served through SGLang on four Intel Arc Pro B70 GPUs. The gate verifies that:

reasoning_content is non-empty and the separate final answer is correct on multi-step arithmetic and syllogistic reasoning;
factual and Python coding responses are coherent and contain their required result;
a basic image input is accepted and its dominant color is identified;
output is printable and does not collapse into repetitive text.

The full prompts, reasoning, final responses, token usage, and pass flags ship in coherence-results.json. This is a release smoke test; it does not estimate benchmark accuracy or claim parity with BF16.

Usage

SGLang on Intel Arc

docker run --rm -d --name ornith15 \
  --device=/dev/dri -v /dev/dri:/dev/dri \
  --group-add video --group-add "$(getent group render | cut -d: -f3)" \
  --cap-add=SYS_PTRACE --security-opt seccomp=unconfined \
  --ipc=host --shm-size=64g --ulimit memlock=-1 \
  -p 30000:30000 -v /path/to/Ornith-1.5-35B-A3B-AWQ-INT4:/model:ro \
  -e ONEAPI_DEVICE_SELECTOR=level_zero:gpu \
  rahulunair/sglang-xpu:latest \
  python -m sglang.launch_server --model-path /model \
    --device xpu --tp-size 4 --moe-runner-backend triton \
    --host 0.0.0.0 --port 30000 \
    --trust-remote-code --attention-backend intel_xpu --page-size 64 \
    --context-length 40960 --max-total-tokens 40960 \
    --max-mamba-cache-size 64 \
    --chunked-prefill-size 4096 --mem-fraction-static 0.85 \
    --reasoning-parser qwen3-thinking \
    --cuda-graph-config '{"decode":{"backend":"full","bs":[1,2,4,8]},"prefill":{"backend":"disabled"}}' \
    --skip-server-warmup

SYS_PTRACE and seccomp=unconfined are required. Without them the symmetric-memory all-reduce declines silently and the slower stock collective runs, with nothing in the log to say so.

Check that max_total_num_tokens equals context_len in the startup log. If it is smaller, the Gated DeltaNet state pool has taken the KV cache and long prompts will fail; --max-mamba-cache-size is what corrects it.

This is a thinking model. Pass --reasoning-parser qwen3-thinking so the trace is returned in reasoning_content and the answer in content.

Recommended sampling, from the base model card: temperature=0.6, top_p=0.95, top_k=20 for general use, temperature=1.0 to reproduce its published benchmark numbers.

Speculative decoding with the MTP head

The head is BF16 and excluded from quantization, so it loads as plain BF16 linears. It is off unless asked for and costs nothing while off.

Support for an MTP head on this architecture depends on your SGLang build registering an MoE MTP model class. Confirm before relying on it.

Limitations

Quality is not benchmarked. The release coherence gate verifies normal reasoning, answer, code, factual, and basic image behavior only.
Vision is lightly exercised. The vision tower is unquantized and the release gate includes one synthetic color image, but no multimodal benchmark was run.
The AWQ smoothing scale for each layer's expert input is shared between the routed experts and the shared expert, and the shared expert is not quantized here, so it contributes to the smoothing objective without benefiting from it. This is what upstream llm-compressor does for every Qwen mixture of experts, and what the reference artifact was produced under.
Treat this as a 4 bit model. Greedy text will diverge from BF16, as it will for any 4 bit build, so it is not a drop in replacement where exact BF16 outputs matter.

License

MIT, inherited from ornith-ai/Ornith-1.5-35B-A3B.

ulkaa/Ornith-1.5-35B-A3B-AWQ-INT4

Ornith-1.5-35B-A3B-AWQ-INT4