Ornith-1.0-35B-Vision is an experimental vision-enabled MLX derivative of deepreinforce-ai/Ornith-1.0-35B, intended for fast local inference on Apple Silicon.

It combines the affine-quantized Ornith language body from wang-yang/Ornith-1.0-35B-MTPLX with the source-BF16 Qwen3.5-compatible vision tower and processor metadata from Shiftedx/ornith-1.0-35b-mxfp4-vision-mtplx. The MTP sidecar and all MTP index entries were deliberately removed: this repository is a standard autoregressive VLM and does not require MTP-specific loader patches.

Highlights

Qwen3.5 MoE multimodal architecture: 35B total / approximately 3B active parameters
40 language layers, 256 experts, hidden size 2,048
Language body: affine 4-bit, group size 64; selected MoE gates use affine 8-bit/G64
Vision tower: 27 layers, BF16, hidden size 1,152, output size 2,048, patch size 16
333 indexed vision tensors preserved from the qualified vision donor
Native image, text, tool-use and agentic-coding workloads through MLX-VLM/oMLX
Thinking disabled in the uploader's local qualification
No mtp.safetensors; mtp_num_hidden_layers=0

Among the 35B-class MLX models evaluated locally by the uploader, this model delivered the highest measured single-request decode speed while retaining vision support. This is a same-machine observation, not a universal cross-hardware ranking.

Local performance

The following measurements were produced with oMLX on an Apple M3 Ultra with 96 GiB unified memory, running macOS 26.5.2. Prompt processing (pp) and text generation (tg) use synthetic token lengths; image encoding time is not represented by these text-only throughput cases.

Single request

Test	TTFT	TPOT	pp TPS	tg TPS	End-to-end	Throughput	Peak memory
pp1024 / tg128	464.8 ms	9.73 ms/tok	2,202.9 tok/s	103.6 tok/s	1.709 s	674.0 tok/s	19.35 GB
pp4096 / tg128	1,769.1 ms	9.83 ms/tok	2,315.3 tok/s	102.5 tok/s	3.030 s	1,394.0 tok/s	20.07 GB

Continuous batching — pp1024 / tg128 per request

Batch	Aggregate tg TPS	Speedup	Aggregate pp TPS	pp TPS/request	Mean TTFT	End-to-end
1x	103.6 tok/s	1.00x	2,202.9 tok/s	2,202.9 tok/s	464.8 ms	1.709 s
2x	192.7 tok/s	1.86x	1,755.3 tok/s	877.6 tok/s	1,166.7 ms	2.495 s
4x	288.3 tok/s	2.78x	1,761.9 tok/s	440.5 tok/s	2,196.5 ms	4.101 s

Results are local point measurements. Runtime version, OS version, thermals, cache state, prompt structure and concurrent workloads can materially affect performance.

https://omlx.ai/benchmarks?model=Ornith-1.0-35B-Vision&chip=&chip_full=&quantization=&context=&pp_min=&tg_min=&page=1

Local intelligence evaluation

All runs below used thinking off. These are sampled oMLX evaluations rather than complete official benchmark submissions.

Benchmark	Score	Correct / evaluated	Full dataset	Time
MMLU	83.0%	83 / 100	14,042	52.6 s
CMMLU	82.0%	82 / 100	11,582	27.3 s
TruthfulQA	90.0%	90 / 100	817	25.4 s
HumanEval	93.0%	93 / 100	164	124.2 s
MBPP	86.0%	86 / 100	500	71.0 s
LiveCodeBench	40.0%	12 / 30	1,055	234.1 s

The smaller samples, especially the 30-problem LiveCodeBench run, have substantial sampling uncertainty. Scores should only be compared with runs using the same harness, sample IDs, random seed, decoding settings and execution environment. These tests do not measure visual reasoning quality.

Usage

Install a recent MLX-VLM build compatible with Qwen3.5 MoE VLM checkpoints, then run:

bash

python -m mlx_vlm.generate \
  --model zwcf5200/Ornith-1.0-35B-Vision \
  --image image.jpg \
  --prompt "Describe this image." \
  --max-tokens 256

The model can also be served by a compatible oMLX release and accessed through its OpenAI-compatible Chat Completions API using image_url content parts.

Suggested no-thinking agent settings

The uploader found the following conservative profile more reliable for Hermes-style tool use:

json

{
  "enable_thinking": false,
  "temperature": 0.4,
  "top_p": 0.9,
  "top_k": 20,
  "min_p": 0.0,
  "repetition_penalty": 1.0,
  "presence_penalty": 0.0,
  "max_tokens": 4096,
  "max_tool_result_tokens": 2048
}

For agent harnesses, use bounded output/tool-turn budgets and reject or deduplicate malformed or repeated tool calls. Do not use </tool_call> itself as a stop string because removing the closing marker can prevent the server from parsing a valid call.

Qualification

Shard index contains no stale or missing files
Strict MLX-VLM lazy weight binding: pass
Vision tower and processor loading: pass
Real image-description request through oMLX: pass
MTP sidecar/index/runtime metadata: absent; MTP depth is explicitly set to 0
Local Hermes tool-use qualification with the recommended no-thinking profile: pass without the previously observed large burst of repeated calls

Provenance

Language model lineage: deepreinforce-ai/Ornith-1.0-35B (MIT)
Quantized language artifact: wang-yang/Ornith-1.0-35B-MTPLX; its MTP sidecar is not included
Vision tensors and processor metadata: Shiftedx/ornith-1.0-35b-mxfp4-vision-mtplx
Vision payload: 333 BF16 tensors, preserved without requantization
This repository is a file-level composition/format conversion; no additional weight training was performed by the uploader

Review all upstream model cards and licenses before redistribution or production use. The language lineage includes MIT-licensed Ornith weights, while the vision/Qwen components are subject to Apache-2.0 terms.

Limitations

Experimental composition; not an official DeepReinforce or Qwen release
Local throughput is specific to the uploader's machine and oMLX setup
Intelligence figures are sampled local evaluations and may not reproduce across harnesses
Visual quality has been functionally validated but not evaluated on a standardized VLM suite
Tool-call reliability depends on matching chat-template/parser behavior and agent guardrails
Upstream safety characteristics remain applicable; no additional safety alignment was trained