ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX

Ornith 1.5 35B-A3B — MTPLX / MLX

Explore the model guide · All public work

Start here

Quick start — documented MTPLX command shape

mtplx pull philipjohnbasile/ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX
mtplx start cli \
  --model philipjohnbasile/ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX \
  --profile sustained --depth 1

These commands resolve mutable repository state. The dry-run and runtime-version qualifications in Usage below apply; re-profile for your hardware and workload.

An Apple-Silicon MTPLX Forge artifact of ornith-ai/Ornith-1.5-35B-A3B, with native multi-token prediction (MTP). The language-model body uses mixed 4-bit/8-bit affine MLX quantization, while the MTP sidecar is stored in BF16. The vision weights and multimodal chat template are included.

This is an MTPLX/MLX runtime artifact, not a drop-in Transformers, vLLM, or SGLang checkpoint. See the upstream model card for the model's training, intended uses, and upstream evaluation results.

Artifact summary

An earlier Hub snapshot reported safetensors.total as 5,865,901,936 stored BF16/U32 elements from the indexed base-model and vision files; it excludes the separately named mtp.safetensors sidecar. Because the quantized weights are packed, this stored-element count is not the logical base-model parameter count; the target index metadata reports 35,107,180,016 parameters, also excluding the MTP sidecar. The rendered Hub badge showed 35B when checked on September 10, 2026; the historical packed-element count should not be presented as the current badge.

mtplx_runtime.json records selected conversion settings, the claimed source repository and revision, an artifact-fingerprint value, the MTP contract, and Forge verification rows. This audit did not reproduce the conversion or independently recompute the fingerprint. Quantization overrides and model structure are recorded in config.json.

Usage

The mtplx start cli argument form below was accepted by --dry-run using the locally installed CLI, which reports 2.8.0 (2.8.1). The mtplx pull command was not executed during this audit, although both command forms are present in the tagged MTPLX 2.9.0 source. Model loading, generation, and visual-input handling were not executed. The artifact manifest records MTPLX 2.9.0 as the forge version; runtime compatibility with other releases is unverified. These unpinned commands resolve mutable repository state.

mtplx pull philipjohnbasile/ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX

mtplx start cli \
  --model philipjohnbasile/ornith-ai-Ornith-1.5-35B-A3B-V2-MTPLX \
  --profile sustained \
  --depth 1

Hugging Face may also display generic MLX-LM examples for this repository. In the inspected MLX-LM 0.31.3 standard loader, weight discovery uses model*.safetensors, so the separately named mtp.safetensors is not loaded and MTP is not activated. No MLX-LM load or generation run was performed. Use MTPLX when MTP decoding is required.

The sampler used in the Forge verification record was:

temperature=0.6, top_p=0.95, top_k=20

The included chat template is a reasoning/tool-use template. Applications should handle thinking and tool-call output according to the MTPLX surface they use.

Forge verification

The Forge record selected D1 for this artifact:

• Publisher-supplied original card label: Apple M5 Max (not independently verified by this card repair)
• Machine-readable environment: macOS-27.0-arm64-arm-64bit-Mach-O; the JSON does not identify the chip model
• Timestamp: 2026-08-22 01:50:23 -04:00
• Artifact-recommended profile: sustained
• Sampler: temperature 0.6, top-p 0.95, top-k 20

These are one recorded, stop-terminated Forge verification row per depth on this artifact and device. They are useful for choosing the local MTP depth; they are not a statistically powered benchmark, a cross-model comparison, or an independent quality evaluation. The record does not provide the prompt, repeat count, output-token count, warmup procedure, RAM/core configuration, or thermal/power state. Each recorded row sets quality_passed: true; because the record omits the evaluated prompt/output, criterion, and exactness baseline, this audit could not interpret or independently validate that flag.

Provenance

mtplx_runtime.json records forged_locally: true from the pinned upstream BF16 checkpoint and the following selected settings:

body_bits=4
body_group_size=64
body_mode=affine
mtp_policy=keep_bf16

The runtime record contains this artifact-fingerprint value:

sha256:f7d71fe2fb6c603c1a8a301556dc340b761a92cfd96d53c15312bce7842a4b77

Audited weight/config/runtime snapshot: 5e00d3b9fb1ef9383b47f20b699975ad33a7cf1c. The manifest's embedded published_to_hf.revision=c3ac80c3486344d0fabf9ca5771fbf04871efdf6 is historical provenance, not the audited or current repository revision.

Limitations and validation scope

• Only the MTPLX/MLX path and the recorded Forge verification are documented here. Direct use with Transformers, vLLM, SGLang, or other runtimes was not validated for this converted artifact.
• The mtplx start cli argument form was accepted by --dry-run; mtplx pull, model loading, and generation were not performed during this audit.
• The conversion has not been independently benchmarked against the upstream BF16 checkpoint for task quality, perplexity, long-context quality, or multimodal quality.
• The 262,144-token context length is configuration metadata, not a completed long-context acceptance test for this conversion.
• Vision weights are present, but the Forge speed record does not exercise a visual-input workload.
• MTP acceptance and throughput depend on prompt, output length, sampling, runtime version, thermal state, and hardware. Re-profile for your workload.
• This conversion inherits the upstream model's behavioral limitations and may produce inaccurate, biased, unsafe, or otherwise undesirable output.

License and attribution

At the audited revisions, the pinned source's Hugging Face metadata declares license: mit, but neither the source nor this conversion contains a standalone LICENSE file, and the source's license_link resolves to a missing file. mtplx_runtime.json separately records published_to_hf.license_spdx: apache-2.0; this audit treats that as conflicting manifest metadata, not a license grant. Confirm all applicable upstream terms before use or redistribution.

Upstream model and research: Ornith AI

Conversion/runtime tooling: MTPLX

Conversion and upload: Philip John Basile