This repository contains Geer's independently produced mixed 4/8-bit MLX conversion of Deep Reinforce AI's Ornith-1.0-35B. It is intended for local agentic coding on Apple Silicon through Geer.

The model identity is intentionally transparent: this is Ornith-1.0-35B, converted by the Geer project. It is not a new foundation model.

Build

Upstream model: deepreinforce-ai/Ornith-1.0-35B
Upstream revision: 5df2ed3f675c7beaa490328cc70bb573b65fb660
Architecture: qwen3_5_moe
Quantization: affine mixed 4/8-bit, group size 64
Four-bit tensors: routed expert gate, up, and down projections
Eight-bit tensors: other eligible language-model matrices, including embeddings, attention and Gated DeltaNet projections, routers, shared experts, and the language-model head
BF16 tensors: vision tower, norms, and other non-quantized parameters
Converted payload: 21,643,961,289 bytes (approximately 20.2 GiB)
Conversion runtime: MLX 0.32.0, MLX-LM 0.31.3, MLX-VLM 0.6.3
Hugging Face tooling: huggingface-hub 1.24.0, hf-xet 1.5.2
Safetensors: 0.8.0

The repository includes:

geer-recipe.toml, the complete pinned conversion recipe;
geer-source-manifest.json, hashes for the downloaded BF16 source;
geer-build-manifest.json, hashes for every converted output; and
the model, tokenizer, chat template, processor configuration, licenses, and attribution notices required to use the converted artifact.

The reproducible conversion tool and tensor-layout transformation are available in the Geer source repository.

Use with Geer

Geer downloads this repository at an immutable revision, verifies every file against geer-build-manifest.json, and activates the verified Hugging Face snapshot without copying the model into another directory. Geer 0.1.0 selects this conversion on 32 GB and 48 GB Macs, with 64K and 128K context windows respectively and BF16 KV cache.

Evaluation status

Deep Reinforce AI publishes results for the upstream BF16 Ornith model in its model card. Geer has not yet reproduced the upstream benchmark suite for this mixed quantization, so upstream scores should not be treated as measured results for this conversion. Although the upstream architecture and repository metadata support image-and-text inputs, Geer 0.1.0 has validated this conversion only for text-based agentic coding workflows.

License and attribution

Deep Reinforce AI declares Ornith-1.0-35B under the MIT license. Its model card states that Ornith-1.0-35B was post-trained on Qwen 3.5; the applicable Qwen Apache License 2.0 text and attribution are preserved here. See LICENSE, LICENSE-QWEN, and NOTICE before using or redistributing the model.

Geer is an independent project and is not affiliated with or endorsed by Deep Reinforce AI or Alibaba Cloud.

ilyakam/Geer-Ornith-1.0-35B-A3B-4-8bit-MLX

Geer Ornith 1.0 35B-A3B (4/8-bit MLX)

Build

Use with Geer

Evaluation status

License and attribution

ilyakam/
Geer-Ornith-1.0-35B-A3B-4-8bit-MLX