ilyakam/Geer-Sage-A1.5-35B-A3B-6bit-MLX

Geer Sage-A1.5 35B-A3B (6-bit MLX)

This is Geer's independent 6-bit MLX conversion of Sage-A1.5 from Shanghai AI Laboratory. It retains the model's native vision tower and processor assets.

Build and changes

• Source: InternScience/Sage-A1.5
• Immutable source revision: 0e3205cf441cf01df0c0d238cc4738d0e43b4ca7
• BF16 source payload: 71,931,809,355 bytes, including the source MTP file
• Architecture: qwen3_5_moe
• Conversion: qwen3_5_moe_unfused_experts in Geer's build tool
• Conversion packages: MLX 0.32.0, MLX-LM 0.31.3, MLX-VLM 0.6.3, huggingface-hub 1.24.0, hf-xet 1.5.2, safetensors 0.8.0
• Converted size and file hashes: recorded by geer-build-manifest.json

Affine 6-bit, group size 64, for eligible language-model matrices. The vision tower, norms, and other non-quantized parameters remain BF16. The transformation fuses the source's individually stored routed experts for MLX-VLM. Its pinned sanitizer omits the source mtp.* tensors. This artifact therefore does not supply native MTP or claim speculative-decoding support.

Geer changes the weight layout and precision, rewrites conversion metadata, and supplies this card, licenses, and attribution notices. MLX-VLM may also normalize tokenizer and processor metadata. NOTICE identifies these changes. The source and build manifests record the exact original and output files, including the tokenizer, chat template, configurations, and tensors.

The artifact includes the pinned geer-recipe.toml, complete geer-source-manifest.json, and geer-build-manifest.json. The completed conversion passed full file-size and SHA-256 verification. Reproduce it with Geer's download, convert, and verify commands using the included recipe.

Evaluation and use

Upstream benchmark reports describe the source model. They are reported quality proxies, not measured results for this quantization. Geer does not claim that conversion preserves any benchmark score.

Image inputs are native to the model. Full serving and computer-use task qualification remains incomplete. Retaining vision tensors alone does not verify image serving or application compatibility. Compatibility checks use Geer-pinned oMLX 0.5.3, revision 31d07a7fdbbb499d4520188fb7e8f7bdbfe4c770; results apply only to the tested artifact and application path.

The source configuration's 262,144-token context limit is not a qualified useful-context or latency result. This artifact does not establish support for a 16 GB Mac or another untested chip and memory configuration.

License and provenance

The pinned source card declares apache-2.0 and describes Qwen3.6-35B-A3B and Agents-A1 as its bases. The pinned Sage snapshot has no separate LICENSE or NOTICE file. LICENSE supplies the standard Apache License 2.0 text. LICENSE-QWEN preserves the Qwen license and attribution. NOTICE records the declared lineage, evidence, and Geer's modifications. Geer's MIT project license does not replace the model's Apache license.

The evidence below identifies source files independently of local paths. The full source manifest supplies the remaining hashes.

Source file	SHA-256
README.md	6018232fad99dd25660741525f2a76d8258dfef6620e930a33e4bbf4a9d0c914
config.json	93a4693fa9d8392fbfccd4b3c9873f4bfdcb14fdede978b123d07d19675efe99
model.safetensors.index.json	c49171ab39575c4b9b1accc2d8525aaba82626f2f1374bd0ed5ae86458d1c6ba
chat_template.jinja	0080393e13df7322bed792a78e75daf8a0374a80928996e5cf3cfd2ab4c1e1be
tokenizer.json	87a7830d63fcf43bf241c3c5242e96e62dd3fdc29224ca26fed8ea333db72de4
tokenizer_config.json	d403d02cf04f618863360c5f6d9043683d54544bab6da2d2e365fbcb089f30ce

The source describes training lineage; it does not identify immutable parent weight revisions. The pinned parent evidence in NOTICE documents the license review, not a claim that those parent snapshots reproduce Sage's training. Geer does not assert rights to upstream trademarks or endorsement by Shanghai AI Laboratory, the Agents-A1 contributors, or Alibaba Cloud.