Qwen3.6-35B-A3B-oQ2e-FP16-MTP-MLX

This is the FP16 precision sibling of treeish/Qwen3.6-35B-A3B-oQ2e-MTP-MLX. It is intended for Sprig on M1 and M2 Macs, where FP16 avoids relying on BF16 arithmetic. M3 and newer Macs should use the original BF16 package.

The oQ2e quantised weights, embedded Multi-Token Prediction head, tokenizer and corrected chat template are all retained. This sibling changes storage precision only for tensors that were BF16 in the source package. It does not reduce the model's disk or memory footprint.

Model

• Base model: Qwen/Qwen3.6-35B-A3B
• Architecture: 35B total parameters, approximately 3B active parameters
• Quantisation: oQ2e imatrix-enhanced mixed precision
• Default quantisation: 2-bit affine, group size 64
• Per-tensor overrides: 3, 4, 5, 6 and 8-bit
• Residual floating-point precision: FP16
• Format: MLX safetensors
• MTP: one embedded layer under language_model.mtp.*
• Context length: 262,144 tokens

The package contains 2,052 indexed tensors, including 42 embedded MTP tensors. It occupies essentially the same 13.6 GB as its BF16 sibling.

Conversion

The source is treeish/Qwen3.6-35B-A3B-oQ2e-MTP-MLX at commit c14177b62ae1303536f08ac9dbd0f72947dceefa.

The conversion used scripts/make_fp16_precision_sibling.py from youssofal/MTPLX at commit 963b923fa8be21f43194203777d5c075fc041ec9 with MLX 0.32.0. Its policy is:

• Cast every BF16 tensor to FP16.
• Preserve packed quantised tensors and every other non-BF16 tensor.
• Preserve safetensors metadata and copy non-weight files byte for byte.

The conversion changed 1,569 BF16 tensors to FP16 and preserved 483 non-BF16 tensors. Of the 42 embedded MTP tensors, 31 became FP16 and the 11 packed uint32 tensors were preserved. MTPLX_FP16_CONVERSION_MANIFEST.json records every tensor, dtype, shape and source/output shard digest.

The parent package's model provenance and Froggeric chat-template attribution remain applicable. See its model card and this package's RELEASE_MANIFEST.json for the pinned upstream revisions.

Runtime compatibility

This package is validated with Treeish's pinned MLX Swift runtime. Another runtime must support the per-tensor quantisation overrides in config.json and the embedded Qwen MTP layout.

This package is Treeish's M1 and M2 variant. M3 and newer Macs use the BF16 sibling.

Treeish uses this model from 24 GB of unified memory and recommends 32 GB. Actual headroom depends on context length, cache settings and other running applications.

Treeish validation

The release candidate was validated on a 36 GB M4 Max using Treeish's release benchmark:

• All 2,052 tensor names, shapes and shard assignments match the BF16 source.
• Each converted tensor exactly matches the source value cast to FP16.
• All 483 non-BF16 tensors match the source value exactly.
• All safetensors metadata matches and all 13 non-weight source files were initially copied byte for byte.
• Embedded-MTP tool use produced a parsed search_text call with the requested query and result count.
• The warmed 1,066-token fixture generated 89.2 tokens/s without MTP, 95.6 tokens/s with block size 2, 105.8 tokens/s with block size 3 and 103.3 tokens/s with block size 4.
• Block size 3 accepted 58 of 73 proposed draft tokens in that fixture.
• Sprig's exact-string edit format produced 9 exact edits and 11 structurally valid edits from 12 fixtures in one trial.

The BF16 comparison generated 92.6 tokens/s without MTP and 115.2 tokens/s with block size 3 on the same M4 Max run.

In this patched runtime, deterministic MTP output diverged from the corresponding non-MTP baseline late in the fixture. The canonical BF16 package showed the same behaviour, so this is not an FP16 conversion regression. Applications requiring byte-identical speculative and baseline output should validate their runtime separately.

These figures describe one machine and one small release fixture. They are not general model benchmarks.

Limitations

This is an aggressive low-bit quantisation. It trades some model quality for memory use and local generation speed. The FP16 sibling changes residual floating-point precision, not quantisation or active parameter count.

The model package contains no custom executable code. File sizes, SHA-256 digests and source revisions are recorded in RELEASE_MANIFEST.json and MTPLX_FP16_CONVERSION_MANIFEST.json.

Licence

Qwen3.6-35B-A3B is licensed under Apache 2.0. The full licence text is included in LICENSE. The Froggeric template repository and MTPLX also declare Apache 2.0 and are attributed above.