Qwen3.5-4B-oQ4e-FP16-MTP-MLX
This is the FP16 precision sibling of
treeish/Qwen3.5-4B-oQ4e-MTP-MLX. It is intended for Sprig on M1 and M2 Macs, where FP16 avoids relying on BF16 arithmetic. M3 and newer Macs use the original BF16 package.
The oQ4e quantised weights, complete vision tower, 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/Qwen3.5-4B
- Architecture: dense 4B vision-language model
- Quantisation: oQ4e imatrix-enhanced mixed precision
- Default quantisation: 4-bit affine, group size 64
- Per-tensor overrides: 5 and 6-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 1,250 indexed tensors, including 297 vision-tower tensors and 29 embedded MTP tensors. It occupies essentially the same 3.3 GB as its BF16 sibling.
Conversion
The source is
treeish/Qwen3.5-4B-oQ4e-MTP-MLX at commit
dfd85130da4b921edf8237ead3103e8bdb801618.
The conversion used MLX 0.32.0 and follows the FP16 precision-sibling policy used by
youssofal/MTPLX at commit
963b923fa8be21f43194203777d5c075fc041ec9:
- 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 994 BF16 tensors to FP16 and preserved 256 packed uint32 tensors. All 297 vision tensors became FP16. Of the 29 embedded MTP tensors, 22 became FP16 and 7 packed uint32 tensors were preserved. MTPLX_FP16_CONVERSION_MANIFEST.json records every tensor, dtype, shape and source/output 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 16 GB of unified memory. 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 1,250 tensor names, shapes and index assignments match the BF16 source.
- Each converted tensor exactly matches the source value cast to FP16.
- All 256 packed tensors match the source value exactly.
- All safetensors metadata matches and all 16 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 averaged 100.2 tokens/s without MTP. MTP blocks 2, 3 and 4 averaged 95.3, 91.9 and 89.9 tokens/s, so Treeish's runtime optimiser can leave MTP disabled for this model and hardware combination.
- An alternating precision comparison averaged 100.0 tokens/s for FP16 and 98.8 tokens/s for BF16 without MTP.
- Sprig's exact-string edit format produced 7 exact edits from 12 fixtures and 10 structurally valid edits from 12 fixtures in one deterministic-format trial.
These figures describe one machine and one small release fixture. They are not general model benchmarks.
Limitations
This is a 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
Qwen3.5-4B 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.
