Skip to content

Releases

v0.3.0: image input and MTP

Choose a tag to compare

@pmichalas pmichalas released this 18 Nov 11:44

Image input and faster decoding for SenseNova-Loop-9B on Apple Silicon.

New in v0.3.0

  • Image input: SenseNova's vision tower in MLX (bfloat16), image processor (about 1 MP per image by default) and 3-axis mRoPE positions through the 56-visit loop. scripts/generate_vision.py --model <build> --image photo.jpg --prompt "...". Text-only prompts take exactly the same path as before (bit-identical logits to v0.2.0).
  • MTP speculative decoding: SenseNova's MTP head drafts tokens that the looped model verifies in one pass, with exact Gated DeltaNet rollback per visit slot. 6-bit build on a MacBook Pro M5 Max: 19.3 → 33.5 tok/s at k=3 (1.74×). Greedy output matches plain decoding except at near-ties of the two best tokens.
  • Float32 norm weights: the 81 RMSNorm weights can be loaded in float32 (model.fp32norms.safetensors), as SenseNova's code computes them; bf16 build top-1 agreement with SenseNova's code 98.6% → 99.1%.

Builds on Hugging Face (updated with vision.safetensors, mtp.safetensors, model.fp32norms.safetensors)

6-bit 8-bit 4-bit bf16
Same top token as SenseNova's code (text) 96.7% 98.6% 84.4% 99.1%
Image prompts: same top token as bf16 98.0% 96.7% 95.4% reference

Apache-2.0. Not affiliated with SenseTime / SenseNova.

v0.2.0: SenseNova-Loop-9B release

Choose a tag to compare

@pmichalas pmichalas released this 17 Nov 18:29

Support for the final SenseNova-Loop-9B (revision 329a65f, released 17 November 2026): the unused loop_gate.logits tensor is dropped on load, YaRN rope scaling gives the 512K-token context setting, and the converter reads the BF16 release weights.

Builds on Hugging Face

Checked on a MacBook Pro M5 Max (128 GB), 17 November 2026

6-bit 8-bit 4-bit bf16
Same top token as SenseNova's PyTorch code 96.7% 97.6% 84.0% 98.6%

Every build answered short test prompts correctly with thinking on and off, and the 6-bit build found a pass code hidden in a 64K-token prompt. For benchmark results, see SenseNova's model card.

The Preview builds keep working with the same code.

Apache-2.0. Not affiliated with SenseTime / SenseNova.

v0.1.0: SenseNova-Loop-9B MLX port

Choose a tag to compare

@pmichalas pmichalas released this 07 Nov 19:29

First release of the MLX port of SenseNova-Loop-9B-Preview (revision 9937e91): looped layers 4–27 (run twice) and two-stream mHC with float32 routers, loaded by mlx-lm through trust_remote_code=True.

Builds on Hugging Face

Checked on a MacBook Pro M5 Max (128 GB), 7 November 2026

8-bit 6-bit 4-bit
Same top token as SenseNova's PyTorch code 97.6% 97.2% 89.2%
ARC-Challenge (first 300) acc / acc_norm 48.7 / 50.3 47.7 / 49.0 45.7 / 45.0
GSM8K (first 50) 94.0 88.0 92.0

Apache-2.0. Not affiliated with SenseTime / SenseNova.