src-e8426e0/docs/reference/models.md

Models

All models are on Hugging Face under txgsync. Each model card has the full results, the generation checks and the known issues.

Lineage

IFM/K2-Horizon-MoVA-36B-A4B                              original (GQA)
├─ K2-Horizon-MoVA-36B-A4B-NSLM-p4mx-q4v                 seeded experts, 22.76 GB
└─ K2-Horizon-MoVA-36B-A4B-MLA                           TransMLA conversion + distillation heal, BF16
   ├─ K2-Horizon-MoVA-36B-A4B-MLA-NanoSeedLM-P4          seeds, 22.20 GB
   ├─ K2-Horizon-MoVA-36B-A4B-MLA-NanoSeedLM-P3          seeds, 20.07 GB
   ├─ K2-Horizon-MoVA-36B-A4B-MLA-oQ4e                   affine (oMLX oQ), 22.89 GB
   ├─ K2-Horizon-MoVA-36B-A4B-MLA-oQ6e                   affine (oMLX oQ), 31.74 GB
   └─ K2-Horizon-MoVA-36B-A4B-MLA-oQ8e                   affine (oMLX oQ), 40.61 GB

IFM/K2-Horizon-0.9B                                      original (dense)
└─ K2-Horizon-0.9B-NSLM-seeds-4bpw                       seeds, 0.54 GB

Qwen/Qwen3.8-Flash-Next                                  original (qwen4_exp)
└─ mlx-community/Qwen3.8-Flash-Next-4bit-mtp             MLX 4-bit (group 32), with the MTP head
   ├─ Qwen3.8-Flash-Next-NanoSeedLM-P4                   seed experts (P = 4), the rest as the 4-bit model, 105.7 GB
   └─ Qwen3.8-Flash-Next-NanoSeedLM-P3                   seed experts (P = 3), the rest as the 4-bit model, 98.2 GB

The oQe models are for comparison. They let us compare seeds with a calibrated affine quantization of the same MLA model.

Runtimes

Model	Loader	oMLX	Metal engine	CUDA engine
NSLM-p4mx-q4v	nanoseedlm_k2.py	Yes	Yes	Yes
MLA-NanoSeedLM-P4	nanoseedlm_k2.py	Yes	Yes	Not checked
MLA-NanoSeedLM-P3	nanoseedlm_k2.py	Yes	Yes	Not yet (P = 3 tensors)
MLA (BF16)	k2_mla.py	Yes	Not checked	Not checked
MLA-oQ4e, oQ6e, oQ8e	k2_mla.py	Yes	No	No
0.9B-NSLM-seeds-4bpw	nanoseedlm_k2.py	Yes	No	No
Qwen3.8-Flash-Next-NanoSeedLM-P4, P3	nanoseedlm_qwen4.py (mlx-vlm)	Not yet	No	Yes (nslm-qwen4-gen)

oMLX removes model_file from qwen4_exp configurations, so it does not load the Qwen3.8-Flash-Next loader yet. Use mlx-vlm 0.7.12 or later with --trust-remote-code. To use a model in oMLX, read The MLX loader. To use a model in the engine, read Run and serve.

Which model to choose

• For the longest context on a Mac: MLA-NanoSeedLM-P4 or MLA-NanoSeedLM-P3 in oMLX.
• For the quality nearest to the original: NSLM-p4mx-q4v. It keeps the original attention.
• For a small download to test the seed format: K2-Horizon-0.9B-NSLM-seeds-4bpw. Its quality is lower than IFM's Q4_K_M, and it is slower.
• To try seed experts on Qwen3.8-Flash-Next: Qwen3.8-Flash-Next-NanoSeedLM-P4 or -P3, in mlx-vlm or with nslm-qwen4-gen on CUDA. Against the 4-bit MLX model, P4 agrees on the top token 91.9% of the time. KL divergence against BF16 is not measured yet.
• For daily use with no interest in seeds: an affine quantization of the original model.

Related data

• BF16 reference log-probs, for nslm-mova-score.
• Seed bank, the searched seed blocks.