add more models

[GandalfTea]

Summary

Add support for new model architectures:

• Olmo3
• GLM4
• Qwen3-MoE

Changes

 src/dnet/core/models/__init__.py  |   6 +++
 src/dnet/core/models/glm4.py      | 119 ++++++++++++++++++++++++++++++++++++++++++
 src/dnet/core/models/olmo3.py     | 120 ++++++++++++++++++++++++++++++++++++++++++
 src/dnet/core/models/qwen3_moe.py | 187 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 4 files changed, 432 insertions(+)

Testing

  Olmo 3
! mlx-community/Olmo-3-1025-7B-4bit          [FAIL] (junk output) 
+ mlx-community/Olmo-3-7B-Think-4bit
+ mlx-community/Olmo-3-7B-Think-SFT-4bit
+ mlx-community/Olmo-3-7B-Instruct-4bit
+ mlx-community/Olmo-3-7B-Instruct-SFT-4bit

+ mlx-community/Olmo-3-1025-7B-8bit
+ mlx-community/Olmo-3-7B-Think-8bit
+ mlx-community/Olmo-3-7B-Think-SFT-8bit
+ mlx-community/Olmo-3-7B-Instruct-8bit
+ mlx-community/Olmo-3-7B-Instruct-SFT-8bit

! mlx-community/Olmo-3-7B-Instruct-bf16.        [FAIL] (bf16 fails sampling)
! mlx-community/Olmo-3-7B-Instruct-SFT-bfloat16 [FAIL] 
! mlx-community/Olmo-3-7B-Think-bfloat16        [FAIL] 
! mlx-community/Olmo-3-7B-Think-SFT-bfloat16
! mlx-community/Olmo-3-1025-7B-bfloat16         [FAIL] 

+ mlx-community/Olmo-3-1125-32B-4bit
+ mlx-community/Olmo-3-1125-32B-8bit


  GLM
+ mlx-community/GLM-4-9B-0414-4bit
+ mlx-community/GLM-Z1-9B-0414-4bit
+ mlx-community/GLM-4-9B-0414-8bit 
+ mlx-community/GLM-Z1-9B-0414-8bit
+ mlx-community/GLM-4-32B-0414-4bit
+ mlx-community/GLM-Z1-32B-0414-4bit
! mlx-community/GLM-Z1-9B-0414-bf16 [FAIL] (failed sampling)
! mlx-community/GLM-4-9B-0414-bf16  [FAIL] (failed sampling)

  Qwen3-MoE
+ mlx-community/Qwen3-30B-A3B-4bit
+ mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit
+ mlx-community/Qwen3-Coder-30B-A3B-Instruct-8bit

Also modified existent catalogue entries:

- mlx-community/Qwen3-8B-bf16 (failed sampling)

Dependencies

Commit is dependent on distilp PRs: firstbatchxyz/distilp#18 and firstbatchxyz/distilp#17

[GandalfTea]

mlx-community/GLM-Z1-9B-0414-bf16 fails with:

2025-11-25 07:52:38,633 - dnet - ERROR - fit_in_memory.py:164 - End-shard sampling failed: [matmul] Last dimension of first input with shape (1,14,4096) must match second to last dimension of second input with shape (151552,4096).

[andthattoo]

Models are now tracked within catalogue. See catalog

[GandalfTea]

Added the working models to the catalogue. I'll look into 6bit and bf16 quantization problems

[ShivaThomas]

I'm curious. Are you not using the native MLX engine that already supports so many more model architectures? It would be wonderfull to be able to use any MLX supported model. Currently the project don't make much sense with my 128GB Mac. Only one supported model would be bigger than what I can run conventionally: Hermes-4-405B-MLX-4bit, at 228GB. But MOE like Qwen3 235B-A22B (132B at 4bit), or GLM 4.6 355B-A32B (198GB at 4bit, 154GB at 3bit) would be much more relevant. Large dense model are too slow for inference on Apple Silicon.

[andthattoo]

I'm curious. Are you not using the native MLX engine that already supports so many more model architectures? It would be wonderfull to be able to use any MLX supported model. Currently the project don't make much sense with my 128GB Mac. Only one supported model would be bigger than what I can run conventionally: Hermes-4-405B-MLX-4bit, at 228GB. But MOE like Qwen3 235B-A22B (132B at 4bit), or GLM 4.6 355B-A32B (198GB at 4bit, 154GB at 3bit) would be much more relevant. Large dense model are too slow for inference on Apple Silicon.

You're right on this. Reason we started very minimum is to test and expand the software itself. Although we are using MLX, models are not directly usable, we need to update model scripts and test them accordingly. In short, we'll add all the models supported by MLX.

We are also working on many optimizations including MoE runtime routing, expert assignments and sparsity.