gmlx maps 48 GGUF architectures to model types, and 47 of them load end-to-end
with no hf_source. scripts/check-coverage.py generates this page from the
architecture table and the mlx-lm that gmlx carries, so do not edit it by hand.
A GGUF loads when gmlx maps its general.architecture to a model type, the
backend has a model class for that type, and a config synthesizer exists for
the architecture. The backend is the mlx-lm that gmlx carries for most
architectures, the mlx-vlm it carries for a few, and gmlx itself for the
encoder embedders. Without a synthesizer, pass hf_source to supply a
config.json. The status column says which case applies:
| Status | Meaning |
|---|---|
| loadable | The model class and the synthesizer are both present. |
| needs synth (hf_source) | The model class is present and no synthesizer exists, so pass hf_source. |
| needs mlx-lm (or another backend) | The backend has no model class for the architecture. |
| disabled (no known-good GGUF) | The code is complete, but each available GGUF is broken at conversion, so the loader refuses the architecture. |
The caveat column names the limit a user is most likely to hit, and an empty cell means none is known. Each mapped GGUF architecture has one row, with its family, backend, status and caveat:
| GGUF arch | Family | Backend | Status | Caveat |
|---|---|---|---|---|
beam |
beam | mlx-lm | loadable | Speculative decoding is not available: the DFlash2 drafter ships only as safetensors. |
deepseek2 |
deepseek | mlx-lm | loadable | DeepSeek-V2, with softmax gating, is not supported. |
deepseek4 |
deepseek | mlx-lm | loadable | The GGUF is not a llama.cpp conversion. |
deepseek41 |
deepseek | mlx-lm | loadable | MTP is not yet implemented. An engram table too large for one GPU buffer, or that alone brings the model under budget, is read from the file. |
diffusion-gemma |
gemma | mlx-vlm | loadable | |
ernie4_5-moe |
ernie | mlx-lm | loadable | |
falcon-h1 |
falcon | mlx-lm | loadable | |
flint |
flint | mlx-lm | loadable | The GGUF has no MTP head, so speculative decoding is not available. --kv-bits has no effect, and the KV cache stays fp16. |
gemma |
gemma | mlx-lm | loadable | |
gemma-embedding |
gemma | gmlx | loadable | |
gemma2 |
gemma | mlx-lm | loadable | Pass hf_source for 27B. Output matches llama.cpp only up to 4096 tokens, the size of the model's sliding attention window. |
gemma3 |
gemma | mlx-lm | loadable | Pass hf_source for 27B. |
gemma3n |
gemma | mlx-lm | disabled (no known-good GGUF) | |
gemma4 |
gemma | mlx-lm | loadable | |
glm-dsa |
deepseek | mlx-lm | loadable | |
glm4 |
glm | mlx-lm | loadable | |
glm4moe |
glm | mlx-lm | loadable | |
glm5-next |
glm | mlx-lm | loadable | The hybrid cache cannot go back in place, so after /retry or /undo, chat reads the earlier conversation again with the next message. |
glm5next |
glm | mlx-lm | loadable | The earlier name of glm5-next, which the first GLM-5.3-Flash uploads carry. Both load the same way. |
gpt-oss |
gpt-oss | mlx-lm | loadable | |
granite |
granite | mlx-lm | loadable | |
granitehybrid |
granite | mlx-lm | loadable | |
graniteswitch |
granite | mlx-lm | loadable | The GGUF has no MTP head, so speculative decoding is not available. |
hunyuan-moe |
hunyuan | mlx-lm | loadable | |
hy_v3 |
hunyuan | mlx-lm | loadable | Early GGUFs whose arch is hy-v3 with a dash are not mapped. Reconvert them. |
hyv4 |
hunyuan | mlx-lm | loadable | The GGUF has no MTP head, so speculative decoding is not available. --kv-bits has no effect, and the KV cache stays fp16. |
k2-horizon |
k2-horizon | mlx-lm | loadable | The GGUF has no MTP head, so speculative decoding is not available. |
kimi-k3 |
kimi | mlx-lm | loadable | The hybrid cache cannot go back in place, so after /retry or /undo, chat reads the earlier conversation again with the next message. |
llama |
llama | mlx-lm | loadable | |
minimax-m2 |
minimax | mlx-lm | loadable | |
minimax-m3 |
minimax_m3 | mlx-lm | loadable | Without indexer tensors or an indexer sidecar file, a GGUF runs dense. Its output is exact to 2048 tokens and degrades beyond that. |
mistral3 |
llama | mlx-lm | loadable | |
muse-glimmer |
muse | mlx-lm | loadable | |
muse-spark |
muse | mlx-lm | loadable | The GGUF has no MTP head, so speculative decoding is not available. |
nemotron_h_moe |
nemotron | mlx-lm | loadable | |
phi3 |
phi3 | mlx-lm | loadable | Pass hf_source for the 128K long-context variants. |
qwen2 |
qwen2 | mlx-lm | loadable | |
qwen2moe |
qwen2 | mlx-lm | loadable | |
qwen3 |
qwen3 | mlx-lm | loadable | |
qwen35 |
qwen3 | mlx-lm | loadable | |
qwen35moe |
qwen3 | mlx-lm | loadable | |
qwen3moe |
qwen3 | mlx-lm | loadable | |
qwen3next |
qwen | mlx-lm | loadable | |
qwen3vl |
qwen3 | mlx-lm | loadable | |
qwen3vlmoe |
qwen3 | mlx-lm | loadable | |
qwen4exp |
qwen3 | mlx-lm | loadable | Speculative decoding needs a companion MTP GGUF beside the model or in an MTP folder, or named with --draft-gguf. |
seed_oss |
seed_oss | mlx-lm | loadable | |
smollm3 |
llama | mlx-lm | loadable |