AFM — local AI infrastructure for Apple Silicon

AFM preflights a checkpoint's config.json before loading it. The model_type field—not the repository name—is normalized, checked against this allowlist, and used to choose a language or vision loading path.

69recognized canonical types

68runnable after safety policy

17vision-capable classifications

6dual-mode classifications

Complete architecture list

`model_type`	Classification	Important behavior
`acereason`	Language	Loads through the MLX language-model path.
`afmoe`	Blocked	Recognized but deliberately blocked because it is known to crash the Metal GPU driver.
`apertus`	Language	Loads through the MLX language-model path.
`baichuan_m1`	Language	Loads through the MLX language-model path.
`bailing_moe`	Language	Loads through the MLX language-model path.
`bitnet`	Language	Loads through the MLX language-model path.
`cohere`	Language	Loads through the MLX language-model path.
`deepseek_v3`	Language	Loads through the MLX language-model path.
`deepseek_v4`	Language	Loads through the MLX language-model path.
`ernie4_5`	Language	Loads through the MLX language-model path.
`exaone4`	Language	Loads through the MLX language-model path.
`falcon_h1`	Language	Loads through the MLX language-model path.
`fastvlm`	Vision	Recognized by the vision-model preflight path.
`gemma`	Language	Loads through the MLX language-model path.
`gemma2`	Language	Loads through the MLX language-model path.
`gemma3`	Dual-mode	The presence of vision_config determines the effective path.
`gemma3_text`	Language	Loads through the MLX language-model path.
`gemma3n`	Language	Loads through the MLX language-model path.
`gemma4`	Language	Loads through the MLX language-model path.
`gemma4_text`	Language	Loads through the MLX language-model path.
`glm_moe_dsa`	Language	Loads through the MLX language-model path.
`glm4`	Language	Loads through the MLX language-model path.
`glm4_moe`	Language	Loads through the MLX language-model path.
`glm4_moe_lite`	Language	Loads through the MLX language-model path.
`gpt_oss`	Language	Loads through the MLX language-model path.
`granite`	Language	Loads through the MLX language-model path.
`granitemoehybrid`	Language	Loads through the MLX language-model path.
`idefics3`	Vision	Recognized by the vision-model preflight path.
`internlm2`	Language	Loads through the MLX language-model path.
`jamba_3b`	Language	Loads through the MLX language-model path.
`joyai_llm_flash`	Language	Loads through the MLX language-model path.
`kimi_k2`	Language	Loads through the MLX language-model path.
`kimi_k25`	Language	Loads through the MLX language-model path.
`lfm2`	Language	Loads through the MLX language-model path.
`lfm2_moe`	Language	Loads through the MLX language-model path.
`lfm2_vl`	Vision	Recognized by the vision-model preflight path.
`lfm2-vl`	Vision	Hyphenated compatibility spelling; lfm2_vl is also accepted.
`lille-130m`	Language	Loads through the MLX language-model path.
`llama`	Language	Loads through the MLX language-model path.
`llava_qwen2`	Vision	Recognized by the vision-model preflight path.
`mimo`	Language	Loads through the MLX language-model path.
`minimax_m2`	Language	Loads through the MLX language-model path.
`mistral`	Language	Loads through the MLX language-model path.
`mistral3`	Dual-mode	The presence of vision_config determines the effective path.
`nanochat`	Language	Loads through the MLX language-model path.
`nemotron_h`	Language	Loads through the MLX language-model path.
`olmo2`	Language	Loads through the MLX language-model path.
`olmo3`	Language	Loads through the MLX language-model path.
`olmoe`	Language	Loads through the MLX language-model path.
`openelm`	Language	Loads through the MLX language-model path.
`paligemma`	Vision	Recognized by the vision-model preflight path.
`phi`	Language	Loads through the MLX language-model path.
`phi3`	Language	Loads through the MLX language-model path.
`phimoe`	Language	Loads through the MLX language-model path.
`pixtral`	Vision	Recognized by the vision-model preflight path.
`qwen2`	Language	Loads through the MLX language-model path.
`qwen2_5_vl`	Vision	Recognized by the vision-model preflight path.
`qwen2_vl`	Vision	Recognized by the vision-model preflight path.
`qwen3`	Language	Loads through the MLX language-model path.
`qwen3_5`	Dual-mode	The presence of vision_config determines the effective path.
`qwen3_5_moe`	Dual-mode	The presence of vision_config determines the effective path.
`qwen3_6`	Dual-mode	The presence of vision_config determines the effective path.
`qwen3_6_moe`	Dual-mode	The presence of vision_config determines the effective path.
`qwen3_moe`	Language	Loads through the MLX language-model path.
`qwen3_next`	Language	Loads through the MLX language-model path.
`qwen3_vl`	Vision	Recognized by the vision-model preflight path.
`smollm3`	Language	Loads through the MLX language-model path.
`smolvlm`	Vision	Recognized by the vision-model preflight path.
`starcoder2`	Language	Loads through the MLX language-model path.

Family view

The runtime works from exact model_type values, but this view makes the breadth easier to scan.

Llama / Mistral

llama · mistral · mistral3 · starcoder2 · jamba_3b · pixtral

Gemma

gemma · gemma2 · gemma3 · gemma3_text · gemma3n · gemma4 · gemma4_text · paligemma

Qwen

qwen2 · qwen3 · qwen3_moe · qwen3_next · qwen3_5 · qwen3_5_moe · qwen3_6 · qwen3_6_moe · qwen2_vl · qwen2_5_vl · qwen3_vl · llava_qwen2

DeepSeek / Kimi

deepseek_v3 · deepseek_v4 · kimi_k2 · kimi_k25 · joyai_llm_flash

GLM

glm4 · glm4_moe · glm4_moe_lite · glm_moe_dsa

Phi / Granite / NVIDIA

phi · phi3 · phimoe · bitnet · granite · granitemoehybrid · nemotron_h

Liquid / Cohere

lfm2 · lfm2_moe · lfm2_vl · lfm2-vl · cohere

Specialized and emerging

acereason · openelm · internlm2 · mimo · falcon_h1 · smollm3 · ernie4_5 · baichuan_m1 · exaone4 · gpt_oss · lille-130m · olmoe · olmo2 · olmo3 · bailing_moe · nanochat · apertus · minimax_m2

Vision-specific

idefics3 · smolvlm · fastvlm

afmoe is omitted from the family view because AFM's preflight safety policy blocks it before model creation.

Normalized aliases

AFM accepts these upstream spelling variants and maps them to one canonical type before checking support:

Declared value	Canonical value
`qwen3.5`	`qwen3_5`
`qwen3.5_moe`	`qwen3_5_moe`
`qwen3.5_next`	`qwen3_next`
`qwen3.5_vl`	`qwen3_vl`
`qwen3.6`	`qwen3_6`
`qwen3.6_moe`	`qwen3_6_moe`
`qwen3_6_next · qwen3.6_next`	`qwen3_next`
`qwen3_6_vl · qwen3.6_vl · qwen3_5_vl`	`qwen3_vl`
`gemma_4 · gemma-4`	`gemma4`
`gemma_4_text · gemma-4-text`	`gemma4_text`

How modality is selected

AFM reads the downloaded or local checkpoint's config.json.
It normalizes known aliases and rejects unknown types before allocating the model.
It inspects nested text_config, vision_config, and processor metadata—not only the type's name.
For dual-mode Qwen3.5/Qwen3.6 and Gemma/Mistral-style configurations, the actual config layout selects the factory.
--media forces VLM intent for single-prompt mode; remote preflight may correct an incorrect text/vision guess.

Architecture-specific behavior

Architecture	Special path	Boundary
`qwen3_6*`	MTP compatibility detection	Also requires an `mtp.safetensors` sidecar and an eligible request.
`gemma4`	Native Gemma 4 tools; dense verifier may use EAGLE3	Adaptive XML is bypassed. EAGLE3 excludes MoE and needs a drafter.
`deepseek_v4`	Native MLX path, checkpoint conversion, or DwarfStar executor	DwarfStar requires a self-contained local layout and accepts fewer controls.
`cohere`-family checkpoints	Thinking tags and template behavior may force serial generation	Correctness takes precedence over requested batch concurrency.
XML-function families	Qwen-compatible parser and optional strict tool grammar	Parser selection changes tool-call interpretation.
`afmoe`	Preflight block	No opt-out: prevents a known Metal driver crash.