MiniMax-M3-Alis-MLX-Dynamic

The first MLX quantization of MiniMax-M3 that keeps the full vision-language model. Sensitivity-graded mixed-precision builds of MiniMaxAI/MiniMax-M3 (427B total / ~23B active MoE VL, MiniMax Sparse Attention, 1M context) for Apple Silicon, sized for 512 GiB and 256 GiB machines.

Exact parameter count, measured across all 59 bf16 source shards: 427.04B (language model 426.18B + vision tower 0.63B + patch-merge 0.19B + projector 0.05B).

Existing MLX quants of M3 are text-only extractions — the vision tower, multimodal projector and patch-merge MLP are deleted. These builds keep all of it: image and video prompts work end-to-end through mlx-vlm.

Builds (branches)

Default branches load on stock mlx-vlm and on oMLX — no fork, no patch.

branch	routed experts	shared expert	attn/dense	embed / head	vision	size	fits
main (T256)	3-bit g64	3-bit (packed)	6-bit	6b / 8b	bf16	192.6 GB (3.65 bpw)	256 GiB Macs, no sysctl needed
t512 (T512)	6-bit g64	6-bit (packed)	8-bit	8b / 8b	bf16	350.8 GB (6.57 bpw)	512 GiB Macs
t512ref (T512REF)	8-bit g64	8-bit	8-bit	8b / 8b	bf16	454.9 GB (8.52 bpw)	512 GiB Macs (max quality / reference)

All three load on stock mlx-vlm and oMLX with no patch (t512ref keeps the unpacked layout but is uniform 8-bit, so its shared expert is already 8-bit and concatenates cleanly).

Never quantized, in every build:

• Router gates + e_score_correction_bias — fp32 (discrete top-4 expert selection)
• MSA index projections (index_q_proj/index_k_proj) — bf16 (they pick the top-16 attention blocks; a flipped selection reads different history, so this control path stays exact)
• Vision tower, multimodal projector, patch-merge MLP — bf16 (~1.4 GB total; multimodality is the point)

Note on the shared expert (compatibility over a marginal lever)

M3's always-on shared expert sees 100% of tokens (a routed expert sees ~3%), so holding it at 8-bit while the routed bank drops to 3/6-bit is a real quality-per-GB lever (+0.4% size). But MLX stores it packed into the same 129-wide SwitchLinear as the routed experts, and a packed bank can hold only one bit-width — so an 8-bit shared expert over a 3/6-bit routed bank needs the unpacked layout, which stock mlx-vlm and oMLX cannot load (they unconditionally re-pack and can't concatenate mixed bit-widths — see discussion #1).

These builds therefore keep the shared expert packed at routed bits so they load everywhere with no fork. If you want the 8-bit-shared variant, rebuild it with pack_shared_expert=false on the patched mlx-vlm — it is not hosted here.

Install & run

Stock, no fork:

pip install mlx-vlm    # >= 0.6.5

# text
mlx_vlm.generate --model avlp12/MiniMax-M3-Alis-MLX-Dynamic \
  --prompt "Explain MoE routing in three sentences." --max-tokens 300

# image
mlx_vlm.generate --model avlp12/MiniMax-M3-Alis-MLX-Dynamic \
  --image photo.jpg --prompt "Describe this image." --max-tokens 300

# video
mlx_vlm.generate --model avlp12/MiniMax-M3-Alis-MLX-Dynamic \
  --video clip.mp4 --prompt "What happens in this clip?" --max-tokens 300

Python:

import mlx.core as mx
from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("avlp12/MiniMax-M3-Alis-MLX-Dynamic")  # or revision="t512"
mx.set_wired_limit(mx.metal.device_info()["max_recommended_working_set_size"])

prompt = apply_chat_template(processor, model.config, "Describe this image.", num_images=1)
out = generate(model, processor, prompt, image=["photo.jpg"], max_tokens=300)
print(out.text)

Sampling: MiniMax recommends temperature=1.0, top_p=0.95. The model thinks in <mm:think>…</mm:think> before answering; budget max_tokens accordingly.

Measured quality (vs T512REF 8-bit reference)

Same-machine (M3 Ultra 512 GB), deterministic contexts (EN/KO/code/econ prose), KL on the reference's top-256 support; long-context on a 16K document exercising the sparse-attention path (MSA only activates beyond ~2.2K tokens — short evals cannot see it).

These numbers were measured on an earlier 8-bit-shared-expert prototype. The published packed builds keep the shared expert at routed bits (T256 → 3-bit, T512 → 6-bit): the T512 delta is negligible (routed already 6-bit), the T256 delta is bounded (one always-on FFN among a 3-bit bank). Figures are indicative; the qualitative ranking holds.

metric	T512	T256 (main)
KL vs REF, short (512 tok)	0.0184 nats	0.1243 nats
top-1 agreement, short	97.4%	90.4%
KL vs REF @16K	0.0087 nats	0.0011 nats
top-1 agreement @16K	99.2%	100.0%
NIAH @16K (3 depths)	3/3	3/3
sparse block-selection overlap vs REF @16K	54%	48%
vision (figure description + OCR of axis labels/annotations)	pass	pass
decode tok/s (short / 2.4K ctx)	22.5 / 17.7	28.1 / 20.6
prefill tok/s @2.4K	344	355
peak memory @2.4K	362 GB	205 GB

Reference itself: PPL 2.851 on the eval slice (T512 2.900, T256 2.994), NIAH 3/3, decode 20.6 tok/s.

Notes, honestly stated:

• T256's short-context KL (0.124) is dominated by a code context (0.217); prose contexts run 0.05–0.12. Target was <0.10 — close but a miss. A v2 with measured per-layer expert promotion may close it.
• Sparse block-selection overlap vs REF @16K (Jaccard on exact top-16 picks) reads low for both builds (T512 54%, T256 48%) — and nearly identically low, despite T512 being 2.9 bpw richer. Combined with 16K KL of 0.0011–0.0087 and 3/3 NIAH, this says the flips happen among near-tied blocks on repetitive text: selection differs, retrieval doesn't degrade.

Memory planning

KV cache ≈ 134 KiB/token fp16 (60L × 4 KV heads × 128d K+V, + 57 index-key caches) → ~13.7 GB per 100K tokens.

machine	default GPU limit (~75% RAM)	with sudo sysctl iogpu.wired_limit_mb=...
512 GiB	~412 GB — T512 fits, REF needs bump	up to ~475 GB: REF + ~20 GB ctx, T512 + 1M-token ctx
256 GiB	206 GB — T256 fits stock, ~11 GB ctx (80K tok)	up to ~240 GB: ~300K+ tok

A 128 GiB build was evaluated and cut: the routed-expert floor alone (2-bit g128 ≈ 117 GB) plus essentials exceeds a 128 GiB Mac's realistic wired ceiling — it cannot load, so we won't ship it.

Provenance & method

• Converted from the bf16 release (869 GB, 59 shards, all files size-verified + sha256 sampled against the Hub) with mlx_vlm.convert + a custom per-module quant predicate.
• MTP heads: not present in the source release (verified against the weight index).
• Playbook: sensitivity-graded mixed precision, 3-lens adversarial design review, ship gates defined before conversion. Same lineage as Hy3-Alis-MLX-Dynamic and GLM-5.2-Alis-MLX-Dynamic-4.5bpw.