A family of OptiQ quants. Ready to load.

Mistral: Devstral, the tool-calling coding model.

Mistral's Devstral is an agentic coding model on the Mistral-Small-3.1 base, which makes it also multimodal. OptiQ quantizes only the language tower and keeps the vision encoder at bf16 in a sidecar, so the one repo loads text-only under stock mlx-lm and full image+text under optiq serve. It clears uniform 4-bit by +1.95 Capability Score, winning five of six benchmarks, with +9 on long-context retrieval. The first OptiQ quant of a Mistral-family model.

Diffusion LLM: OptiQ's first non-autoregressive family.

Diffusion language models are a different kind of model. They don't decode left-to-right; they iteratively un-mask a block of tokens over a handful of denoising steps. The family has two very different members: Google's DiffusionGemma-26B-A4B (block-diffusion, 128-expert MoE, image-text), and dhara-250m (tri-mode: AR + diffusion + self-speculation), a 250M model small enough to fine-tune on a laptop. OptiQ ships native, dependency-free decoders for both; for DiffusionGemma it lands a higher Capability Score on a smaller artifact, while dhara is held at 8-bit mixed precision, where it reproduces the bf16 reference exactly.

Nemotron 3: NVIDIA's Mamba-attention hybrid.

NVIDIA's Nemotron 3 Nano interleaves Mamba2 state-space blocks with a handful of full-attention layers, and the larger model adds a 128-expert sparse MoE. The 30B-A3B (≈3 B active per token) gains the most: OptiQ assigns per-layer 4/8-bit across the fused routed experts and clears uniform 4-bit by a full +2.0 Capability Score, winning or tying all six benchmarks. The dense 4B gains less but still beats uniform 4-bit.

MiniCPM5: a strong sub-gigabyte 1B.

OpenBMB's 1.08B-parameter Llama-architecture base, fully Apache-2.0. Hybrid-reasoning chat template with an enable_thinking flag. In non-thinking mode (the OptiQ benchmark recipe) it posts 52% MMLU, 65% IFEval, and 58% HumanEval on a model that weighs less than a gigabyte on disk. The OptiQ-4bit quant beats stock uniform-4 by 12 points on HumanEval and rescues HashHop from a 0% floor, same sensitivity-aware allocation story as the larger families.

Gemma-4: Google's instruct series.

Two small dense (e2b, e4b), the new 12 B (the unified text+vision Gemma-4, now with image input), and two large (31 B dense, 26 B-A4B sparse-MoE). Mixed-precision recovery is large here, gemma-4-e4b posts a +13.6-point Capability Score gain over uniform 4-bit, and the 12 B adds +6.4. Pair e4b or 31B with their matching -assistant-bf16 drafter for speculative decoding. All five also take image input through a bf16 vision sidecar; see the vision guide.

Google also publishes QAT-unquantized Gemma-4 checkpoints, bf16 weights tuned so that 4-bit quantization is near-lossless. OptiQ runs its per-layer sensitivity pass on those bases and still beats a uniform 4-bit quant of the same QAT base, by +0.2 to +2.1 Capability points: the mixed-precision allocation keeps finding headroom even after QAT has hardened the weights. The 31B model gains the most in absolute terms (+1.65, driven by a +9-point jump on long-context retrieval). Each scores against a genuine uniform-4 quant of its own QAT base, holding the base fixed.

Qwen3.6: strong reasoning.

Qwen3.6 in two configurations: a dense 27 B and a 256-expert MoE with 3 B active per token. Both quantized with the same mlx-optiq pass; both ship a bundled MTP head for ~1.4× decode via optiq serve --mtp. Both beat uniform 4-bit on the six-metric Capability Score, and both take image input via a bf16 vision sidecar (vision guide).

Qwen3.5: the daily-driver series.

From 0.8 B for prompt-rewriters and toy agents up to 27 B for serious reasoning, plus a 35 B-A3B sparse MoE. All quantized with the same mlx-optiq pass; all ship a bundled MTP head for speculative decoding via optiq serve --mtp; all beat uniform 4-bit on the six-metric Capability Score. Every Qwen3.5 size also takes image input via a bf16 vision sidecar (vision guide).

The same snippet loads any model on this page.

Every OptiQ quant follows the same load contract. Swap the repo name; the rest stays.

from mlx_lm import load, generate

# Pick any of the 12. Stock mlx-lm, no special loader needed.
model, tok = load("mlx-community/Qwen3.6-27B-OptiQ-4bit")

prompt = tok.apply_chat_template(
    [{"role": "user", "content": "Explain mixed-precision quantization in 3 sentences."}],
    tokenize=False,
    add_generation_prompt=True,
)
out = generate(model, tok, prompt=prompt, max_tokens=300)
print(out)