baa-ai/MiniMax-M2.7-RAM-91GB-MLX

MiniMax-M2.7 — 91 GB (MLX)

Mixed-precision MLX build of MiniMaxAI/MiniMax-M2.7, prepared by baa.ai.

This is the smallest healthy M2.7 build in the series — ~91 GiB of unified memory at inference time. If you have 112 GB or more of unified memory, see the 100 GB variant below for a meaningful quality bump.

Metrics

Metric	Value
In-memory footprint	~91 GiB (measured RSS after load)
Size on disk	96.4 GB (19 shards)
Group size	64
Framework	MLX (Apple Silicon)

Benchmarks

Benchmark	Score	Notes
HumanEval pass@1 (single-shot)	79.9% (131/164)	164/164 completed, 0 skipped
HumanEval pass@1 (best-of-2)	90.2% (148/164)	Retry of the 33 single-shot failures recovered 17
Decode throughput (Apple Silicon)	36.2 tok/s (wall-gen) / 36.7 tok/s (task-mean)	317,753 tokens generated over 146.3 min

Settings for both runs match the Recommended inference settings below.

Quality vs. memory — what you get if you step up

If you have	Variant	HumanEval best-of-2	Δ vs this build
~91 GiB	this build (91 GB)	90.2%	—
~100 GiB	100 GB	94.5%	+4.3 pp
~111 GiB	111 GB	94.5%	+4.3 pp
~116 GiB	116 GB	93.3%¹	+3.1 pp
~120 GiB	120 GB	93.9%	+3.7 pp

¹ 116 GB single-shot is 93.3%; a best-of-2 retry has not yet been measured for that build.

In short: jumping from 91 GB to 100 GB buys you ~4 points of best-of-2 HumanEval for 9 GB more memory. Anything above 100 GB plateaus — the 100–120 GB tier all lands in the 93–94.5% band. If your machine can spare the memory, 100 GB is the sweet spot.

Recommended inference settings

sampler_params = {
    "temperature": 1.0,
    "top_p": 0.95,
    "top_k": 40,
    "repetition_penalty": 1.1,
    "max_tokens": 8192,
}

Chat template — thinking mode

MiniMax-M2.7 uses a <think>…</think> reasoning block. Important: the base chat template injects <think>\n at the end of the prompt before generation, so the model's output begins inside the reasoning block with no opening tag. Strip everything up to and including the first </think>:

def strip_thinking(text: str) -> str:
    if "</think>" in text:
        return text.split("</think>", 1)[1].strip()
    return text.strip()

Give the model enough token budget that it can finish reasoning and emit the </think> closing tag — we recommend at least 4096, and 8192 for harder problems.

Usage

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler, make_logits_processors

model, tokenizer = load("baa-ai/MiniMax-M2.7-RAM-91GB-MLX")

sampler = make_sampler(temp=1.0, top_p=0.95, top_k=40)
logits_processors = make_logits_processors(repetition_penalty=1.1)

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Write a Python function that reverses a string."}],
    tokenize=False,
    add_generation_prompt=True,
)

response = generate(
    model,
    tokenizer,
    prompt=prompt,
    max_tokens=8192,
    sampler=sampler,
    logits_processors=logits_processors,
)

if "</think>" in response:
    response = response.split("</think>", 1)[1].strip()
print(response)

Hardware

• Apple Silicon Mac with ~104 GB unified memory recommended for comfortable inference (model + context + headroom).
• Fits in 96 GB Macs with minimal context headroom.
• Runs on less with swap, at substantially reduced throughput.

Variants

Variant	Size	Link
91 GB	96.4 GB	baa-ai/MiniMax-M2.7-RAM-91GB-MLX
100 GB	100.1 GB	baa-ai/MiniMax-M2.7-RAM-100GB-MLX
111 GB	110.9 GB	baa-ai/MiniMax-M2.7-RAM-111GB-MLX
116 GB	116.0 GB	baa-ai/MiniMax-M2.7-RAM-116GB-MLX
120 GB	120.1 GB	baa-ai/MiniMax-M2.7-RAM-120GB-MLX

Black Sheep AI Products

Black Sheep AI builds private AI with the Bounded AI Architecture (BAA): three layers, each bound to something you control, on infrastructure you control. Nothing leaves your boundary.

Ram (Reasoning Layer) — Full-capability frontier models, shrunk to fit smaller GPUs with reasoning kept and certified. They rely strictly on your Data Modules for knowledge, so you swap the model as better ones ship.

Paddock (Knowledge Plane) — Binds your Data Modules directly to the reasoning layer: one per topic, period, person, permission or tenant. Permission, freeze, hand back or delete each one on its own.

Watchman (GRC Layer) — Governance, risk and compliance in a single, complete audit log: every module permission, every bind, every model's provenance. Five certificates before anything goes live.

Shepherd — Takes a frontier model, compresses it into a Ram reasoning model, certifies it with Watchman, binds it to your Paddock, and deploys the Bounded AI Architecture to your infrastructure: your cloud, a Mac Studio, or an air-gapped facility.

Learn more at baa.ai — Private AI for enterprises whose data cannot leave the building.

License

Inherited from the upstream MiniMax-M2.7 license: non-commercial use permitted; commercial use requires written authorization from MiniMax.

---

Quantized by baa.ai