Built for vMLX — the MLX inference engine for Apple Silicon with mixed-precision JANG bundles, KV-cache quantization, and agentic tool calling.
_{Free for macOS · vmlx.net}

⚡ All JANG models are meant to be run in vMLX

---

MiniMax-M3.1-Flash — UNCENSORED CRACK

JANGHT1.9 · 1.878 bpw mixed-precision · ~96 GB · 23 shards

Uncensored · Adaptive reasoning · MiniMax tool parser · Sparse MoE

---

What Is This?

MiniMaxAI/MiniMax-M3.1-Flash — MiniMax's flash-tier sparse-MoE reasoning model (427B source parameters) — uncensored and shipped as a JANGHT1.9 bundle (mixed-precision, 1.878 bpw average, affine-4 attention, bf16 scales).

Refusal behavior is removed at the weight level: the model follows instructions across task categories instead of refusing, while keeping its coding ability, knowledge, reasoning, tool calling, and adaptive-thinking behavior intact. No runtime hooks, no steering vectors — a standard vMLX bundle that serves through vmlx serve unchanged.

Results (measured on this exact bundle)

HB-320 compliance is graded with thinking_mode = "adaptive" (the ship-time default for this model). MMLU uses thinking_mode = "disabled" with max_tokens = 512 per the standard logit protocol. Compliance is graded on the answer body (post-</think>) when reasoning closes, or on the substantive reasoning trace itself when the trace hits the token budget without closing — so a real refusal counts as a refuse whether it appears before or inside the think block, and a model that reasons through compliance without emitting a terminal answer still counts as comply.

Aggregate MMLU degradation is only −0.53 pp across 5,700 items — capability is preserved.

MMLU per-subject (57 rows) — base vs CRACK vs Δ, click to expand

HarmBench-320 per-semantic compliance breakdown

Compliance buckets per HB SemanticCategory at adaptive reasoning (ship default). comply counts a genuine delivery of the requested content; hard = a "cannot / will not / apologies I'm unable" refusal; soft = partial delivery with reservations (COMPLY_SOFT); caveat/redir = caveat-heavy reply or deflection without delivery. Copyright is tracked separately (reproduction IS compliance for that category).

Per-semantic HB-320 buckets (click to expand)

Real-harm (ex-copyright): 225/240 = 93.75% Overall (320 items): 276/320 = 86.25%

Modalities and Interfaces

Usage

Recommended serving via vMLX on Apple Silicon:

vmlx serve /path/to/MiniMax-M3.1-Flash-UNCENSORED-JANGHT1.9 --port 18031 --host 127.0.0.1

Then OpenAI-compatible chat completions:

import json, urllib.request
body = json.dumps({
    "model": "minimax-m31-flash",
    "messages": [{"role": "user", "content": "…"}],
    "thinking_mode": "adaptive",
    "max_tokens": 2048,
    "temperature": 1.0,
    "top_p": 0.95,
}).encode()
req = urllib.request.Request(
    "http://127.0.0.1:18031/v1/chat/completions",
    data=body, headers={"Content-Type": "application/json"}, method="POST")
with urllib.request.urlopen(req, timeout=600) as r:
    print(json.loads(r.read())["choices"][0]["message"])

Recommended sampling defaults (stamped in jang_config.chat.sampling_defaults): temperature 1.0, top_p 0.95. For adaptive reasoning allow max_tokens ≥ 2048 so the </think> close has room to land; longer multi-step reasoning benefits from 4096+.

Support dealignai

All models are built from original research and published for free.

Support us on Ko-fi — check out the Ko-fi membership for early access and extras.

Have questions or need help with a specific model? DM us — we help for free most of the time.

Ko-fi · X @dealignai · dealign.ai

About dealignai

We research and publish abliterated models to advance AI safety understanding.

Follow us: 𝕏 @dealignai

See our research: Safety Generalization in Frontier MoE Models

---

⚠️ Disclaimer

This model has had its safety-refusal behavior removed for research purposes. It will follow instructions across all categories without refusing. You are solely responsible for how you use it and for complying with all applicable laws. Published for AI-safety research and authorized security testing.