dealignai/GPT-OSS-120B-MLX-CRACK

Built for vMLX — the only MLX inferencer with VL support, KV cache quantization, prefix cache reuse, agentic tool calling, and speculative decoding.
_{Free for macOS · vmlx.net}

---

GPT OSS 120B — CRACK Abliterated (MLX)

OpenAI's open-weight 120B MoE model with refusal behavior removed via CRACK surgery.

Overview

This is openai/gpt-oss-120b with safety refusal behavior removed using the CRACK (Controlled Refusal Ablation via Calibrated Knockout) method. The model retains full intelligence, reasoning ability, and Harmony chain-of-thought while responding to all requests without refusal.

• Architecture: 36-layer Mixture-of-Experts (128 experts, 4 active per token)
• Parameters: ~117B total, ~5.7B active per token
• Format: Native MLX — BF16 attention + mxfp4 experts (as trained by OpenAI)
• Size: 61 GB
• Speed: ~80 tok/s on M3 Ultra 256GB

Note on precision: GPT OSS was trained natively in mxfp4 for expert weights. This is NOT post-training quantization — the experts were trained directly in 4-bit microscaling format. There is no higher-precision version. This is the native precision as released by OpenAI.

Test Results

Evaluated across 5 independent trials (1 greedy + 4 sampled at temp=0.6) with 32 prompts per trial:

Category	Result
Compliance (12 harmful prompts)	✅ 11.8/12 average (4/5 trials perfect 12/12)
Coherence (20 diverse prompts)	✅ 19.0/20 average (2/5 trials perfect 20/20)
Factual accuracy	✅ Correct (geography, science, math, history)
Code generation	✅ Working Python, algorithms, data structures
Creative writing	✅ Poetry, stories, recipes, summaries
Technical explanation	✅ Physics, biology, computing, economics

Thinking Depth Validation

Complexity	Greedy	Sampled
Simple factual	✅ 5/5	✅ 10/10
Multi-step reasoning	✅ 5/5	✅ 9/10
Complex creative/analytical	✅ 4/5	✅ 8/10

Overall: 91% pass rate across 45 thinking-depth tests at 3 temperatures.

Method

CRACK uses a calibrated, signal-proportional approach to weight surgery:

1. Probe the model to identify per-layer refusal directions across all 36 layers
2. Tier the intervention strength based on each layer's measured refusal signal — high-signal layers receive stronger surgery, low-signal layers receive gentler treatment
3. Apply targeted weight modifications to attention output projections and expert MLP weights in the early-to-mid layers
4. Validate across multiple trials, temperatures, and prompt categories

The tiered approach preserves the model's Harmony chain-of-thought channel-switching mechanism while effectively removing refusal behavior. Standard (non-norm-preserving) ablation is used, as empirical testing showed it outperforms magnitude-preserving alternatives on this architecture.

Recommended Settings

# Best results at greedy (temp=0) or warm (temp≥0.8)
# Moderate temperatures (0.5-0.7) occasionally produce channel loops on short creative prompts
temperature = 0.0  # or 0.8+

Usage

import mlx_lm
from mlx_lm.sample_utils import make_sampler

model, tokenizer = mlx_lm.load("dealignai/GPT-OSS-120B-MLX-CRACK")

prompt = "<|start|>user<|message|>Your prompt here<|end|><|start|>assistant<|message|>"

# Greedy (most reliable)
response = mlx_lm.generate(model, tokenizer, prompt=prompt, max_tokens=2048)

# Or with sampling
sampler = make_sampler(temp=0.8, top_p=0.95)
response = mlx_lm.generate(model, tokenizer, prompt=prompt, max_tokens=2048, sampler=sampler)
print(response)

Requirements

• Apple Silicon Mac with 64GB+ unified memory (model is 61 GB)
• Python 3.10+
• mlx-lm >= 0.30.0

Architecture Details

GPT OSS 120B is a unique MoE architecture with:

• 128 mxfp4 experts with 4 active per token (SwiGLU activation with hard clamp)
• Alternating sliding/full attention (64 GQA heads, 8 KV heads, 128-token sliding window)
• Harmony chat format with analysis/commentary/final channels for structured chain-of-thought
• Per-head attention sinks for persistent token attention
• YaRN RoPE for 128K context window

Other Models by dealignai

Qwen 3.5 CRACK Series

Model	Sizes
Qwen 3.5 VL (Vision+Language)	0.8B · 2B · 4B · 9B · 35B · 122B
Qwen 3.5 397B REAP CRACK	Text · VL

MiniMax M2.5 CRACK Series

Model	Quants
MiniMax 172B REAP CRACK	Q4 · Q6 · Q8
MiniMax 139B REAP CRACK	Q4 · Q6 · Q8

---

Support dealignai

All models are built from original research and published for free. These models are specifically crafted to be excellent coders and general-purpose assistants.

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

---

This model is provided for research and educational purposes. Users are responsible for compliance with applicable laws and regulations. The creators are not responsible for misuse.