Sign up

dealignai/Ling-2.6-flash-MXFP4-CRACK

🤗 On Hugging Facetext-generationmit107B params215 GBsafetensors✓ HF checksums availableupdated today
Magnet

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

Ling 2.6 Flash — MXFP4 + CRACK

MXFP4 quantized | CRACK abliterated | Hybrid MLA + Linear-Attn MoE | EN + ZH | 63 GB

What Is This?

This is Ling 2.6 Flash by inclusionAI — a 35B-parameter Mixture-of-Experts model with 256 routed experts (8 active per token) + 1 always-active shared expert, hybrid MLA + Lightning Linear-Attention architecture, native English + Chinese, 131K context.

It has been:

1. MXFP4 quantized — uniform 4-bit affine, group_size=32 — 63 GB

2. CRACK abliterated — permanent weight-level removal of safety refusal

| | |

|---|---|

| Base model | inclusionAI/Ling-2.6-flash (35B total, 1 shared + 8 routed active) |

| Architecture | bailing_hybrid — Multi-Latent Attention (MLA) every 8th layer + Lightning Linear-Attn elsewhere |

| Quantization | MXFP4 (Q4 g=32 affine) — 63 GB |

| MMLU-200 | 78.5% (base 80.0% — within −1.5pp) |

| HarmBench-320 | 97.8% comply (base 50.3% — +47.5pp) |

| Context | 131,072 native |

| Languages | English + Chinese (probed bilingual) |

| Speed | 30+ tok/s on M4 Max 128 GB |

| Fits on | 96 GB+ Macs |

MMLU-200 Results (thinking OFF)

| Model | Correct | Accuracy | No-match |

|---|:---:|:---:|:---:|

| MXFP4 Base | 160/200 | 80.00% | 6 |

| MXFP4 + CRACK | 157/200 | 78.50% | 10 |

| Δ | −3 | −1.5pp | +4 |

CRACK delta of −1.5pp is well inside the noise floor for a 200-question sample — capability essentially preserved.

HarmBench-320 Results

| Model | COMPLY | REFUSE | EMPTY |

|---|:---:|:---:|:---:|

| MXFP4 Base | 161 (50.3%) | 157 (49.1%) | 2 (0.6%) |

| MXFP4 + CRACK | 313 (97.8%) | 5 (1.6%) | 2 (0.6%) |

| Δ comply | | +47.5pp | |

Refusal directional removal lifts HarmBench compliance from ~50% to ~98% with negligible MMLU regression.

Ling 2.6 Flash CRACK Series

| Model | Format | Size | MMLU-200 | HarmBench-320 | Fits on |

|-------|:---:|:---:|:---:|:---:|:---:|

| MXFP4 + CRACK (this model) | affine 4-bit g=32 | 63 GB | 78.5% | 97.8% | 96 GB Mac |

| JANGTQ2 + CRACK | TurboQuant 2-bit experts + 8-bit affine | 29 GB | 81.0% | 100.0% | 48 GB Mac |

The JANGTQ2 variant is smaller and scores higher on both benchmarks — quant noise on the 2-bit routed experts ends up helping rather than hurting.

Usage

from mlx_lm import load, generate

model, tokenizer = load("dealignai/Ling-2.6-flash-MXFP4-CRACK")

messages = [{"role": "user", "content": "Hello — what can you do?"}]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=400, verbose=True))

mlx_lm >= 0.20 with bailing_hybrid model class is required.

About This Model

Ling 2.6 Flash is the latency-tier sibling in the Ling 2.6 family — fast multilingual instruction-follow + tool use. The chat template includes a ... reasoning block, but in practice this Flash variant is best treated as a non-reasoning instruct model: leave thinking OFF (the default) for benchmark-style work and short-form responses. For chain-of-thought reasoning prefer the larger Ling 2.6 Plus / Ring / Pro tier.

CRACK is a permanent weight-level abliteration that removes safety refusal from the always-active residual-stream writers. Multilingual (EN + ZH) refusal direction extraction means the model complies on both English and Chinese prompts. The vision tower (none here — Ling 2.6 Flash is text-only) and MoE routing/expert internals are untouched.

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.

Twitter ·

HF ·

Ko-fi

Disclaimer

This model has had its safety refusal circuits removed. It will produce responses that would normally be refused, including technical content on security testing, dual-use research, and sensitive topics. You are responsible for how you use it.

The CRACK abliteration process does not add new capabilities — it only removes the model's learned refusal patterns. All knowledge, including the knowledge used to produce unsafe outputs, was already present in the base Ling model.