license: apache-2.0 base_model:
- swdq/MiniCPM-V-4.7-35B-A3B-abliterated
- openbmb/MiniCPM-V-4.7-35B-A3B tags:
- mlx
- mlx-vlm
- abliterated
- uncensored
- minicpm-v
- moe
- vision
- image-text-to-text language:
- en
- zh library_name: mlx-vlm pipeline_tag: image-text-to-text
MiniCPM-V-4.7-35B-A3B-abliterated-MLX-4bit
Refusal-ablated MiniCPM-V-4.7-35B-A3B as an MLX 4-bit quant for Apple Silicon — the first MLX build of an abliterated MiniCPM-V-4.7. Vision still works; only the language model was edited.
| Base | openbmb/MiniCPM-V-4.7-35B-A3B (Apache-2.0) |
| Ablation | swdq/MiniCPM-V-4.7-35B-A3B-abliterated (bf16, abliterix LoRA steering incl. MoE expert down_proj, Optuna trial 28/40) |
| This repo | bf16 → MLX 4-bit affine, group size 64 (mlx_vlm @ commit e928101a, the same recipe as mlx-community/MiniCPM-V-4.7-35B-A3B-4bit of the unablated base) |
| Size | ~23 GB, 5 shards; vision tower, embeddings, heads and linear attention kept bf16 |
Refusal removal, measured
House red-team v2 (50 probes: 8 controls, 18 dual-use, 16 sensitive, 8 fiction; greedy, thinking off, 400-token cap; keyword + manual verdict review), identical protocol for both models:
| tier | base 4-bit | this model |
|---|---|---|
| controls (must still answer) | 8/8 comply | 8/8 comply |
| dual-use refusals | 8/18 | 0/18 |
| sensitive refusals | 15/16 | 0/16 |
| fiction refusals | 2/8 | 0/8 |
| legacy-21 cut | 0/5 · 4/9 · 7/7 | 0/5 · 0/9 · 0/7 |
Upstream's own card numbers (100 harmful prompts, keyword detector): 100/100 → 5/100, KL divergence vs base on harmless prompts 0.101.
Capability spot-check (this quant vs the base 4-bit, same engine)
| instrument | base 4-bit | this model |
|---|---|---|
| tool calls, 280 BFCL-style cases (greedy, thinking off) | 247/280 = 0.879 | 248/280 = 0.886 |
| 30-case vision suite, exact-token scoring (screens/charts/documents) | 28/30 | 26/30 |
| writing, 24 cases, blind judges, both orders (position-consistent) | better 8 / human 8 / wrapped 0.33 | better 3 / human 7 / wrapped 0.42 (13 ties; base's edge is within noise, p ≈ 0.15) |
No statistically significant regression on any instrument; the point estimates lean slightly negative on two of three — the uniform small drift (KLD 0.101) the ablation leaves, with nothing collapsed. Tool calls, the load-bearing agentic skill, are fully intact.
Serving
pip install mlx-vlm
python -m mlx_vlm.server --model MiniCPM-V-4.7-35B-A3B-abliterated-MLX-4bit --port 8131
# or direct generation:
python -m mlx_vlm.generate --model MiniCPM-V-4.7-35B-A3B-abliterated-MLX-4bit --image img.png --prompt "What is this?"
Needs mlx-vlm ≥ 0.7.8 (minicpmv4_7 support). Peak memory ≈ 24 GiB with images → 32 GB Macs
minimum, 64 GB comfortable. enable_thinking is read as a top-level body field
(chat_template_kwargs also accepted).
Verification chain
- Source bf16: 36/36 files byte-matched to HF LFS metadata.
- Quant output: all shards size-matched between build host and release copy.
- Red-team results file and scoring scripts: Avicennasis/model-evals
runs/2026-10-14-minicpm-v47/.
Warning
Safety alignment is intentionally weakened. This model will produce harmful, offensive or illegal content on request. You are responsible for how you use it and for complying with applicable law and the base model's license. Apache-2.0.