MiniCPM-V-4.7-35B-A3B-abliterated

Refusal-ablated version of openbmb/MiniCPM-V-4.7-35B-A3B, made with abliterix (LoRA steering incl. MoE expert down_proj, merged into bf16 weights; no quantization). Trial 28 of 40 (Optuna TPE).

	refusals (keyword detector, 100 harmful eval prompts)	KL divergence (harmless prompts)
base	100/100	0
this model	5/100	0.101

Only the language model was edited; the vision tower is unchanged and image input was not tested. Refusal counting is keyword-based (no LLM judge). Weights keep the original fused-expert layout (experts.gate_up_proj/down_proj).

Warning: safety alignment is intentionally weakened. The model may produce harmful, offensive or illegal content. You are responsible for its use and for complying with applicable laws and the base model license.

Serving with vLLM + DSpark (speculative decoding)

dspark/ contains the unmodified official draft model openbmb/MiniCPM-V-4.7-35B-A3B-DSpark (Apache-2.0), trained against the base model. Tested with the vLLM nightly (0.32.0rc1) on one DGX Spark (GB10): MiniCPM-V 4.7 is not in vLLM v0.31.0 and SGLang support is not merged yet.

huggingface-cli download swdq/MiniCPM-V-4.7-35B-A3B-abliterated --local-dir ./model
vllm serve ./model --served-model-name minicpm-abliterated --dtype bfloat16 \
  --max-model-len 131072 --max-num-seqs 8 --trust-remote-code \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3 \
  --default-chat-template-kwargs '{"enable_thinking": false}' \
  --speculative-config '{"method":"dspark","model":"./model/dspark","num_speculative_tokens":15,"enable_adaptive_verification":false}'

Measured decode speed (single stream, greedy): 25.5 tok/s without DSpark; with DSpark about 86 tok/s on code and 102 tok/s on math, but ~25-30 tok/s on prose/chat. At ~100k-token context decode was slower with DSpark (10 vs 20 tok/s). Claude Code works via ANTHROPIC_BASE_URL=http://localhost:8000.