MiniCPM-V 4.7 35B A3B Abliterated - NanoPlus V2.1

APEX Quantization Series by IsValorum

[!IMPORTANT] Independently computed using dedicated cloud compute clusters, funded 100% out-of-pocket.

If you consider these releases to be of high quality and they help you in your day-to-day workflow, consider supporting on Ko-fi. Every contribution directly sustains future cloud compute hours and helps support my university studies.

[!NOTE] Abliterated source: distributedcognition edited 74 tensors; no Heretic applied.

Vision: use the companion mmproj-Q8_0.gguf.

DSpark: the main GGUF has 733 tensors and no MTP tensors; the DSpark Q4_K_M/Q6_K draft is a separate companion.

Family cross-links

Tier	File	Base	Status
NanoPlus	`MiniCPM-V-4.7-35B-A3B-Abliterated.APEX-I-NanoPlus.gguf`	Abliterated BF16	Published
MiniPlus	see repo	Abliterated BF16	Published

Collection: apex-i-nanoplus-standard-moe-6acfee79c288c5d8e7307785.

Architecture and precision

Base model: distributedcognition/MiniCPM-V-4.7-35B-A3B-abliterated (Apache-2.0), abliterated by distributedcognition from openbmb/MiniCPM-V-4.7-35B-A3B.
Architecture: 40 layers, 30 linear-attention/SSM + 10 full-attention; 256 routed experts + 1 shared, 8 active; hidden 2048; 262,144-token context; vision tower 27 layers.
Main GGUF: 733 tensors. Vision is a separate Q8_0 projector; DSpark is a separate draft.
iMatrix: public mradermacher matrix, 319 x 512; covers 400 of 402 quantized text tensors. output.weight and token_embd.weight do not use it.
Recurrent shielding: ssm_alpha, ssm_a, ssm_dt.bias, ssm_conv1d and RMSNorms remain F32.
Layer 0 uses Q4_0 for attn_gate, attn_qkv, ssm_beta and ssm_out.

Audited tensor map

Tensor group	Layers	Count	Precision
`output.weight`	global	1	Q6_K
`output_norm.weight`	global	1	F32
`token_embd.weight`	global	1	Q4_K
`blk.0.attn_gate.weight`	layer 0	1	Q8_0
`blk..attn_norm.weight, blk..post_attention_norm.weight`	all 40	80	F32
`blk.0.attn_qkv.weight`	layer 0	1	Q4_K
`blk.*.ffn_down_exps.weight`	0-9	2	Q3_K
`blk.*.ffn_down_shexp.weight`	all 40	40	Q4_K
`blk.*.ffn_gate_exps.weight`	0-9	2	IQ3_XXS
`blk.*.ffn_gate_inp.weight`	all 40	40	F32
`blk.*.ffn_gate_inp_shexp.weight`	all 40	40	F32
`blk.*.ffn_gate_shexp.weight`	all 40	40	Q4_K
`blk.*.ffn_up_exps.weight`	0-9	2	IQ3_XXS
`blk.*.ffn_up_shexp.weight`	all 40	40	Q4_K
`blk.*.ssm_a`	SSM layers	30	F32
`blk.*.ssm_alpha.weight`	SSM layers	30	F32
`blk.0.ssm_beta.weight`	layer 0	1	Q4_K
`blk.*.ssm_conv1d.weight`	SSM layers	30	F32
`blk.*.ssm_dt.bias`	SSM layers	30	F32
`blk.*.ssm_norm.weight`	SSM layers	30	F32
`blk.0.ssm_out.weight`	layer 0	1	Q4_K
`blk.*.attn_gate.weight`	SSM layers except 0	29	Q8_0
`blk.*.attn_qkv.weight`	SSM layers except 0	29	Q4_K
`blk.*.ssm_beta.weight`	SSM layers except 0	29	Q4_K
`blk.*.ssm_out.weight`	SSM layers except 0	29	Q4_K
`blk.*.ffn_down_exps.weight`	0-9	8	IQ3_XXS
`blk.*.ffn_gate_exps.weight`	0-9	8	IQ2_S
`blk.*.ffn_up_exps.weight`	0-9	8	IQ2_S
`blk.*.attn_k.weight`	full attention 3,7,...,39	10	Q4_K
`blk.*.attn_k_norm.weight`	full attention 3,7,...,39	10	F32
`blk.*.attn_output.weight`	full attention 3,7,...,39	10	Q6_K
`blk.*.attn_q.weight`	full attention 3,7,...,39	10	Q4_K
`blk.*.attn_q_norm.weight`	full attention 3,7,...,39	10	F32
`blk.*.attn_v.weight`	full attention 3,7,...,39	10	Q4_K
`blk.*.ffn_down_exps.weight`	10-29	20	IQ3_XXS
`blk.*.ffn_gate_exps.weight`	10-29	20	IQ2_S
`blk.*.ffn_up_exps.weight`	10-29	6	IQ2_S
`blk.*.ffn_up_exps.weight`	10-29	14	IQ2_XXS
`blk.*.ffn_down_exps.weight`	30-39	8	IQ3_XXS
`blk.*.ffn_gate_exps.weight`	30-39	8	IQ2_S
`blk.*.ffn_up_exps.weight`	30-39	8	IQ2_XXS
`blk.*.ffn_down_exps.weight`	30-39	2	Q3_K
`blk.*.ffn_gate_exps.weight`	30-39	2	IQ3_XXS
`blk.*.ffn_up_exps.weight`	30-39	2	IQ3_XXS

The machine-readable version is tensor-map.tsv.

Files

File	Size	Notes
`MiniCPM-V-4.7-35B-A3B-Abliterated.APEX-I-NanoPlus.gguf`	12,553,899,680 bytes	Main NanoPlus V2.1 GGUF
tensor-map.tsv	33,214 bytes	Audited per-tensor types
chat_template.jinja	from source	MiniCPM-V 4.7 template
mmproj-Q8_0.gguf	733,004,960 bytes	Vision projector
dspark Q4_K_M	237,004,800 bytes	DSpark draft
dspark Q6_K	330,315,072 bytes	DSpark draft

Usage

llama-cli -m MiniCPM-V-4.7-35B-A3B-Abliterated.APEX-I-NanoPlus.gguf -p "Hello" -ngl 99 -c 32768
llama-server -m MiniCPM-V-4.7-35B-A3B-Abliterated.APEX-I-NanoPlus.gguf --mmproj mmproj-Q8_0.gguf --jinja -ngl 99 -c 32768

The main file has no MTP tensors. Use the DSpark companion separately if your runtime supports it.

Limitations

Base benchmarks are self-reported by OpenBMB; abliteration evaluation is from distributedcognition.
Vision and DSpark were not independently evaluated for this tier.
This is a quantized deployment tier, not the BF16 reference.

License

Apache-2.0, inherited from the source model. See the source model card.

IsValorum/MiniCPM-V-4.7-35B-A3B-Abliterated-APEX-I-NanoPlus-GGUF