How to Run MiniCPM V 4.7 35B A3B Locally

Built from openbmb's original weights with our own importance matrix. The calibration corpora behind our builds are public.

• You can now run MiniCPM in Atomic Chat.
• See our quantization analysis below for measurements and instructions.

---

Pick a file

Every number below is measured, not estimated. How we measured it is at the bottom, and the raw logs are in the metrics repo so you can check any of it yourself.

KL divergence is how far the quantized model's predictions drift from the original weights. Lower is better, and zero means identical. top-1 is how often it picks the same next word the original would have picked.

AD-IQ2_XXS-IQ1_M is a 1-bit experts layout: it matches the original's token choice 78.6% of the time. It fits a 12 GB card, but treat it as an experiment.

AD- marks an Atomic Dynamic layout. The name says what the two largest tensor groups got: AD-<ffn_down>-<ffn_up>, collapsed to one name when both match. Nothing is named after a type it does not contain.

Against llama.cpp's plain presets

The same reference, held-out text, machine and llama.cpp build; the plain presets are made with no importance matrix:

AD-Q4_K-IQ4_XS is a gigabyte smaller than Q4_K_M and drifts less than half as far: as far as Q5_K_M, which is 4.6 GB larger. Each AD file in this table is smaller than the plain preset next to it and closer to the original, so the plain presets are only here, not in the repo. Every AD layout here keeps attention, the shared expert, the embedding and the output head at q8_0 and spends the savings on the routed experts, which are 93% of the weights.

Which one fits your card

The file has to fit, and so does the context. This model keeps 20 KB of attention cache per token, which is 0.2 GB at 8k context and 0.6 GB at 32k. Budget for both.

[!TIP] If you are choosing between two neighbouring files, take the larger one. The steps between them cost one or two gigabytes and buy noticeably fewer wrong words.

Running it

llama-server -m MiniCPM-V-4.7-35B-A3B-AD-Q4_K-IQ4_XS.gguf -ngl 99 -c 8192 --jinja

The chat template is inside every file, llama.cpp applies it on its own. With --jinja it starts every answer in thinking mode; turn it off per request with "chat_template_kwargs": {"enable_thinking": false}.

Images and video

This model reads images, not only text. That needs one extra file, the vision projector, which is separate from the quant and shared by all of them:

• mmproj-MiniCPM-V-4.7-35B-A3B-F16.gguf for most people
• mmproj-MiniCPM-V-4.7-35B-A3B-BF16.gguf if your setup prefers bfloat16

Download it once alongside whichever quant you picked:

llama-mtmd-cli \
  -m MiniCPM-V-4.7-35B-A3B-AD-Q4_K-IQ4_XS.gguf \
  --mmproj mmproj-MiniCPM-V-4.7-35B-A3B-F16.gguf \
  --image your-photo.jpg -ngl 99 -c 8192 \
  -p "What is in this image?"

The calibration data

Our importance matrix comes from a corpus we built for this model and published: AtomicChat/calib-corpora, recipe minicpm-v-4.7-moe. It is 4,906,667 tokens across 2,896 documents:

Every conversation is rendered through this model's own chat template, so the chat markup appears as the single tokens the model actually sees.

[!IMPORTANT] If you build your own importance matrix from this corpus, pass --parse-special to llama-imatrix. Without it the chat markup is read as ordinary punctuation.

How we measured

Reference: the original BF16 weights, converted to GGUF and run unquantized. Perplexity on the held-out set is 8.6549 plus or minus 0.06668.

Held-out text: eval_neutral from our calibration dataset, never used for calibration. 87 chunks at 4096 context.

Metric: per-token KL divergence of each file's predictions against the reference, plus top-1 agreement.

Hardware: 2x H100 80GB HBM3.

The raw logits of the reference are published in the metrics repo, split into parts because of the file size limit. With them you can measure your own build against exactly the same point we did:

cat base-neutral.kld.*.part > base-neutral.kld

llama-perplexity -m your-quant.gguf -f eval_neutral.txt \
  --kl-divergence-base base-neutral.kld --kl-divergence -c 4096 -ngl 99

Reproducing a file

The importance matrix was collected on the BF16 weights rather than on a quantized copy, split across 8 workers and merged. Splitting by chunk range gives the same result as one long run, because the statistic is a sum:

llama-imatrix -m MiniCPM-V-4.7-35B-A3B-BF16.gguf -f calib_train.txt -o shard-0.gguf \
  -ngl 99 -c 512 -b 2048 -ub 2048 --parse-special --from-chunk 0 --chunks 1198

llama-imatrix -m MiniCPM-V-4.7-35B-A3B-BF16.gguf --in-file shard-0.gguf,shard-1.gguf,... \
  -o imatrix.gguf

Then, for AD-Q4_K-IQ4_XS:

llama-quantize \
  --imatrix imatrix.gguf \
  --tensor-type '^token_embd\.weight$=q8_0' \
  --tensor-type '^output\.weight$=q8_0' \
  --tensor-type '^blk\.\d+\.ffn_down_exps\.weight$=q4_k' \
  --tensor-type '^blk\.\d+\.ffn_(gate|up)_exps\.weight$=iq4_xs' \
  --tensor-type '^blk\.\d+\.ffn_(gate|down|up)_shexp\.weight$=q8_0' \
  --tensor-type '^blk\.\d+\.attn_(q|k|v|qkv|gate|output)\.weight$=q8_0' \
  --tensor-type '^blk\.\d+\.ssm_(out|alpha|beta)\.weight$=q8_0' \
  --override-kv general.file_type=int:30 \
  MiniCPM-V-4.7-35B-A3B-BF16.gguf MiniCPM-V-4.7-35B-A3B-AD-Q4_K-IQ4_XS.gguf Q8_0

The rules for every other file are in ladder/ in the metrics repo, next to its quantization log. The first rule that matches a tensor wins.

Model details

40 layers, hidden size 2048, vocabulary 248,144, context 262,144.

Needs llama.cpp at or after 69f201a (MiniCPM-V 4.7 support, ggml-org/llama.cpp#29416); older builds run text only and get image positions wrong. Built and tested with llama.cpp b11574 (8bb3cf0d2).