AtomicChat/MiniCPM-V-4.7-1B-GGUF

How to Run MiniCPM V 4.7 1B 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.

File	Size	KL divergence	top-1
Q8_0	1,082 MB	0.00348	96.64%
AD-Q6_K	1,003 MB	0.00630	95.51%
Q6_K (plain preset)	838 MB	0.01686	92.66%
AD-Q5_K_M	735 MB	0.04418	89.08%
AD-Q5_K_M-Q4_K_M	693 MB	0.07105	85.97%
AD-Q4_K_M	666 MB	0.09674	83.70%
AD-IQ4_XS	654 MB	0.10690	82.89%
AD-IQ4_XS-IQ3_S	587 MB	0.19859	76.71%

A 1B model loses more per bit than a large one: even Q8_0 agrees with the original on 96.6% of tokens here. Below AD-IQ4_XS-IQ3_S the files drop under 75% top-1, so they are not published (their numbers are in the metrics repo).

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:

Plain preset	Size	KL divergence	top-1	AD file near that size	Size	KL divergence	top-1
Q4_K_M	672 MB	0.19053	77.42%	AD-Q4_K_M	666 MB	0.09674	83.70%
Q5_K_M	752 MB	0.05753	87.06%	AD-Q5_K_M	735 MB	0.04418	89.08%
Q6_K	838 MB	0.01686	92.66%	AD-Q6_K-Q5_K	875 MB	0.01887	92.52%

At 4 and 5 bits the AD files are smaller and drift half to three quarters as far. At 6 bits the plain preset wins. Half of this model is the embedding and the output head (a 248k vocabulary on 1,024 dimensions, a quarter of the weights each), and our 6-bit layouts have not found a better split for them yet. So Q6_K stays in this repo, and Q4_K_M, Q5_K_M and AD-Q6_K-Q5_K are only in this table.

Which one fits your card

The file has to fit, and so does the context. This model keeps 12 KB of attention cache per token, which is 0.1 GB at 8k context and 0.4 GB at 32k. Budget for both.

Every file here fits a 2 GB card with 32k of context.

Running it

llama-server -m MiniCPM-V-4.7-1B-AD-Q6_K.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-4.7-1B-F16.gguf for most people
• mmproj-MiniCPM-V-4.7-1B-BF16.gguf if your setup prefers bfloat16

Download it once alongside whichever quant you picked:

llama-mtmd-cli \
  -m MiniCPM-V-4.7-1B-AD-Q6_K.gguf \
  --mmproj mmproj-MiniCPM-V-4.7-1B-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. It is 4,942,284 tokens across 3,107 documents:

Part	Share
agentic and tool use	24.6%
code	17.7%
reasoning	14.7%
multilingual	13.8%
long context	12.2%
vocabulary sweep	10.1%
structured data	3.9%
graphics	3.0%

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

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 15.9138 plus or minus 0.11294.

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: 4x 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, in one run:

llama-imatrix -m MiniCPM-V-4.7-1B-BF16.gguf -f calib_train.txt -o imatrix.gguf \
  -ngl 99 -c 512 -b 8192 -ub 8192 --parse-special

Then, for AD-Q6_K:

llama-quantize \
  --imatrix imatrix.gguf \
  --tensor-type '^blk\.(0|1|19|20|21|22|23)\.ffn_(down|gate|up)\.weight$=q8_0' \
  --tensor-type '^blk\.\d+\.ffn_down\.weight$=q6_k' \
  --tensor-type '^blk\.\d+\.ffn_gate\.weight$=q6_k' \
  --tensor-type '^blk\.\d+\.ffn_up\.weight$=q6_k' \
  --tensor-type '^blk\.\d+\.attn_(q|qkv)\.weight$=q6_k' \
  --tensor-type '^blk\.\d+\.attn_gate\.weight$=q8_0' \
  --tensor-type '^blk\.\d+\.ssm_out\.weight$=q8_0' \
  --tensor-type '^output\.weight$=q8_0' \
  --tensor-type '^blk\.\d+\.attn_output\.weight$=q8_0' \
  --tensor-type '^token_embd\.weight$=q8_0' \
  --override-kv general.file_type=int:18 \
  MiniCPM-V-4.7-1B-BF16.gguf MiniCPM-V-4.7-1B-AD-Q6_K.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

24 layers, hidden size 1024, feed-forward size 3584, vocabulary 248,094, 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).