MiniCPM-V 2.6
MiniCPM-V 是由面壁智能与清华大学自然语言处理实验室联合研发的多模态视觉语言模型系列。 该系列侧重于在端侧设备上实现多模态能力的部署，通过特定的模型架构设计，使模型能够处理图像输入并根据文本指令进行响应，涵盖了图像理解、多轮对话及视频解析等应用场景。
- 核心特点：模型支持处理不同长宽比的图像输入，并具备视频理解能力，能够对视频内容进行总结和问答。 其技术重点在于提升了像素级的空间感知能力，支持坐标定位与物体追踪。同时，模型针对端侧运行进行了优化，在降低显存占用的前提下，能够处理复杂表格、长图以及各类文本 OCR 提取任务。
- 版本说明：本模型 MiniCPM V 2.6 是该系列中的一个具体型号，参数量约为 80 亿。 它在保持较小参数规模的基础上，集成了单图、多图以及短视频的理解能力。该版本适用于移动端设备部署，以及对推理延迟和计算成本有特定要求的实时交互与自动化办公场景。
环境配置
参考 llama.cpp 文档准备好 llama.cpp 工具。
快速开始
下载模型
O6 / O6N
pip3 install modelscope
cd llama.cpp
modelscope download --model radxa/minicpm-v-2_6-gguf ggml-model-Q5_K_M.gguf --local_dir ./
modelscope download --model radxa/minicpm-v-2_6-gguf mmproj-model-f16.gguf --local_dir ./
运行模型
O6 / O6N
./build/bin/llama-mtmd-cli -m ./ggml-model-Q5_K_M.gguf --mmproj ./mmproj-model-f16.gguf -p "What is this picture about?" --image ./tools/mtmd/test-1.jpeg
完整转换流程
克隆模型仓库
O6 / O6N
cd llama.cpp
hf download openbmb/MiniCPM-V-2_6 --local-dir ./MiniCPM-V-2_6
创建虚拟环境
O6 / O6N
python3 -m venv .venv
source .venv/bin/activate
pip3 install -r requirements.txt
模型转换
分离模型
O6 / O6N
python3 ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ./MiniCPM-V-2_6
转换视觉模块
O6 / O6N
python3 ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ./MiniCPM-V-2_6 --minicpmv-projector ./MiniCPM-V-2_6/minicpmv.projector --output-dir ./MiniCPM-V-2_6/ --minicpmv_version 3
转换文本模块
O6 / O6N
python3 ./convert_hf_to_gguf.py ./MiniCPM-V-2_6/model
模型量化
这里采用 Q5_K_M 量化。
O6 / O6N
./build/bin/llama-quantize ./MiniCPM-V-2_6/model/Model-7.6B-F16.gguf ./MiniCPM-V-2_6/model/ggml-model-Q5_K_M.gguf Q5_K_M
模型测试
模型测试输入
O6 / O6N
./build/bin/llama-mtmd-cli -m ./MiniCPM-V-2_6/model/ggml-model-Q5_K_M.gguf --mmproj ./MiniCPM-V-2_6/mmproj-model-f16.gguf -p "What is this picture about?" --image ./tools/mtmd/test-1.jpeg
模型输出：
$ ./build/bin/llama-mtmd-cli -m ./MiniCPM-V-2_6/model/ggml-model-Q5_K_M.gguf --mmproj ./MiniCPM-V-2_6/mmproj-model-f16.gguf -p "What is this picture about?" --image ./tools/mtmd/test-1.jpeg
build: 7110 (3ae282a06) with cc (Debian 12.2.0-14+deb12u1) 12.2.0 for aarch64-linux-gnu
llama_model_loader: loaded meta data with 24 key-value pairs and 339 tensors from ./MiniCPM-V-2_6/model/ggml-model-Q5_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv 0: general.architecture str = qwen2
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.name str = Model
llama_model_loader: - kv 3: general.size_label str = 7.6B
llama_model_loader: - kv 4: qwen2.block_count u32 = 28
llama_model_loader: - kv 5: qwen2.context_length u32 = 32768
llama_model_loader: - kv 6: qwen2.embedding_length u32 = 3584
llama_model_loader: - kv 7: qwen2.feed_forward_length u32 = 18944
llama_model_loader: - kv 8: qwen2.attention.head_count u32 = 28
llama_model_loader: - kv 9: qwen2.attention.head_count_kv u32 = 4
llama_model_loader: - kv 10: qwen2.rope.freq_base f32 = 1000000.000000
llama_model_loader: - kv 11: qwen2.attention.layer_norm_rms_epsilon f32 = 0.000001
llama_model_loader: - kv 12: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 13: tokenizer.ggml.pre str = qwen2
llama_model_loader: - kv 14: tokenizer.ggml.tokens arr[str,151666] = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv 15: tokenizer.ggml.token_type arr[i32,151666] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 16: tokenizer.ggml.merges arr[str,151387] = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv 17: tokenizer.ggml.bos_token_id u32 = 151644
llama_model_loader: - kv 18: tokenizer.ggml.eos_token_id u32 = 151645
llama_model_loader: - kv 19: tokenizer.ggml.unknown_token_id u32 = 128244
llama_model_loader: - kv 20: tokenizer.ggml.padding_token_id u32 = 151643
llama_model_loader: - kv 21: tokenizer.chat_template str = {% for message in messages %}{% if lo...
llama_model_loader: - kv 22: general.quantization_version u32 = 2
llama_model_loader: - kv 23: general.file_type u32 = 17
llama_model_loader: - type f32: 141 tensors
llama_model_loader: - type q5_K: 169 tensors
llama_model_loader: - type q6_K: 29 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q5_K - Medium
print_info: file size = 5.06 GiB (5.71 BPW)
load: printing all EOG tokens:
load: - 151643 ('<|endoftext|>')
load: - 151645 ('<|im_end|>')
load: special tokens cache size = 24
load: token to piece cache size = 0.9310 MB
print_info: arch = qwen2
print_info: vocab_only = 0
print_info: n_ctx_train = 32768
print_info: n_embd = 3584
print_info: n_embd_inp = 3584
print_info: n_layer = 28
print_info: n_head = 28
print_info: n_head_kv = 4
print_info: n_rot = 128
print_info: n_swa = 0
print_info: is_swa_any = 0
print_info: n_embd_head_k = 128
print_info: n_embd_head_v = 128
print_info: n_gqa = 7
print_info: n_embd_k_gqa = 512
print_info: n_embd_v_gqa = 512
print_info: f_norm_eps = 0.0e+00
print_info: f_norm_rms_eps = 1.0e-06
print_info: f_clamp_kqv = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale = 0.0e+00
print_info: f_attn_scale = 0.0e+00
print_info: n_ff = 18944
print_info: n_expert = 0
print_info: n_expert_used = 0
print_info: n_expert_groups = 0
print_info: n_group_used = 0
print_info: causal attn = 1
print_info: pooling type = -1
print_info: rope type = 2
print_info: rope scaling = linear
print_info: freq_base_train = 1000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn = 32768
print_info: rope_finetuned = unknown
print_info: model type = 7B
print_info: model params = 7.61 B
print_info: general.name = Model
print_info: vocab type = BPE
print_info: n_vocab = 151666
print_info: n_merges = 151387
print_info: BOS token = 151644 '<|im_start|>'
print_info: EOS token = 151645 '<|im_end|>'
print_info: EOT token = 151645 '<|im_end|>'
print_info: UNK token = 128244 '<unk>'
print_info: PAD token = 151643 '<|endoftext|>'
print_info: LF token = 198 'Ċ'
print_info: EOG token = 151643 '<|endoftext|>'
print_info: EOG token = 151645 '<|im_end|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: CPU_Mapped model buffer size = 5184.87 MiB
......................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max = 1
llama_context: n_ctx = 4096
llama_context: n_ctx_seq = 4096
llama_context: n_batch = 2048
llama_context: n_ubatch = 512
llama_context: causal_attn = 1
llama_context: flash_attn = auto
llama_context: kv_unified = false
llama_context: freq_base = 1000000.0
llama_context: freq_scale = 1
llama_context: n_ctx_seq (4096) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_context: CPU output buffer size = 0.58 MiB
llama_kv_cache: CPU KV buffer size = 224.00 MiB
llama_kv_cache: size = 224.00 MiB ( 4096 cells, 28 layers, 1/1 seqs), K (f16): 112.00 MiB, V (f16): 112.00 MiB
llama_context: Flash Attention was auto, set to enabled
llama_context: CPU compute buffer size = 303.22 MiB
llama_context: graph nodes = 959
llama_context: graph splits = 1
common_init_from_params: added <|endoftext|> logit bias = -inf
common_init_from_params: added <|im_end|> logit bias = -inf
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
mtmd_cli_context: chat template example:
<|im_start|>system
You are a helpful assistant<|im_end|>
<|im_start|>user
Hello<|im_end|>
<|im_start|>assistant
Hi there<|im_end|>
<|im_start|>user
How are you?<|im_end|>
<|im_start|>assistant
clip_model_loader: model name:
clip_model_loader: description: image encoder for MiniCPM-V
clip_model_loader: GGUF version: 3
clip_model_loader: alignment: 32
clip_model_loader: n_tensors: 455
clip_model_loader: n_kv: 20
clip_model_loader: has vision encoder
clip_ctx: CLIP using CPU backend
load_hparams: projector: resampler
load_hparams: n_embd: 1152
load_hparams: n_head: 16
load_hparams: n_ff: 4304
load_hparams: n_layer: 27
load_hparams: ffn_op: gelu
load_hparams: projection_dim: 0
--- vision hparams ---
load_hparams: image_size: 448
load_hparams: patch_size: 14
load_hparams: has_llava_proj: 0
load_hparams: minicpmv_version: 3
load_hparams: n_merge: 0
load_hparams: n_wa_pattern: 0
load_hparams: model size: 996.02 MiB
load_hparams: metadata size: 0.16 MiB
load_tensors: ffn up/down are swapped
alloc_compute_meta: warmup with image size = 448 x 448
alloc_compute_meta: CPU compute buffer size = 55.81 MiB
alloc_compute_meta: graph splits = 1, nodes = 893
warmup: flash attention is enabled
main: loading model: ./MiniCPM-V-2_6/model/ggml-model-Q5_K_M.gguf
encoding image slice...
image slice encoded in 5523 ms
decoding image batch 1/1, n_tokens_batch = 64
image decoded (batch 1/1) in 5046 ms
encoding image slice...
image slice encoded in 5550 ms
decoding image batch 1/1, n_tokens_batch = 64
image decoded (batch 1/1) in 5063 ms
encoding image slice...
image slice encoded in 5540 ms
decoding image batch 1/1, n_tokens_batch = 64
image decoded (batch 1/1) in 5083 ms
The image is a black and white photograph of a newspaper, specifically the front page of The New York Times from July 21, 1969. The headline, written in bold, large font, proclaims "Men Walk on Moon". Below the headline, there's a subheading that reads "Astronauts Land on Plain; Collect Rocks, Plant Flag". The newspaper is open to a page with a photograph of a man walking on the moon. The man is wearing a spacesuit and is seen walking on the lunar surface. The photograph is accompanied by a caption that reads "A Powdery Surface Is Closely Explored". The newspaper is a significant historical artifact, marking the momentous event of the first human landing on the moon.
llama_perf_context_print: load time = 936.29 ms
llama_perf_context_print: prompt eval time = 33699.48 ms / 212 tokens ( 158.96 ms per token, 6.29 tokens per second)
llama_perf_context_print: eval time = 27574.13 ms / 156 runs ( 176.76 ms per token, 5.66 tokens per second)
llama_perf_context_print: total time = 61703.52 ms / 368 tokens
llama_perf_context_print: graphs reused = 154