Kimi K3 vs MiniCPM-V 4.7 1B

Kimi K3 50

Memory earns nothing, because the only build of 4-bit precision or better from Moonshot or a runtime's own library is Moonshot's MXFP4 release of 1,561 GB, about 1,873 GB with the fifth added, far above 256 GB. Ollama and LM Studio list it only as a cloud model. Runtimes earn 15, for model code in released vLLM, SGLang and llama.cpp; Transformers has no kimi_k3 code of its own and loads Moonshot's remote code, so it doesn't count. Builds earn 10 for one publisher precision, the MXFP4 files the model was trained for. Terms earn 10, a custom MIT-based licence that is free for commercial use below a revenue threshold. Download earns 10, with no gate and no account. Hosted earns 5, since Moonshot's API serves it and five inference providers are live for it on Hugging Face.

MiniCPM-V 4.7 1B 95

Memory 30, OpenBMB's own IQ4_XS GGUF is 0.64 GB, 1.54 GB with the 0.90 GB vision projector, and 1.54 GB × 1.2 = 1.9 GB, which fits 16 GB with room to spare. Runtimes 25, MiniCPMV4_7ForConditionalGeneration in vLLM's registry at v0.33.0, the minicpmv4_7 model type in SGLang's minicpmv.py at v0.5.23, the minicpmv4_7 model in Transformers at v5.22.2, a MiniCPM-V 4.7 guide and conversion code in llama.cpp at b12242, and an openbmb/minicpm-v4.7:1b tag in OpenBMB's namespace on Ollama, read as the 35B-A3B listing's is. Builds 15, OpenBMB publishes AWQ, GPTQ and int4 safetensors builds and a GGUF repository from IQ3_XS to F16, so quantised builds at two precisions or more. Terms 15, Apache 2.0 in the card's metadata and as the LICENSE of OpenBMB's MiniCPM-V repository, an OSI licence. Download 10, gated is false on the main repository and on the GGUF, AWQ, GPTQ and int4 repositories, so the files download with no account or terms. Hosted 0, Hugging Face lists no inference provider for the model (an empty provider mapping on 10 November 2026), and no provider's own list was read. 30 + 25 + 15 + 15 + 10 + 0 = 95.

Kimi K3 65

File safety earns 15, safetensors only with no pickle files, less 10 because the documented loading path runs Moonshot's Python with trust_remote_code and builds prompts with encoding_k3.py. Provenance earns the full 15, files in the organisation Moonshot links to with a SHA-256 for every shard. Safety evaluation earns 20 for the cyber assessment the UK AI Security Institute and CAISI ran and published, which found its safeguards didn't stop exploit development. Agent guidance earns 5 for the blog's general cautions about proactiveness and passing back thinking. Licence risk earns 10, losing 5 for the model-as-a-service revenue threshold and nothing else. The security programme earns nothing, with no security.txt, disclosure policy, SECURITY.md or public advisories found.

MiniCPM-V 4.7 1B 55

File safety 25, safetensors in the main repository and the AWQ, GPTQ and int4 builds, GGUF in the GGUF build, with no .bin, .pt, .pth or .py files in the main repository. Provenance 15, files in the openbmb organisation that OpenBMB's GitHub README links to (10), with a SHA-256 for both large files in the main repository (5). Safety evaluation 0, no safety results on the card. Agent guidance 0, the card has no limitations or disclaimer section, and no prompt-injection guidance, instruction hierarchy or statement of trained refusals. Licence risk 15, Apache 2.0 has no use policy, field-of-use limit, threshold or change clause, nothing is added on the card, and the LICENSE in OpenBMB's GitHub repository resolved on 10 November 2026. Security programme 0, no SECURITY.md in OpenBMB/MiniCPM-V and no security.txt on openbmb.cn (both 404 on 10 November 2026), and no advisories found. 25 + 15 + 0 + 0 + 15 + 0 = 55.

Kimi K3 41

The model card earns 11, covering intended use, architecture, parameters and context, but not limitations, languages or a knowledge cut-off. Training data earns 5 for a general description with no cut-off. The last change earns 15, an encoding_k3.py update on 20 August, 74 days before 2 November and within 90; the GitHub repository's last commit is 6 August. Responsiveness earns nothing, with no Moonshot reply in the 40 most recent discussions. Version policy earns 5 for a first release. File health earns 5, because a report that the tokeniser goes past the declared vocabulary waits without a reply.

MiniCPM-V 4.7 1B 48

Model card 8, Intended use (image, multi-image and video understanding on device), architecture (early visual-token compression inside the vision encoder) and parameters (1B) are there. Training data 0, no description of the training data on the card. Last change 20, the last commit, on 30 October 2026, updated the card and removed an unused asset folder, eleven days before this check. Responsiveness 0, the repository has no discussions that aren't pull requests (two pull requests in all), so nobody has asked anything yet and the line reads as none answered, as settled in calibration. Version policy 10, MiniCPM-V 4.5 and 4.6 stay in their own repositories and the GitHub README dates each release. File health 10, Config, template and card agree on the architecture and the thinking switch, and no report waits without a reply, since there are no discussions. 8 + 0 + 20 + 0 + 10 + 10 = 48.

Kimi K3 65

The chat template earns 10, because there is no Jinja template and the prompt format exists only as encoding_k3.py, code you run. The tool-calling format earns 10, since Moonshot documents the tools channel but the parser name comes only from vLLM's and SGLang's own recipes. Trained tool calling earns 20 for indexed parallel calls in one turn. Structured output earns 5, as the card names it among its linked guides without saying how to ask the weights for it, and the how sits in hosted-API docs. Reasoning control earns 10, with per-request effort documented on the card and reasoning returned as reasoning_content. Context earns 10 for 1,048,576 tokens trained in.

MiniCPM-V 4.7 1B 70

Chat template 20, chat_template.jinja in the repository, 7,249 characters. Tool-calling format 10, the card shows tools declared in the OpenAI tools array in a request to Transformers' server, but not how a call comes back, and names no parser for vLLM or SGLang. Trained tool calling 20, the template loops over tool_calls, so it renders several calls in one assistant turn, and the card gives a tool-calling request as a supported use. Structured output 0, Neither JSON nor JSON Schema output is documented or claimed for the model on the card. Reasoning control 10, the card passes enable_thinking per request in its examples, the template has been off by default since 19 October 2026, and the GGUF repository keeps thinking builds in their own folder. Context 10, 262,144 tokens, max_position_embeddings in the text config. 20 + 10 + 20 + 0 + 10 + 10 = 70.