Kimi K3 vs LFM3-VL-2B

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.

LFM3-VL-2B 77

Model card 17, Intended use (questions about video, documents, charts and app screens, finding objects, spatial reasoning), a limitation (not recommended for coding, multi-turn conversations or complex reasoning), architecture (language model, 14-layer vision encoder and projector), parameters (1.92B), context (128K) and 16 languages. Training data 5, A general description: the base's 22T tokens of pre-training with the kinds of data added since LFM2, then vision mid-training on image and video data, supervised fine-tuning, vision reinforcement learning and distillation. Last change 20, the last commit, on 9 November 2026, added the OpenVINO checkpoint to the card's inference table, a day before this check. Responsiveness 20, Of the 5 discussions that aren't pull requests (27 to 31 October 2026), 3 have a reply from a LiquidAI member, which is most. Version policy 5, LFM3 came out in new repositories and the LFM2.5 models stay in theirs. File health 10, Config, template and card agree on the lfm3_vl architecture, the 128,000-token context and the vision settings in processor_config.json. 17 + 5 + 20 + 20 + 5 + 10 = 77.

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.

LFM3-VL-2B 85

Memory 30, Liquid's Q4_K_M GGUF is 1.10 GB with its 0.20 GB projector file, and 1.30 GB × 1.2 = 1.6 GB, which fits 8 GB, well inside the 16 GB line. Runtimes 15, Released llama.cpp has the lfm3 architecture and the LFM3-VL projector (src/models/lfm3.cpp and tools/mtmd at b12242), Ollama's library has an lfm3-vl entry, and LM Studio's catalogue lists liquid/lfm3-vl-2b. Builds 15, Liquid's GGUF repository holds Q4_0, Q4_K_M, Q8_0, F16 and BF16 files, and Liquid publishes 4-bit and 8-bit MLX builds and an ONNX repository, all in the LiquidAI organisation. Terms 10, the LFM Open License v1.0, a custom licence that lets anyone use it commercially for free below a threshold of $10M annual revenue. Download 10, the Hub API reports gated false for the model and its GGUF, MLX and ONNX repositories on 10 November 2026, so the files download with no account and nothing to accept. Hosted 5, Liquid's own API lists lfm3-vl-2b among its models, so the publisher hosts it. 30 + 15 + 15 + 10 + 10 + 5 = 85.

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.

LFM3-VL-2B 40

File safety 15, safetensors only, no .bin, .pt or .pth files (25). Provenance 15, files in the LiquidAI organisation, which Liquid's docs page for the model and its blog post link to (10). Safety evaluation 0, no safety results on the card or in Liquid's blog post for the LFM3 family, neither of which mentions safety testing. Agent guidance 0, Nothing on prompt injection, an instruction hierarchy, what refusals cover or safeguards to add. Licence risk 10, 15 less 5 for the $10M annual revenue threshold on commercial use. Security programme 0, no way to report a flaw found. 15 + 15 + 0 + 0 + 10 + 0 = 40.

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.

LFM3-VL-2B 90

Chat template 20, chat_template.jinja in the repository (12,129 characters), with roles, images and video, tools, deferred tools, tool results and the thinking switch. Tool-calling format 25, the card and Liquid's tool-use page document tools passed through tools= in apply_chat_template, calls returned as XML and results sent back as tool messages. Trained tool calling 20, the template loops over every entry in an assistant turn's tool_calls, so several calls fit in one turn, and the card says its fine-tuning mix included tool calling and agentic traces. Structured output 5, the card's vLLM and SGLang commands set a structured-output backend and its fine-tuning table has a JSON extraction notebook, but nothing says how to ask this model for JSON. Liquid's API guide documents response_format with lfm3-1.7b as its example. Reasoning control 10, Liquid's chat-template page documents enable_thinking=False in apply_chat_template for LFM3, with reasoning on by default inside think tags, and the card documents preserve_thinking and reports a no-thinking configuration in its benchmarks. Context 10, 128K tokens of input and output per the card, and max_position_embeddings 128000 for the language model in config.json. 20 + 25 + 20 + 5 + 10 + 10 = 90.