LFM3-Audio-2B vs Nemotron 3.5 Lightning 30B A3B

LFM3-Audio-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.

Nemotron 3.5 Lightning 30B A3B 75

File safety earns 25, safetensors only with no Python files, and nothing from the repository runs to load it or parse its output. Provenance earns 15, files in NVIDIA's linked organisation with a hash for every file. Safety evaluation earns 0, since NVIDIA says it evaluated content safety but publishes no results. Agent guidance earns 10, a subcard that names prompt injection and the safeguards to add, without an instruction hierarchy. Licence risk keeps all 15 under OpenMDW-1.1. The security programme earns 10 for the reporting route on the card and NVIDIA's public bulletins.

LFM3-Audio-2B 52

Model card 17, Intended use (local agents, tool calling, structured outputs, fine-tuning), limitations (not recommended for agentic coding or complex multi-step reasoning, voice chat in English only), an architecture section, parameters (2B), context (32K) and 19 speech languages and locales. Training data 5, the card has no training section. Last change 20, the last commit, on 3 November 2026, changed the card's navigation icons, seven days before this check, and the one before it, on 31 October, edited the card. Responsiveness 0, Neither of the 2 discussions that aren't pull requests (27 and 31 October 2026) has a reply from a LiquidAI member. Version policy 5, LFM3 came out in new repositories and the LFM2.5 models stay in theirs. File health 5, the card gives a 32K context while config.json gives the language model 128,000 positions and Liquid's API lists 128,000 for the hosted model, so the card and the files don't agree on context. 17 + 5 + 20 + 0 + 5 + 5 = 52.

Nemotron 3.5 Lightning 30B A3B 75

The model card earns 20, all seven items with limitations in the linked subcard. Training data earns 20, datasets listed with sources and token counts, cut-offs and filtering, and the corpora published. Last change earns 20, counting NVIDIA's own inference code as for the other Nemotron listings. The repositories last changed on 10 September, 53 days ago, and TensorRT-LLM, which the card's serving section runs, changed its Nemotron-H model code on 14 October, 19 days ago. Responsiveness earns 5, one NVIDIA reply across 14 questions. Version policy earns 5 as a first release. File health earns 5, since reports about the card's recipes wait without a reply and the llama.cpp recipe names a quantisation the linked repository doesn't hold.

LFM3-Audio-2B 80

Memory 30, Liquid's Q4_K_M GGUF is 1.10 GB with its 0.17 GB projector file, and 1.27 GB × 1.2 = 1.5 GB, which fits 8 GB, well inside the 16 GB line. Runtimes 10, Released llama.cpp has the lfm3 architecture with the LFM3 audio projector and speech output (src/models/lfm3.cpp and tools/mtmd at b12242, with a realtime client example that starts this model), and Ollama's library has an lfm3-audio entry. 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-audio-2b among its models, so the publisher hosts it. 30 + 10 + 15 + 10 + 10 + 5 = 80.

Nemotron 3.5 Lightning 30B A3B 85

Memory earns 25, since ggml-org's 18.9 GB Q4_0 GGUF and Ollama's 19 GB nvfp4 build both need under 24 GB with the fifth added. Runtimes earn the full 25 from released vLLM, SGLang, llama.cpp, Transformers, Ollama and LM Studio. Builds earn 10, as NVIDIA's only quantised build is NVFP4. Terms earn 10, because OpenMDW-1.1 is permissive with no threshold or use policy but isn't an OSI licence. Download earns 10, with no gate or account. Hosted earns 5, since NVIDIA's own API serves it and Fireworks hosts it.

LFM3-Audio-2B 85

Chat template 20, chat_template.jinja in the repository (9,328 characters), with roles, audio, tools, deferred tools and tool results. 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 names turning spoken requests into tool calls as one of its four uses. Structured output 5, the card says the model is designed for structured outputs without saying how to ask for JSON. Liquid's API guide documents response_format with lfm3-1.7b as its example. Reasoning control 10, the card documents reasoning per request for text replies: the caller prefills the assistant turn with an open think tag to let the model reason, or a closed one for a direct answer, with a use_thinking example. Context 5, 32K tokens of input and output per the card. 20 + 25 + 20 + 5 + 10 + 5 = 85.

Nemotron 3.5 Lightning 30B A3B 90

The chat template earns 20, a Jinja file in the repository. The tool-calling format earns 25, since the card names the qwen3_coder parser and shows a tool call declared and read back. Trained tool calling earns 20, with tool-use training and several calls per turn in the template. Structured output earns 5, claimed from training with no instructions. Reasoning control earns 10, thinking off per request with runtime parsers that return the reasoning apart. Context earns 10 for up to 1M tokens.