LFM3-Audio-2B vs Nanbeige4.3-33B

LFM3-Audio-2B

No score change since it was listed.

Nanbeige4.3-33B

• -0.2 on 3 November 2026: product, new evidence

LFM3-Audio-2B

A 2B model that listens and speaks on an 8 GB machine, transcribes and translates 19 languages and turns spoken requests into tool calls. Only llama.cpp and Ollama run it in a release. The context is 32K, voice chat is English only, Liquid hasn't answered the repository's questions, and there are no safety results or data sources.

Nanbeige4.3-33B

A dense 33B Llama-architecture cousin for long agentic coding and office tasks, with per-request thinking, XML tool calls that stock vLLM and SGLang parse, and 512K of context. Nanbeige's INT4 build needs about 27 GB. Artificial Analysis's first independent runs give it 20.4 on its Intelligence Index, and there are no safety results, training data notes or licence file.

LFM3-Audio-2B

1. Start it with llama-server -hf LiquidAI/LFM3-Audio-2B-GGUF:Q4_K_M --jinja or ollama run lfm3-audio
2. llama.cpp ships examples/realtime_client.py for live speech with this model
3. Prefill the assistant turn with an open think tag to let it reason before a text reply
4. The context is 32K
5. Voice chat is English only, while transcription, synthesis and translation cover 19 languages and locales
6. Liquid doesn't recommend it for agentic coding or complex multi-step reasoning

Nanbeige4.3-33B

1. Serve with vllm serve Nanbeige/Nanbeige4.3-33B --enable-auto-tool-choice --reasoning-parser qwen3 --tool-call-parser qwen3_coder, as the card says, or SGLang with the same two parsers
2. On a 32 GB GPU use Nanbeige/Nanbeige4.3-33B-INT4 with the same vLLM or SGLang flags; the INT8 and FP8 builds need about 43 GB
3. Keep tool_call_format at its default xml, since qwen3_coder parses that format; the json option writes Hermes-style JSON calls, which need a different parser such as vLLM's hermes
4. Pass chat_template_kwargs {"preserve_thinking": true} for multi-turn tool use and coding, and {"enable_thinking": false} to skip reasoning, as the card recommends
5. Use temperature 1.0, top_p 0.95 and up to 65,536 new tokens for agent work, and temperature 0.6 for reasoning and chat, per the card
6. Send your own system prompt, since without one the template injects a Chinese identity line, or a Chinese tool-expert prompt when tools are passed
7. Send one system message at the start only, as the template raises an error for a later one
8. In plain Transformers pass use_cache=True to generate, since config.json and generation_config.json set it false; trust_remote_code isn't needed