LFM3-Audio-2B vs Qwen-Audio-3.1: One Checkpoint, Many Meters

Ask what Qwen-Audio-3.1 costs and the first honest answer is another question: which one? The name covers a tiered hosted speech line — a realtime duplex model, a realtime flash tier, streaming and file transcription, and two synthesis models — and each tier has its own rate card. LFM3-Audio-2B is the opposite kind of object: one 2B checkpoint from Liquid AI, shipped on 27 October 2026 inside the six-model LFM3 family, downloadable under the LFM Open License v1.0, with a 2.1 GB Ollama build and no per-call meter anywhere. Comparing them is not a comparison of audio quality. It is a comparison of two billing shapes, and the arithmetic lands differently depending on whether your traffic is steady, spiky, or still hypothetical.

The release shapes matter too, because they set expectations for documentation. Qwen-Audio-3.1 is catalogue-first: model IDs documented in Alibaba Cloud Model Studio, described in a line or two each, priced per token, with no weights published and no public parameter count. LFM3-Audio-2B is weights-first: a repository, a model card, a merged llama.cpp pull request, and one paragraph in its vendor's family post. One is a product line you rent. The other is a file you own. Every practical difference below follows from that.

What the name actually covers

Qwen-Audio-3.1 is not one endpoint, and treating it as one is the most common mistake in comparisons. The tiers do different jobs:

• qwen-audio-3.1-realtime-flash — real-time duplex speech, text and audio in and out, with function calling, web search, voice cloning and three turn-detection modes. This is the closest analogue to LFM3-Audio-2B, and the tier this article prices.

• qwen-audio-3.1-realtime-plus — the fuller realtime tier. Alibaba documents a 12.5-tokens-per-second audio conversion for it and charges a materially higher output rate than the flash tier. Different product, different bill.

• qwen-audio-3.1-asr-flash-streaming — recognition over a stream, with intermediate results, VAD, hotwords and disfluency removal. No speech output.

• qwen-audio-3.1-asr-flash-filetrans — the same recognition task batched against a file.

• Qwen-Audio-3.1-TTS and TTS-Next — synthesis, with the Next model marketed as an "Audiogen" model for voice, sound effects and ambience in one pass at 48 kHz.

LFM3-Audio-2B collapses those jobs into one checkpoint. Recognition, synthesis, translation, audio understanding and spoken tool calls are modes of the same 2B model, selected by prompt and by serving profile. There is no SKU to pick and no separate meter for speaking versus listening. The trade is granularity for simplicity: you cannot pay only for recognition when recognition is all you need, because you are running the whole model either way.

The bill, worked out

Alibaba publishes rates for qwen-audio-3.1-realtime-flash in USD per million tokens. In Singapore: text input $0.23, audio input $0.93, text output $0.70, audio output $1.87. Beijing is slightly cheaper at $0.212, $0.848, $0.636 and $1.696. Rate limits are 60 requests per minute and 100,000 tokens per minute, with no separate cap on concurrent WebSocket connections. The context window is 262,144 tokens, with a maximum input of 245,760 and output of 16,384 — but audio history is capped at 50 turns or 300 cumulative seconds, which is the limit a long conversation actually hits.

LFM3-Audio-2B has no rate card because Liquid does not sell it as an endpoint. The cost is the machine you run it on, the 2.1 GB download, and the licence threshold: free for commercial use below $10M annual revenue, paid above it. On a laptop that is near zero; on a GPU it is whatever you already pay for idle capacity. The codec is the efficiency claim worth knowing — LiquidCodec packs 80 ms of speech into 12 tokens and decodes at about 87× real time on a single CPU thread of an M4 Max — but that measures the codec, not the language model, and should not be read as end-to-end throughput on your hardware.

Now the part that decides it. A hosted realtime model bills per token of conversation, including the audio it generates, and a voice agent's cost lives in output. A self-hosted model bills per hour of hardware, whether or not anyone is talking. For a contact centre running at steady load, the fixed cost amortises and the per-hour number can fall below any rate card. For a product with spiky traffic — a compliance team that transcribes a thousand calls on the last day of the quarter and nothing for eleven weeks — per-call billing wins, because the GPUs bill through the silence. Neither shape is cheaper in general. They are cheaper at different utilisations.

Coverage, latency, and the claims that do not meet

• Languages — LFM3-Audio-2B: 18 languages, 19 locales, with live ASR limited to English and translation strongest to or from English. Qwen-Audio-3.1: Alibaba's documentation for the realtime flash tier does not publish a language list; its recognition tiers are documented around Mandarin, Cantonese, Sichuanese and other Chinese dialects, with automatic language detection. Treat coverage as unknown for Qwen until the docs enumerate it.

• Latency — LFM3-Audio-2B: "as low as 50 ms" time to first token, vendor-measured on named Apple hardware, with no percentile or harness. Qwen-Audio-3.1: Alibaba's family headline of 92 ms first-token comes from its own high-concurrency benchmark on a 0.6B specification, not third-party testing. Neither is a typical number.

• Duplex — LFM3-Audio-2B: request-and-response with streaming; a causal encoder allows prefill while you speak, but there is no barge-in claim. Qwen-Audio-3.1 realtime-flash: explicitly duplex, with server VAD and a smart-turn mode that handles interruptions.

• Context — LFM3-Audio-2B: 32K tokens, input plus output. This is not the 256K figure repeated in some LFM3 coverage; that belongs to the text model. Qwen-Audio-3.1 realtime-flash: 262K tokens, with the audio-history cap above.

That duplex line is the one that changes architecture. If your product needs to be interrupted mid-sentence, LFM3-Audio-2B is the wrong shape of model and no prompt fixes it. If your product reads a recording and answers a question about it, duplex is irrelevant and the 2B checkpoint is the simpler object.

Who has measured what

Liquid publishes six figures for LFM3-Audio-2B, all against its own previous model, LFM2.5-Audio-1.5B: English ASR WER 6.23 (versus 6.13, flat), multilingual WER/CER 6.03, UTMOSv2 3.83 (versus 3.22), VoiceBench 73.13 (versus 55.17), MMAR 50.60 (versus 25.85) and an internal tool-calling eval at 58.50. No test set, harness or variance is named. We could find no independent reproduction of any row.

Qwen's evidence is different in kind. There is a technical report, "Qwen-Audio-3.1-Realtime: Towards Reliable Agentic Voice Interaction", which reports task success on a half-duplex speech-to-text adaptation of τ-Voice improving from 78.4% to 82.0%, and the rate of responding to background speech on Full-Duplex-Bench v1.5 falling from 73.0% to 13.0%. Those are the two figures stated in the abstract; the body carries model-level tables, but I have not verified them against a harness, and the paper is a vendor report rather than an independent evaluation.

So the honest scoreboard is: one model has a vendor table with no paper, the other has a vendor paper with no reproducible harness, and neither appears on a neutral board under its own name. Any article that puts their numbers in adjacent columns is arranging two unverified claims and letting the layout do the persuading.

Where a router fits, precisely

Neither LFM3-Audio-2B nor Qwen-Audio-3.1 is in the OrcaRouter catalogue, so there is no routing claim to make about the speech layer itself. The layer above it is a different story. A voice product built on either model still needs a text model to summarise, classify, extract or decide what the agent says next, and that call is where cost and reliability actually vary.

OrcaRouter puts more than 200 text and multimodal models behind one key at provider list price with no markup, with automatic failover across providers and a routing DSL that can compose several models into one call. For this matchup the useful pattern is an A/B: run the same recorded conversations through qwen-audio-3.1-realtime-flash and through LFM3-Audio-2B on your own hardware, then send both transcripts to the same downstream model on the same endpoint. The comparison stays clean, and the text half never becomes the variable you forgot to control.

Which bill to choose

Pick Qwen-Audio-3.1 realtime-flash if you need duplex conversation, a Chinese-first recognition stack, a 262K context, or a vendor SLA — and if your traffic is spiky enough that per-token billing is a feature. Read the tier names carefully before you quote a price: realtime-plus and realtime-flash are not the same product, and the output rate is where a voice agent's bill is decided.

Pick LFM3-Audio-2B if you need to own the model, run offline, avoid a per-minute meter, or synthesise 18 languages on a laptop. Accept the 32K context, the English-only live ASR, the licence threshold, and the fact that its recognition accuracy is mid-table in its own vendor comparison.

Wait if you need a number you can defend to a procurement team. Neither model has one yet. The first independent evaluation on a shared audio corpus — not a vendor table, not a vendor paper — will be worth more than every figure above, and it is the thing to watch before either of these becomes a standard.