LFM3-Audio-2B vs StepAudio 3: The Open 2B Against the Scored Product Line

StepAudio 3 Realtime has a number attached to the thing voice products are judged on: 98.9% on Artificial Analysis's Conversational Dynamics board, first place. LFM3-Audio-2B has a licence. That is not a rhetorical contrast — it is the actual state of the evidence. StepFun's StepAudio 3 family went out on 15 September 2026 as five commercial audio models on one platform, with third-party placements and a preview price of zero. Liquid AI's LFM3-Audio-2B appeared on 27 October inside the six-model LFM3 drop as a 2B open-weight checkpoint with a vendor table and no independent score. One is a measured product you rent. The other is an unmeasured file you can own. The decision between them is less about audio quality than about which of those two sentences describes your constraint.

There is a temptation to declare the scored model the winner and move on. Resist it for a paragraph. StepAudio 3 Realtime's score measures conversational dynamics — turn-taking, interruption recovery, backchannels — on a hosted system with a persistent WebSocket and a Chinese-first training set. LFM3-Audio-2B is not on that board, is not duplex, and would not be measured by it even if someone entered it. The two models are not competing for the same slot in a stack. The useful comparison is what each one can do that the other cannot, and what each one costs you to find out.

What StepAudio 3 actually is

The name covers a family, and the tier you pick changes the comparison. StepFun shipped five audio models on 15 September: Realtime, ASR, TTS, Gen and Music, all live on its open platform as commercial APIs. A Chat tier with a 256K context and text output rounds out the developer surface.

• StepAudio 3 Realtime — native full-duplex conversation, reasoning directly over speech rather than transcribing first, with tool calling and asynchronous execution of long tasks while the conversation continues. The flagship, and the closest analogue to a general voice model.

• StepAudio 3 ASR — recognition, with a 1.7% word error rate on Artificial Analysis's non-streaming ASR accuracy board, tied for first.

• StepAudio 3 TTS — streaming synthesis, six languages (English, Chinese, Japanese, Korean, French, Spanish), 12 system voices, up to 1,000 characters of input.

• StepAudio 3 Gen — unified generation of voice, sound effects, ambience, background music and singing from natural language, currently on a limited free preview.

• StepAudio 3 Music — zero-shot song generation with multi-turn control via ABC notation, vocal-to-arrangement and covers.

LFM3-Audio-2B is one checkpoint that attempts several of those jobs at once: recognition, synthesis, translation, audio understanding and spoken tool calls, selected by prompt and serving profile. It will not generate music, and its sound-effect repertoire is whatever falls out of a general audio model. What it does have is a weight file. StepAudio 3 has none — it is a commercial API family, not an open release, and no tier ships weights.

The scoreboard, and who produced it

StepFun's numbers are the strongest evidence base in this matchup, and they are also the easiest to over-read. The 98.9% Conversational Dynamics figure is a third-party placement on Artificial Analysis, which makes it more credible than a vendor table. The 99.7% speech-reasoning figure is cited by StepFun, not independently reproduced. The 1.7% ASR word error rate is an Artificial Analysis placement, tied for first, and it belongs to the ASR tier rather than the Realtime model. Those are three different kinds of claim wearing the same uniform.

LFM3-Audio-2B's table is uniformly vendor-reported and uniformly relative to its own predecessor, LFM2.5-Audio-1.5B: English ASR WER 6.23 against 6.13, multilingual WER/CER 6.03, UTMOSv2 3.83 against 3.22, VoiceBench 73.13 against 55.17, MMAR 50.60 against 25.85, and an internal tool-calling eval at 58.50. No test set, harness or variance is named, and we could find no independent reproduction of any row.

Put the two side by side and the honest observation is that they barely overlap. StepAudio 3 Realtime is scored on conversation timing; LFM3-Audio-2B is scored on recognition, synthesis quality and audio reasoning. The one axis where both publish something is recognition, and even there the comparison is contaminated: StepAudio's 1.7% is an independent board score for the ASR tier, while Liquid's 6.23 is a vendor figure for a general audio model on an unnamed set. The gap is large enough that direction is probably real. The magnitude is not a measurement of anything two teams ran together.

Duplex is the architectural line

StepAudio 3 Realtime is built for a phone call. It listens while it speaks, handles interruption, and reasons over raw audio rather than a transcript, which StepFun argues preserves tone and emphasis a text hop would flatten. Its docs make the split explicit: the Realtime API brings built-in speech recognition, context management, VAD, web search and retrieval; the Chat API makes you implement each of those yourself.

LFM3-Audio-2B is built for a request. A causal audio encoder lets it prefill while you speak, and LiquidCodec streams speech back — 80 ms of audio into 12 tokens, decoded at about 87× real time on one CPU thread of an M4 Max. The "as low as 50 ms" time-to-first-token figure comes from that design. But streaming a reply is not the same as handling barge-in, and Liquid makes no duplex claim. If your product needs to be talked over, the architecture decides this before any benchmark does.

There is a latency catch on the other side. StepAudio 3 Realtime posts a time to first audio of 8.83 seconds on the same leaderboard where it wins conversational dynamics — against 1.18 seconds for Gemini 3.8 Live and 1.24 to 1.34 seconds for GPT-Live-1. A model that wins turn-taking but takes most of a sentence's worth of time to start speaking is a model whose best quality you may never demonstrate in production. Test latency before conversation quality, and test it outside StepFun's own serving setup if you can.

Price, licence, and what you are actually buying

StepAudio 3 is in limited-time free preview across the family, with no post-preview rate announced for the Realtime tier. The TTS tier has a published list rate of roughly $0.36 per 10,000 characters through a third-party catalogue, but that is not the Realtime price and should not be used as one. The practical position is that StepAudio 3 Realtime is free to evaluate today and unpriced for production — a good time to test, a bad time to build a unit-economics model.

LFM3-Audio-2B costs the machine it runs on. The LFM Open License v1.0 is free for commercial use below $10M annual revenue and paid above it. The Ollama build is 2.1 GB and CPU-capable; llama.cpp merged support on 27 October; Transformers needs 5.20.2 or newer with trust_remote_code, and the codec weights load from a separate directory, so pin the environment. There is no per-minute meter. There is also no SLA, no pager, and no vendor to call when the output degrades.

• What you get — StepAudio 3 Realtime: a commercial endpoint, current flagship, with support. LFM3-Audio-2B: a checkpoint, a licence, and the operational burden.

• What you can inspect — StepAudio 3: documentation, leaderboard placements, no weights. LFM3-Audio-2B: architecture, card, weights, codec, fine-tuning path.

• Cost shape — StepAudio 3: per call, zero when idle, unpriced after preview. LFM3-Audio-2B: fixed hardware, near-zero marginal cost, no meter.

• Failure mode — StepAudio 3: a quota, a region, an account problem, a support ticket. LFM3-Audio-2B: you own the pager.

Language coverage is not a tie

LFM3-Audio-2B lists 18 languages across 19 locales, with English appearing twice as US and UK variants. That is a broad synthesis list for a 2B model, and it is the strongest practical argument for it in a European or multilingual product. The limits are equally specific: live ASR is English-only, and translation is strongest to or from English.

StepAudio 3 is Chinese-first. The Realtime tier's documentation does not publish a full language list; the family's TTS tier covers six languages, and the ASR tier is built around Chinese and English with dialect handling, code-switching, long audio and singing. If your traffic is Mandarin with English mixed in, StepFun's stack is built for that workload and Liquid's is not. If your traffic is Swedish and Dutch, the 18-language checkpoint is the only one of the two that claims it.

Where a router fits, and where it does not

StepAudio 3 is not in the OrcaRouter catalogue, and LFM3-Audio-2B is not either — one is a closed API, the other a download. So there is no failover-between-these-two story to tell, and no reason to pretend otherwise.

The routing layer matters one level up, where voice agents actually spend their complexity. StepAudio 3 Realtime supports tool calling and long-running asynchronous tasks, which means the conversation is constantly handing work to ordinary text models: a lookup, an extraction, a policy check. That delegation layer is where provider outages and cost variance live. OrcaRouter puts more than 200 text and multimodal models behind one key at provider list price with no markup, with automatic failover, a routing DSL that composes several models into one call, and model fusion when one model's judgment is not enough. For a LFM3-Audio-2B prototype, that is the half of the stack you can put behind an endpoint today while the audio model stays local.

Three questions that decide it

Do you need a conversation that can be interrupted?

Then StepAudio 3 Realtime is the only one of these two that claims the behaviour, and it is the right starting point despite the latency question. LFM3-Audio-2B is request-and-response; no amount of fine-tuning changes the interface. Test the 8.83-second first-audio figure on your own network before you commit.

Do you need to own the model or run offline?

Then LFM3-Audio-2B is the only option here, and the licence threshold is the number to check against your revenue. Accept the 32K context, the English-only live ASR, and the absence of an independent evaluation. Run your own recordings through it before you trust the UTMOS gain.

Is a vendor score enough to choose?

Not yet, and the reason is not that StepFun's numbers are weak. It is that the scores measure different things from different harnesses, and the model without a score is not eligible for the board that produced the most quoted number. A neutral head-to-head on the same conversation set would settle more than either vendor's table. Until it exists, the choice is a constraint question, not a quality question.

The verdict

StepAudio 3 Realtime wins on the axis the market currently measures — conversational dynamics — and has the stronger evidence base because a third party produced part of it. It also has an unresolved latency problem, no published post-preview price, no weights and a Chinese-first language profile. It is the better choice for a real-time voice agent aimed at Chinese and English users, provided the latency test passes and the preview pricing does not turn into a surprise.

LFM3-Audio-2B wins on ownership and breadth of synthesis, and loses on evidence. It is four days old, its numbers are vendor-reported, its ASR is mid-table in its own table, and its licence has a revenue threshold. It is the better choice for an offline product, a multilingual voice, or a team that cannot accept a per-call meter — and the only one of the two you can evaluate entirely on your own hardware this week.

What to watch is the same on both sides: a neutral evaluation that puts a duplex model and a request-and-response model on the same conversation set, and a published production price for StepAudio 3. The first would tell you whether Liquid's 2B checkpoint is closer to the frontier than its paperwork suggests. The second would tell you whether StepFun's lead survives contact with a bill.