LFM3-Omni-2B vs Qwen3.8-Omni-Flash: Only One of Them Can Speak

Put LFM3-Omni-2B and Qwen3.8-Omni-Flash in the same room and the difference that decides most products is not on the benchmark sheet. Both accept audio. Both accept video. Only one of them can answer out loud. Alibaba's model returns text and nothing else — its own documentation is explicit that spoken replies require a different model in the Qwen line — while Liquid AI's 2.17-billion-parameter checkpoint emits 24 kHz speech directly, or interleaves speech and text in a single response.

Everything else about the pair is a study in scale. Qwen3.8-Omni-Flash is a closed model on Alibaba's API since 18 September 2026, built on a 125-billion-parameter mixture of experts with roughly 6 billion active, carrying a one-million-token context. LFM3-Omni-2B is open weights, 2.17 billion parameters, 32K of context, and was announced on 27 October 2026 as one of six models in the LFM3 family post — with its repository public since 21 October and no launch post of its own. One you rent by the token; one you download and pay for in hardware.

The two products are not the same product

Stripped to what each is for, Qwen3.8-Omni-Flash is an understanding engine with an enormous window: transcribe, summarise, reason across hours of media, call tools, and hand back text. LFM3-Omni-2B is a conversational endpoint that happens to see and hear: it takes a question by voice, looks at a frame, and answers in a voice, on a device.

The consequence is that building a voice assistant on Qwen3.8-Omni-Flash means building two more stages — speech recognition on the way in, synthesis on the way out — and paying for three models where Liquid charges you for one. Building a two-hour meeting analyser on LFM3-Omni-2B means feeding it in chunks, because a 32K window is not an hour of audio and Liquid does not publish an audio-to-token rate that would let you calculate the limit before trying it.

Context is the widest gap on the sheet

• Context — Qwen3.8-Omni-Flash 1M tokens, with a maximum input around 991K and output up to 131K; LFM3-Omni-2B 32K tokens covering input and output together.

• Media limits — Alibaba documents up to three hours of audio and two hours of video in a single request, with audio input in 113 languages and dialects; Liquid documents 16 kHz mono audio and video sampled at 2 frames per second, 4 to 384 frames.

• Output — text only against text, 24 kHz speech, or interleaved text and speech.

• Scale — a 125B mixture of experts with about 6B active, hosted, against 2.17B dense running on a phone.

• Weights — closed against open, under the LFM Open License v1.0.

• Price — $0.15 per million input tokens, $0.016 cached, $0.47 per million output for Qwen3.8-Omni-Flash; no per-token price for LFM3-Omni-2B because there is nothing to meter.

The 1M context is not a spec-sheet ornament. It is the difference between "summarise this meeting" and "summarise this meeting, then answer questions about the third hour". Alibaba's launch framing was a bill rather than a score: 98% cheaper per hour of audio input than Qwen3.5-Omni-Plus, 93% cheaper per hour of audio-video, figures derived by pricing a two-minute sample and multiplying. The savings are real and they are also a vendor's own arithmetic.

What the money actually buys

At $0.15 in and $0.47 out, an hour of transcription is cheap enough that most teams stop modelling it. That is the point of the price cut, and it is the strongest argument for renting: the marginal cost of a request is near zero and somebody else owns the GPUs.

LFM3-Omni-2B inverts the curve. The weights are free, the marginal cost per request is your user's battery, and the fixed cost is the engineering time to ship a 2.17B model into an app. If you already have a mobile team, that is not a large number. If you do not, it is the whole project.

This is also where a router earns its place rather than being a slogan. The Qwen3.8 family's text models are already on OrcaRouter's catalogue at list price with 0% markup, which means a vendor price cut is live on our side the same day it lands; Qwen3.8-Omni-Flash itself is not on the catalogue, and neither is LFM3-Omni-2B. What the catalogue does give you is the ability to keep the hosted half of a hybrid — the long-audio batch job, the fallback when a phone is offline — behind one key and one invoice, with automatic failover if a provider goes down.

Benchmarks: one number in common, two harnesses

RealWorldQA is the only benchmark both cards report, and the gap is 21 points: Qwen3.8-Omni-Flash at 87.7, LFM3-Omni-2B at 66.54 in vision mode and 64.68 when the question is spoken.

Do not read that as a like-for-like defeat. Qwen's 87.7 was measured by Alibaba on a 125B model with a reasoning budget; Liquid's numbers were measured by Liquid with greedy decoding and each model's own chat template, and its spoken variant feeds the question as audio rather than text. Same benchmark name, different conditions, different model classes. The directional conclusion is safe — a model twenty times larger, with thinking enabled, should be better at visual question answering — but the magnitude is not a measurement of anything.

The rest of each card is unshared. Qwen3.8-Omni-Flash reports LiveCodeBench v6 at 92.6, GPQA Diamond at 91, SpotSoundBench 67.2, WildClawBench-MM 71.0 and OmniVideoBench 67.8. LFM3-Omni-2B reports Daily-Omni 70.09, WorldSense 47.10, OmniBench 46.32, OmniBench speech 48.25, MMAR 51.40 and a tool-calling score of 52.20. The overlap between those lists is close to empty, which is what you would expect from two models built for different jobs — and a reminder that neither vendor has published an independent evaluation.

Latency, and the round trip nobody prices

A hosted omni model has a network round trip in front of every answer and a queue behind it. For batch work that is invisible. For a conversation it is the difference between a product that feels alive and one that feels like a form submission.

LFM3-Omni-2B's advantage here is structural rather than measured: the audio never leaves the device, the model is 2.17B, and the output is speech tokens generated in the same pass as the text. There is no ASR call, no synthesis call, and no orchestration between them. Whether that beats a well-provisioned hosted endpoint on wall-clock latency is an empirical question Liquid has not answered — it publishes no latency figures, and there is no independent test.

The honest statement is that the local path removes three failure modes (upload, queue, download) and adds one (a phone that is thermally throttling).

Who should pick which

Pick Qwen3.8-Omni-Flash if your input is long, your output is text, and your users are not waiting on a voice. Meeting transcription, media monitoring, call-centre analytics, video search, anything that benefits from a million-token window and a reasoning budget — this is the model class built for it, and the price makes the decision easy. Accept that you are renting, that the weights will never be yours, and that a spoken interface means adding models rather than configuring one.

Pick LFM3-Omni-2B if the interaction is spoken and the device is the product. It is smaller, it sees less, it remembers 32K instead of a million, and its licence has a revenue clause that Apache 2.0 does not. In exchange, it is the only one of these two that can end a sentence with a voice.

That is the whole trade, and it is worth saying plainly because most comparisons of these two models spend their length on scores that only one of them was ever trying to win. If your product talks, one of these models was built for you and the other was not.