Ming-3.0-audio-flash vs Gemini 3.8 Live: Open Weights, Eight GPUs, and a Bill You Can Actually Calculate

One hour of continuous audio into G​emini 3.8 Live costs $0.30 at G​oogle's published list price. One hour of audio back out costs $1.08. A two-way voice session, both directions running for the full hour, is $1.38. Ming-3.0-audio-flash has no per-minute price at all, because there is no endpoint to buy: it is a 126-billion-parameter mixture-of-experts checkpoint under MIT, and its cost is whatever eight H800s cost you for the hour. That is the entire comparison in two sentences, and almost every other difference between these two systems follows from it. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

The dating matters, because this is not a launch. inclusionAI put the Ming-3.0-audio-flash weights on Hugging Face on September 3, 2026 and said nothing about it. For six weeks the repository was effectively a folder of shards: no serving code, no examples, no usable card. The GitHub repository was created on October 15, the vLLM-Omni and SGLang profiles, the voice presets and a rewritten model card landed between October 15 and October 19, and as of this writing the Hugging Face listing shows 22 downloads. G​emini 3.8 Live, by contrast, is a stable, generally available Live API model that G​oogle describes as the default for low-latency voice agents. One of these is a product you can call this afternoon. The other became runnable this week and still has no operator.

What the two systems actually are

G​emini 3.8 Live is a hosted streaming model behind the Live API. The API accepts raw 16-bit PCM audio at 16 kHz, images at up to one frame per second, and text, and returns raw 24 kHz PCM audio. G​oogle's Live API documentation puts language coverage at 70 languages. You open a session, stream, and get audio back; there is no checkpoint, no GPU reservation, and no licence to read.

Ming-3.0-audio-flash is an audio-language model in the more literal sense. An audio encoder and projector feed a mixture-of-experts language backbone, and an optional Talker turns the model's text output into speech. Inputs are audio plus a text instruction. Outputs are text, or text with synthesized speech. The card lists Chinese and English as primary languages, "including varied accents and dialects." The licence is MIT.

• Parameters — Ming: 126B total, 6.5B activated per inference step (the Hugging Face sidebar separately says 128B, a discrepancy the card does not explain). G​emini 3.8 Live: not disclosed.

• Languages — Ming: Chinese and English. G​emini 3.8 Live: 70 languages per the Live API docs.

• Input — Ming: audio and a text instruction. G​emini 3.8 Live: audio, images, text.

• Output — Ming: text, optionally speech. G​emini 3.8 Live: audio, and text through the same session.

• Licence — Ming: MIT on weights and code. G​emini 3.8 Live: a commercial API agreement.

• Price — Ming: no rate card; you pay for GPUs. G​emini 3.8 Live: $3.00 per million audio input tokens or $0.005 per minute; $12.00 per million audio output tokens or $0.018 per minute; text at $0.75 in and $4.50 out per million.

The hour-of-audio calculation, done properly

The per-minute figures above are the ones worth internalising, and they are also the ones most comparisons get wrong by quoting token prices at a reader who has no idea how many tokens a minute of speech burns. Use the minute rates instead, because G​oogle publishes them directly.

• Listening only, one hour — 60 minutes at $0.005 is $0.30.

• Speaking only, one hour — 60 minutes at $0.018 is $1.08.

• A full two-way hour — $1.38, before any text tokens or tool calls.

• A thousand two-way hours — $1,380.

Now the same hour on Ming-3.0-audio-flash. The reference serving profile for the full 128K context with speech output is tensor parallelism across eight 80 GB H800 GPUs, per the repository's own deployment notes. Four GPUs run a validated SGLang profile, but that profile is text output only. So the honest self-hosted comparison is eight GPUs against a metered session. At a hypothetical $2 per GPU-hour, one hour of capacity costs $16 — and that is for the whole GPU, which you can amortise across concurrent requests rather than one conversation. At ten concurrent streams, the per-conversation cost falls under $1.60 an hour; at a hundred, under $0.16. Those are arithmetic, not measurements, and they ignore utilisation, batching efficiency and the fact that a 126B model at 6.5B active parameters will not saturate eight H800s on short audio clips.

The crossover is real but it is not where people assume. Below roughly a dozen concurrent conversations, G​emini 3.8 Live is cheaper than renting eight H800s and vastly cheaper than buying them. Above that, and assuming you can keep the batch full, the open checkpoint wins on raw compute cost — which is the entire economic argument for open weights, and the reason the eight-GPU requirement is the number to remember rather than the parameter count.

What "audio understanding" means on each side

G​emini 3.8 Live is built for conversation. Its job is to hear you, answer quickly, and handle the turn-taking. G​oogle's model list positions it as the default Live model "for most low-latency voice agent experiences without reasoning delays," which is a statement about latency budget, not about depth of audio analysis.

Ming-3.0-audio-flash is pointed at a different task. The card's five highlights are complex speech recognition across dialects, accents, noise and multi-speaker recordings; long-audio understanding; perception of non-speech sound, acoustic scenes and music; reasoning over audio content; and spoken interaction with optional synthesized replies. That is a transcription-and-analysis list with a chat mode bolted on, not a voice-agent list.

The practical consequence: if your product is a phone agent that must respond in 400 milliseconds, G​emini 3.8 Live is the shape of the thing and Ming-3.0-audio-flash is not, regardless of benchmark scores. If your product is "ingest a 90-minute earnings call, tell me who said what, and flag the moment the CFO's tone changed," the Ming card is describing your job and the Live API is describing a different one.

The benchmark problem nobody has solved

Ming-3.0-audio-flash ships three benchmark panels in its repository: speech recognition, audio intelligence, and text. All values are percentages, higher is better unless marked with a downward arrow, and results marked with an asterisk are the team's own evaluation. The repository adds a note that comparison-model results were reproduced under inclusionAI's own evaluation setup and may differ from those models' official reports because of prompts, decoding settings, model versions and pipelines.

That note is unusually candid and it also means the charts cannot be used to rank Ming against G​emini 3.8 Live. The numbers live in images, not in the README text, and G​oogle publishes no comparable audio-understanding suite for 3.8 Live. There is no third-party leaderboard that has scored both. Anyone showing you a bar chart with these two models side by side has built it from vendor claims on one axis and nothing on the other.

The one honest statement available today: Ming-3.0-audio-flash has vendor-run numbers and no independent ones, G​emini 3.8 Live has a production track record and no published audio-reasoning benchmark. That is a tie in the only category that matters for a purchasing decision, which is evidence.

Where OrcaRouter fits, and where it does not

Ming-3.0-audio-flash is not on our platform. We are not going to pretend otherwise: if you want to run it, you clone the repository, install the vLLM-Omni or SGLang branch it names, and bring the GPUs. What a router is good for in this comparison is the other side of the decision — the part where you have already decided that a hosted audio model handles your live traffic, and the question becomes which hosted model, at what price, with what fallback when a session fails. That is a routing problem across more than 200 models behind one key, at provider list price with no markup added, so a vendor price change reaches you the same day rather than at the next contract renewal.

For a team testing whether an open audio checkpoint deserves production traffic, the useful pattern is to keep the hosted path as the fallback and move volume onto self-hosted Ming gradually. Automatic failover is what makes that migration reversible instead of a bet.

Choosing

Take G​emini 3.8 Live if you need a voice agent in production this quarter, need more than two languages, need images in the same stream, or expect fewer than a dozen concurrent conversations. The per-minute bill is small, the operational surface is zero, and the model is stable and documented.

Take Ming-3.0-audio-flash if you need to own the weights, need to run on-premises or air-gapped, need MIT terms, are working primarily in Chinese and English, and have eight H800s or a scheduler that can give you them. Expect to do your own evaluation, because nobody has done it for you. Expect the toolchain to be days old, because it is. And expect the model card to keep changing under you: its last commit was the day this was written.

There is a third case, and it is the largest one. If you do not yet know which of those two you are, you are not ready to choose between a metered API and an eight-GPU deployment. Run a hundred hours of your own audio through a hosted model first, measure what accuracy you actually need, and let that number decide whether the open checkpoint is worth the rack.

What to watch

The thing that would change this comparison is not a benchmark. It is an operator. The moment a serving provider lists Ming-3.0-audio-flash behind an OpenAI-compatible endpoint with a per-hour price, the eight-GPU argument stops being the reader's problem and the $1.38 hour gets a competitor. Until then, the honest summary is that G​oogle sells a minute and inclusionAI gives away a machine, and the two are only comparable if you already own the machine.