Qwen3.8-LiveTranslate vs Qwen 3.8: A Naming Collision, Not a Fair Fight

Search for one of these models and you will be shown the other. Qwen3.8-LiveTranslate, released September 19, 2026, and Qwen3.8-Max — the model most people mean when they write "Qwen 3.8" with no suffix, generally available since August 3, 2026 and refreshed as Qwen3.8-Max-0902 on September 2 — share a generation prefix and almost nothing else. One is a simultaneous interpretation system that takes audio in and returns translated audio and text out, billed over a WebSocket under the ID qwen3.8-livetranslate-flash-realtime. The other is a 2.4-trillion-parameter sparse mixture-of-experts generalist with a 1M-token context window that cannot hear anything at all. Putting them head-to-head is a category error, and the fact that so many people arrive at that error is the actual story here: the vendor now ships at least four distinct things under the Qwen 3.8 name, and the naming scheme does not tell you which one you want.

What each one actually is

The Qwen 3.8 generation shipped in layers through the second half of 2026, and the layers are not interchangeable.

Qwen3.8-Max is the hosted flagship: a sparse MoE model with roughly 2.4 trillion total parameters and about 95 billion active per forward pass, a 1-million-token context, and text, image and video input with text-only output. It is priced at $2.00 per million input tokens and $6.00 per million output tokens, with cache reads at $0.250 per million. Vendor-reported scores put it at 86.6 on Terminal-Bench 2.1, 92.6 on GPQA Diamond, 67.7 on SWE-bench Pro and 43.6 on Humanity's Last Exam.

Qwen3.8-2.4T-A95B is the same generation's open-weight release, published August 12, 2026: 512 routed experts across 92 layers, 69 of those layers using linear attention, and roughly 4.89 TB of weights. It is the model most people mean by "Qwen 3.8" when they are talking about running something themselves rather than calling an API.

Qwen3.8-27B is the small open checkpoint in the same family, and Qwen3.8-LiveTranslate is the specialist — an Interleave-architecture interpreter built on a Hybrid MoE Thinker–Talker design, with 60 source languages recognized and 29 available for spoken output.

The axis that separates them is not size. It is modality. Qwen3.8-Max has no audio input path and no audio output at all. Asking it to interpret a live conversation is not a matter of prompting it better; the capability is not in the model.

The spec contrast

Laid side by side, the two models barely overlap on any dimension a buyer cares about:

• Primary job — Qwen3.8-LiveTranslate: simultaneous speech-to-speech interpretation. Qwen3.8-Max: general reasoning, coding, agentic and multimodal text work.

• Input modality — Qwen3.8-LiveTranslate: audio and image. Qwen3.8-Max: text, image and video.

• Output modality — Qwen3.8-LiveTranslate: text and synthesized audio. Qwen3.8-Max: text only.

• Context window — Qwen3.8-LiveTranslate: 53,248 tokens (49,152 in / 4,096 out). Qwen3.8-Max: 1,000,000 tokens.

• Weights — Qwen3.8-LiveTranslate: closed, API-only. Qwen3.8-Max: hosted, with the open sibling Qwen3.8-2.4T-A95B published under the same generation.

• Rate limits — Qwen3.8-LiveTranslate: 10 requests and 100,000 tokens per minute. Qwen3.8-Max: standard hosted throughput, no per-request realtime ceiling.

The context disparity is the one that surprises people. Qwen3.8-Max holds a million tokens; the interpreter holds about five percent of that. But a realtime interpreter does not need a long context — it needs a short memory and a very fast one. Its 49,152-token input ceiling is sized for carrying the last several minutes of a conversation forward to keep names and terminology consistent, which is exactly the job "long-context disambiguation" does. Judging the interpreter by its context number is like judging a racing engine by its fuel tank.

Why the price comparison is a trap

On a per-million-token basis, the interpreter looks dramatically more expensive than the flagship: $7.50 per million audio input tokens against $2.00 per million text input tokens, and $30.00 per million audio output against $6.00 per million text output.

That comparison is meaningless, and it is the single most common mistake made about this class of model. Audio and text tokens are not the same unit. Speech is encoded into audio tokens at a density that has no relationship to the same content written down — a minute of conversation consumes vastly more audio tokens than its transcript consumes text tokens, and the output audio adds a second stream on top. Rendering the two rates side by side implies a cost ratio that does not exist in practice.

The structural difference matters more than the price difference. Qwen bills the interpreter per token, while much of the realtime speech market bills per minute of session. Those two models of pricing behave completely differently as your audio gets quieter, your sessions get shorter, or your speakers pause — a per-minute meter charges for silence, and a per-token meter does not. If you are building a cost model, the question is not which rate is lower; it is which meter punishes your usage shape. And note the region split: Beijing pricing runs ¥40 / ¥3.3 / ¥100 / ¥160 per million for audio input, image input, text output and audio output respectively, materially below the Singapore rates, which is the kind of difference that only becomes visible when you compare the two line by line.

One further point in the interpreter's favour on cost: Qwen kept these rates essentially flat against Qwen3.5-LiveTranslate, with Beijing unchanged and Singapore slightly cheaper. A new generation arriving without a price increase is not the norm in this market.

Open weights versus API-only is the real dividing line

If you are choosing between these two on principle rather than on capability, the deciding fact is licensing.

Qwen3.8-Max is available hosted, and the generation's Max-class weights were open-sourced as Qwen3.8-2.4T-A95B in August. That path exists, but it is not a casual one: 4.89 TB of weights is datacentre hardware, and the open-weight licence requires organisations earning above $50 million in a twelve-month period to obtain a commercial licence from Alibaba Cloud.

Qwen3.8-LiveTranslate has no such path. It is closed-weight and API-only, reachable through a WebSocket Realtime API that requires target_language to be set in a session.update event before any audio moves, supports no function calling, no structured output, no context caching, no batch inference and no fine-tuning. If your requirement is running the interpreter on your own metal, this model does not satisfy it and no amount of engineering will change that.

That matters for a specific kind of buyer: the one who cannot send audio to a third party — a hospital, a law firm, a government contractor. For everyone else, the practical question is whether the interpreter is good enough to justify the dependency, and the answer to that is a measurement Qwen has not yet allowed anyone to make.

Picking by job, not by name

The correct answer to "Qwen3.8-LiveTranslate or Qwen 3.8" is that you probably need the answer to a different question.

• If the job is live interpretation — a meeting, a call, a broadcast — only one of these can do it, and it is not the flagship.

• If the job is summarizing what was said, extracting action items, answering questions about the transcript, or writing the follow-up — only the flagship can do it, and it is not the interpreter.

• If the job is both, you are building a pipeline, and you will be paying two vendors with two contracts and two sets of credentials unless you consolidate deliberately.

That third case is the common one, and it is where running both halves behind a single endpoint stops being a convenience and starts being architecture. Qwen3.8-Max is on OrcaRouter at the provider's list price of $2.00 per million input and $6.00 per million output with zero markup passed through, so a Qwen price cut reaches your bill the same day rather than at the next contract renewal, and automatic failover means the summarizer half of the pipeline does not go dark when one provider has a bad hour.

To be explicit about the other half, because the distinction matters: Qwen3.8-LiveTranslate is not on our catalogue. It is available through Alibaba's own Model Studio API and the vendor's trial endpoint at omni.qwen.ai/live-translate, and we are not going to imply we route a model we do not. What you can put behind one key today is the downstream stack — the models that consume what the interpreter produces.

The naming problem is the actual finding

Four models, one prefix, no suffix convention that a reader can rely on. "Qwen 3.8" resolves to the hosted flagship, the open 2.4T checkpoint, the 27B small model, or the interpreter depending on who is typing and what they read last. The published benchmark tables do not help, because the flagship has them and the interpreter mostly does not: Qwen3.8-Max can be placed on Terminal-Bench or GPQA Diamond, while Qwen3.8-LiveTranslate's evidence is average lagging of 2.3 seconds, a vendor-reported 85.7 xCOMET-XXL on FLEURS, and a self-reported 9.7% diarization error rate with no independent replication.

So the comparison this page exists to answer is really a warning. Before you compare Qwen 3.8 to anything, establish which Qwen 3.8 you mean. If it takes audio, it is the interpreter, and it is a specialist you buy for one job. If it takes video, it is the flagship, and it is a generalist you buy for the other twenty. Any evaluation that mixes them will produce a conclusion about neither.