Frontline Hotspot
Frontline Hotspot

Hy Realtime Lands in Yuanbao, but the Cloud API Is Still Coming

Checks the official wording and channel timing of Tencent Hunyuan Hy Realtime, released on November 13, 2026: voice calls in Yuanbao and Tencent Customer Service are already live, while the Tencent Cloud API remains at the official will-soon stage with no public interface documentation, endpoints, or pricing. It stops at the web-search decision, the six song-request dimensions from the TME Co-Design, and the two entry layers of the Yuanbao voice preset and the partial web demo. Full-duplex interaction and the other two items remain on the roadmap. Internal A/B Test and chart figures are not turned into third-party results.

Published November 15, 20269 min read
<!-- hy-realtime-voice-model-hotspot | hotspot | Hy Realtime Lands in Yuanbao, but the Cloud API Is Still Coming -->

As of November 14, 2026, the easiest mistake to make when talking about Tencent Hunyuan's Hy Realtime is to compress two facts with different time status into a single sentence. This real-time voice model genuinely is running inside consumer products already, while the Cloud API that would expose it on Tencent Cloud remains at the stage the official wording calls "soon": it is still coming, and it has not opened. On one side there is the statement that the model has already been deployed in Yuanbao voice calls, Tencent customer service, and other products; on the other side there is the statement that it will soon be available on Tencent Cloud to provide API services. The first is past and present; the second is future. Reading those two clauses separately matters more than memorising any single capability name. This piece checks only the official release wording and the tense of each channel. It does not supply conclusions on the vendor's behalf, it does not score benchmarks, and it does not dress up internal evaluations as third-party tests. The contrast in the headline is not rhetoric; it is the factual boundary the whole article has to hold: Yuanbao has shipped, the Cloud API is still coming.

Publication and identity: three dates to keep apart

Tencent released the real-time voice model Hy Realtime on November 13, 2026. The channel deserves a note first. The primary release venue was an official WeChat account post; there was no separate standalone release page on a dedicated website. The verifiable pages on the cloud side are a web demo hosted on a Tencent Cloud subdomain and the Yuanbao download page. The first-hand source is a single account post rather than a product interface document, and that alone determines why much of what follows can only be stated qualitatively rather than pinned to parameters.

Three dates need to be kept apart. The first is the publication date, November 13, 2026, the day the official account made the model public. The second is the press date, also November 13, 2026, when a number of technology outlets and aggregator sites carried the news the same day. Those pickups show how fast the news travelled; they do not show when the model shipped, and a repost date must not be read backwards as a release date, still less as a feature-availability date. The third is the verification point of this article, As of November 14, 2026, the moment at which the wording above was checked line by line. Mixing these three dates is the most common source of distortion in fast-moving coverage: some accounts treat the repost date as the release date, some read "released" straight into "fully available," and some write about things that had not appeared by the verification point as though they should already exist. This article reconstructs the original wording as of the publication date, does not infer capability state from the press date, does not fetch extra numbers on its own, and does not convert quantities the vendor never gave.

The object also needs to be identified before anything else. Hy Realtime is a real-time voice model, not a general text large language model, and not the name of a chat assistant product. The official description sits firmly on the voice side: strong performance on basic conversational abilities such as voice understanding, instruction following, and human-like expression, plus newly added Agentic tool use and singing. It is presented as a foundation capability for real-time voice interaction, aimed at the loop of speaking, being understood, and replying, rather than at a text generation leaderboard. Getting that identity clear keeps the later word "deployed" from drifting: what has shipped is the model's capability already live inside several products, not a public interface open to every developer.

Capability boundaries: three basics, one tool chain, one singing scenario

The official description groups the basic abilities into three: voice understanding, instruction following, and human-like expression. Voice understanding concerns making sense of real spoken language; instruction following concerns remembering and keeping up with requirements across multiple turns; human-like expression concerns the tone and pacing of a reply. These are qualitative descriptions. The vendor says they are strong, but gives no verifiable per-item scores, so this article reports them as stated and converts them into no metric.

Of the two additions, Agentic tool use is the one most easily overstated. The official description anchors this capability first and foremost on one chain: judging, from the conversational context and the user's intent, whether to run a web search, and unifying the search decision, the tool call, and the answer generation. In other words, it describes how a single search decision is folded into generating a reply, and the boundary stops there. Stretching it into automatic transfers, placing orders, or replacing professional arrangement exceeds the original wording. This article writes only as far as the "whether to run a web search" chain and no further. Tool use here is one link in the answer-generation chain, not a permission list onto which arbitrary external actions may be hung.

The singing capability rests on deep Co-Design with the TME AI team at Tencent Music. The official interaction is combinatorial song requests across six dimensions: singer, album, genre, language, release year, and popularity. A user combines them to request a song, and can also continue a choice from earlier in the conversation. Six is the number to use; some retellings list only a subset, or shorten "release year" to "year," and both are imprecise. As for catalogue scope, whether it can be used commercially, and how licensing would work, the official material says nothing about outbound commercial authorisation, so this article makes no assertion on its behalf and does not read the song-request feature as a catalogue inventory. The ability to continue from earlier context suggests a multi-turn conversation design rather than one-off keyword matching, but it still sits inside the scenario of picking a song to sing.

On the engineering side the official disclosure is method names and structure, which can be unpacked, but no verifiable figures. Training is split into four stages, CT, SFT, OPD (also written MOPD in the text), and RL. There is a dual-teacher setup, with a Hunyuan text model and a Hunyuan voice Agent teacher participating together. The reward side is a dual track of a general preference reward model plus a rule list, backed by a bottom-line gate the vendor describes as having a "one-vote veto" role. For redundant retrieval, the official wording is "stepwise down-weighting." Low latency comes from three-stage training that "generates text and synthesises speech at the same time." These are restatements at the method level; this article infers no metric from them and supplies no evaluation numbers on the vendor's behalf. They help explain why the model is positioned as real-time, but they are not performance proof.

The experience entrances come in two layers. On mobile there is the latest Yuanbao app, where a dedicated voice preset is chosen inside the voice-call feature. On the web there is the demo on a Tencent Cloud subdomain, which the official wording describes as offering "a partial experience of some capabilities." The two are not equivalent: Yuanbao carries the full voice call, while the web demo opens only part of the capability set. Neither should the demo's behaviour be used to infer the full Yuanbao experience, nor should a capability already live inside Yuanbao be counted as a public interface.

The tense question: what has shipped and what is still coming

The line that deserves the closest reading is the deployment sentence. In the official wording, Hy Realtime "has already been deployed in products including Yuanbao voice calls and Tencent customer service, and will soon be available on Tencent Cloud to provide API services." The first half, "has already been deployed," is present perfect, and its landing points are concrete products: Yuanbao voice calls and Tencent customer service. The second half, "will soon be available," is future, and its landing points are Tencent Cloud and the provision of API services. Within one sentence something has happened and something has not, and the two must not both be read as "deployed" simply because they sit next to each other. The headline uses "shipped" against "still coming" precisely to put that half-clause difference in plain view.

That is why saying the Cloud API is still coming, and has not opened, is a judgement about the present state rather than a guess. The official wording is "will soon," gives no concrete date, and attaches no public interface documentation, endpoints, or pricing. At the verification point there is no public interface material a developer could plug into directly. "Coming to Tencent Cloud" names a channel on the roadmap, which is not the same thing as being callable by developers right now. What a reader can actually use is the voice call with that preset in the Yuanbao app and the partial capability experience in the web demo. What cannot be done yet is to pick up a Tencent Cloud interface document and start writing calls, checking endpoints, or comparing prices. Reading "soon" as "already," or treating a planned channel as a ready-to-call endpoint, smuggles future tense into past tense. To check whether a retelling has drifted, it is often enough to see which half of that sentence it rendered as completed: turning the future half into a finished one, inventing a date the vendor never gave, or filling in pricing that does not exist all cross the line.

For developers this is a practical boundary: as of the verification point, the sensible actions are to try the capabilities already live on the product side and watch for a formal Tencent Cloud availability notice, not to schedule an integration around a public interface document that does not exist.

The roadmap items also need to be marked off separately. The official text adds that Hy Realtime is still being iterated and that the next step will focus on improving full-duplex interaction, long-term personalised memory, and foreground and background Agent interaction. "The next step will" is an explicit planning statement, and all three belong to the roadmap rather than to delivered capabilities. "Full-duplex interaction" in particular is easily conflated with the existing full-duplex voice comparison on this site, but that comparison discusses a technical direction, not a feature Hy Realtime has shipped at this moment. Writing roadmap items as available now would repeat the tense error once more. The phrase "still being iterated" is itself a reminder that the current version is not the endpoint.

What can be checked and what cannot

Lay out what the vendor provided and what it did not, and the boundary becomes clear. What was provided is a capability description plus a self-reported evaluation framing: three basic conversational abilities, two newly added capabilities in tool use and singing, and the engineering methods above. On evaluation, the vendor says it used an internal expert A/B Test and an internally built basic intelligence test set, and states that its overall objective performance is at a leading level across several public benchmarks in voice interaction, without naming specific leaderboards or giving a complete score table. This article treats "leading level" qualitatively as the vendor's own claim; it does not convert that phrase into a ranking and does not compare across leaderboards.

What was not provided should be stated just as plainly. The internal expert A/B Test is a vendor self-assessment, not a third-party test, and must not be relabelled as an independent evaluation. Scores from charts, figures from internally built evaluations, third-party leaderboard conclusions, and the catalogue and commercial-licensing scope behind singing all lack verifiable public material, and this article introduces none of them. Those gaps should not be filled by a release explainer. What can be checked is the wording the vendor wrote in its first-hand channel and the tense of each channel; what cannot be checked is left blank. This article supplies no conclusions on the vendor's behalf, and it does not turn "the official page did not mention X" into "the vendor announced there is no X." What separates a vendor claim from a third-party test is precisely verifiable material, and that material is not on the public pages at this point.

Where this piece sits

This is a single-product release and channel-tense check, not a buying guide. The full-duplex voice model comparison and the cloud versus local voice AI selection guide on this site discuss contrasts between different models and deployment trade-offs, and their point is "how to choose." This piece looks at one object only, and its point is "what has shipped and what is still in the future tense." There is also a release article on another Tencent Hunyuan product line; that is a different product line, not the same real-time voice line as Hy Realtime, and the tenses of the two should not be read together. For comparisons or deployment trade-offs, see those dedicated articles; for what this one model has actually opened up right now, the third section above is enough.

Related reading

This article is AI-assisted and human-edited. Last updated: 2026-11-15

Related

Frontline Hotspot

One-Eighth the Cost of Opus 5: Can Step 5 Preview Deliver?

Per ai-bot.cn on 2026-09-20, StepFun released Step 5 Preview, a next-generation flagship base model: sparse MoE with 600B total parameters and only 27B activated per inference, a native 1M-token context with text and vision multimodality, designed for real-world agentic tasks. It scores 44 on the Artificial Analysis Intelligence Index, top three among open models, with a claimed per-task cost one-eighth that of Claude Opus 5, GPU kernel optimization at 508 TFLOPS against 493, and a 22-hour continuous autonomous agent run. The article labels its sources honestly: the figures come from a vendor comparison table, per-token API pricing and stability remain Preview-stage unknowns, and weights are only promised for 2026-10-15.

Sep 20, 20267 min read
Frontline Hotspot

GPT-Live-1 API: real-time voice signals and a cold look

OpenAI shipped the real-time speech model GPT-Live-1 to API on 2026-09-11: full-duplex dialogue (simultaneous speech in and out), handling interruptions, pauses and background noise, aimed at phone voice agents such as restaurant booking and customer service; the model fuses speech understanding and generation in one network to cut latency, and offloads complex reasoning to a backend text model. This piece breaks down each release claim, reads the two-part pattern of "fused understanding and generation" plus "a real-time speech shell around a strong reasoning core" (echoing the 9-10 ChatGPT voice-mode hand-off to GPT-5.6 Sol / GPT-6 Astra), flattens the traditional IVR / ASR+NLU pipeline into a comparison table, and closes with cold takes: quota cost includes backend model hand-offs, Chinese multi-dialect robustness is unverified, the cloud-versus-local boundary, and vendor-claim caveats. Note that GPT-Live-1 is a closed-source API model with no public code repository.

Sep 13, 20269 min read
Frontline Hotspot

DeepSeek App APIs: Five Tools, Fees on Top of Tokens

Only what the DeepSeek November 13, 2026 changelog, guides and pricing page actually say about the full suite of application APIs: web search, deep-thinking search, file input, URL reading and model safety are now exposed on the cloud API, while Search, Fetch, OCR, ASR and TTS remain standalone APIs. Tool calls and context assembly happen server-side. Fees are quoted verbatim as tool usage fees in addition to standard token charges; no price conversion, no cross-vendor comparison, no invented discount end date, and the chat-demo repository and MIT license are described as published.

Nov 16, 20269 min read