Language Models

Tencent's Hunyuan Team Answers Developers: Next-Gen Hy Will Be Bigger, Hy4-mini and a 24GB-VRAM Model Are on the Roadmap — But No Dates

November 22, 20260 views阅读中文原文
Tencent's Hunyuan Team Answers Developers: Next-Gen Hy Will Be Bigger, Hy4-mini and a 24GB-VRAM Model Are on the Roadmap — But No Dates

The short version: In the early hours of November 21, Tencent chief AI scientist Shunyu Yao and the Hunyuan team held a public AMA on Reddit's r/LocalLLaMA. No new model launched. What they gave was a roadmap briefing: the next Hy generation gets substantially bigger, smaller and locally-runnable variants are all planned, and every timeline is "stay tuned" with not a single date attached.

A note on sourcing first: This article is compiled from multiple media reports. The original Reddit thread could not be fetched directly during verification (it returned 403), so every quote here comes from on-site write-ups published on the morning of November 21 by Zhidongxi, Sina Tech, and NetEase. Zhidongxi was one of the outlets actually asking questions in the thread.

What this AMA actually was

Put simply: developers asked questions, the model team answered them in public. The Hunyuan team had just shipped two models back to back — the flagship Hy4 on November 4, followed by the cost-focused Hy4 flash. According to figures Tencent provided to media, Hy4 ranks second among Chinese models on Artificial Analysis's intelligence index and first among open-weight models on the Agent Arena leaderboard; Hy4 flash brings per-task cost down to roughly one-twentieth of Hy4's.

With two fresh models on the shelf, the questions converged on three things: will the next version be bigger, when do the small models arrive, and can I run this locally. For ordinary users those three questions matter more than any benchmark table — they decide whether you keep renting an API next year or finally put the GPU you already own to work.

Upward: the next Hy generation will scale up "significantly"

The team's wording was that the next Hy model's scale will "definitely expand significantly further," and that they will keep investing heavily in large-scale reinforcement learning.

But when a developer put a number on it — asking whether Hy4.5 would reach 1.5 to 2 trillion parameters — the team did not confirm. That detail matters: the direction is stated, the number is not. Any retelling that presents "1.5T/2T" as a settled spec has overread the exchange.

On rivals such as ByteDance and Kimi pushing toward 3T or even 5T models, the team's answer was that a smaller model trained solidly can outperform a larger one, at lower cost. That is consistent with their own position: the Hy line still sits around 700B.

Downward: Hy4-mini, Flash-Lite, and 24B–31B dense models

This is where the substantive news is. Item by item:

  • Hy4-mini: already on the development roadmap. The team's reply was an unambiguous yes.
  • Flash-Lite: explicitly teased, positioned for low-latency, low-overhead work such as customer-facing chatbots, document parsing, classification, and simple automation.
  • 24B–31B dense models: under consideration, not yet committed.
  • Smaller image models: the team will consider an open image model in the 5B–12B range for local use, and said to watch for a next-generation video model.
  • Local deployment: confirmed a model aimed at 24GB-VRAM devices, and continued work on versions for machines with 64GB–128GB of memory. They also said yes to an open-weight model at roughly current scale that fits a 128GB machine after Q4 quantization.
  • On-device vision: the team believes a model at roughly 3B active parameters (A3B) is already capable of reading screenshots and charts and reasoning about them, not just OCR.

Put together, the strategy is legible: push the ceiling with the flagship while building a whole ladder of small models for the "runs on my own machine" market. That second part is what the open-source community has actually been waiting for.

Why the architecture trade-offs went this way: how it works

The AMA was the first time the team went into real detail on Hy4's architecture decisions, and that part has practical value for anyone deploying inference.

Between GQA and MLA attention, the team weighed the options for a long time. MLA delivers better model quality — beating GQA at the same KV cache size — but its compute intensity is markedly higher, and speculative decoding on inference GPUs runs straight into a compute bottleneck. What finally let them adopt MLA was DSA, which cut MLA's compute cost during decoding.

Hy4 flash took a different route: hybrid sliding-window attention (SWA) shrinks the KV cache to a quarter of its previous size, cutting serving cost substantially. The team also introduced components including LatentMoE and Engram. On MoE sparsity, their view is that sparser is not automatically better: an overly sparse mixture-of-experts model eats VRAM without fully using the GPU's compute.

Three weaknesses the team admitted

This section is arguably the most telling, because the team volunteered all of it.

One: overthinking. The team said the hardest behavior to fix in post-training is overthinking. Hy4 targets maximum intelligence, so long chains of thought are a known problem; on Hy4 flash they will use reinforcement learning to shorten them.

Two: over-refusal. On cybersecurity training, the team kept only data related to vulnerability reproduction and remediation, stripping out exploitation and reverse engineering. The trade-off is stronger vulnerability discovery and comparatively weaker exploitation, and they acknowledged that reducing unnecessary refusals remains an important area to improve.

Three: recursive self-improvement is early. Asked whether Hy4 could develop its successor through RSI, the team said the capability is still early, that they have several internal benchmarks, and that the current priority is fundamentals. On the AGI capability scale, one team member's personal estimate put Hy4 at "about level 2.5."

On ecosystem support, the team said llama.cpp support is in progress and that they are working with the vLLM and SGLang communities.

Cost: one anchor worth remembering

Hy4 flash has 224B total parameters including 64B of Engram parameters, activating only 12B per token. According to Tencent's figures, on its internal WorkIndex benchmark for everyday work tasks, Hy4 flash costs about 0.26 yuan per task — one-twentieth of Hy4 — based on a blind test with 85 internal users across 259 work tasks.

Mind the boundaries of that number: it is a Tencent-built benchmark run blind with internal users, not an independent reproduction. Treat it as an order-of-magnitude reference, not an absolute cross-vendor comparison.

Who this is for — and when not to bother

  • People who want to run models on their own hardware: both the 24GB-VRAM and 64–128GB-memory lines got a positive answer. Worth waiting for.
  • Anyone building cost-sensitive products: a per-task cost around 0.26 yuan determines which use cases can actually be commercialized.
  • Teams choosing an open base model: Hy4 flash already ships open weights, and the small-model ladder will make the product range more complete.
  • Those watching the Chinese model race: this is a first-hand statement of how a major lab is thinking, more concrete than any launch keynote.

Who it is not for: when not to bother

  • Anyone hunting for a release date: there isn't one anywhere. The next Hy, Hy4-mini, and Flash-Lite are all "stay tuned."
  • Anyone wanting confirmed parameter counts: 1.5T–2T was a developer's question, not the team's answer.
  • Anyone looking for independent evaluation: the benchmarks and cost figures come from Tencent or from media retellings; there is no new third-party test here.

Pitfalls: three easy ways to misread this

  1. "Next-gen at 1.5T–2T" is a question, not an answer. In the original exchange that figure came from a developer. The team only said the next generation will expand significantly. Presenting it as a spec is the most common error in circulation.
  2. A roadmap is not a schedule. The three items differ in maturity: Hy4-mini is "on the roadmap," Flash-Lite is "coming," and 24B–31B dense is merely "under consideration."
  3. This was not a model launch. It was a follow-up Q&A after Hy4 and Hy4 flash. The actual releases happened on November 4 and with flash afterward.

Where to start

  • First step: if you are waiting on local deployment, plan around the 24GB-VRAM line — that is the direction the team explicitly confirmed. Treat 24B–31B dense as a possibility only.
  • Second step: what you can do today is take the already-open Hy4 flash weights, run them on vLLM or SGLang against your own task set, and replace that 0.26 yuan figure with your own real number.
  • Third step: put llama.cpp support on your watch list. The team says it is in progress, and once it lands the barrier to local inference drops noticeably.
  • Fourth step: when the next Hy arrives, look at post-training and RL gains rather than parameter count alone — the team itself said Hy4 has substantial room to improve in post-training.

Sources

  • Zhidongxi / ifeng: https://tech.ifeng.com/c/8yE5loSnUps
  • Sina Tech: https://finance.sina.com.cn/tech/roll/2026-11-21/doc-iniznhsh6798925.shtml
  • NetEase: https://www.163.com/dy/article/LC8ODKI90534A4SC.html
  • Original Reddit AMA thread (not directly reachable during verification): https://www.reddit.com/r/LocalLLaMA/comments/1y15w4a/tencent_hy_team_shunyu_yao_here_ask_us_anything/

Sources

Comments

0 published
Log inorsign up to leave a comment.

Loading comments...