Tencent Hy4 preview: 770B open weights, 1M context, Apache 2.0

Tencent open-sourced Hy4 preview on August 28, 2026 with a launch post that ended "Use it. Tell us what breaks." The headline numbers: 770B total parameters, 49B active, a 1M-token context window, and Apache 2.0 weights. The number that actually moves adoption: $0.834 per million input tokens.

The most striking thing is the pace. Hy3 shipped on July 6 at 295B total parameters, 21B active, and a 256K context. Fifty-three days later Tencent released a model 2.6x larger with 4x the context. Whatever else is true about this release, the iteration speed in Chinese open weights has not slowed.

TL;DR

The architecture, and what's borrowed

Hy4 preview is a sparse MoE, and the config is worth reading closely because several choices are lifted from elsewhere in the open-weight ecosystem.

Two things stand out. First, the attention is explicitly DeepSeek's sparse attention design with an IndexCache added on top so sparse indices are reused across layers rather than recomputed — that is what makes a 1M context economically serveable at all, and it is a direct borrowing from the lab whose pricing has set the floor for this entire market. Second, the MTP layer means speculative decoding is shipped in the weights rather than bolted on by the serving stack — vLLM and SGLang both enable it in Tencent's prebuilt images.

Tencent's own framing of the jump: "We scaled Hy4 preview on three fronts: model size, context length, and training data." No claimed algorithmic breakthrough. This is a scaling release.

Benchmarks

The published scores, as reported on the model card and aggregators:

Those are frontier-adjacent numbers for an Apache 2.0 model, and the agentic column — Terminal-Bench 2.1 at 85.4 and MCP-Atlas at 83.7 — is the one worth caring about if you run tool-calling loops rather than chat.

The comparison claim is weaker than the headline suggests. Tencent ran a blind evaluation with 163 internal experts scoring 203 engineering tasks. Hy4 preview averaged 2.99 out of 4, against Kimi K3 at 2.94 and GLM-5.3 at 2.92. A 0.05-point margin, on tasks the vendor selected, graded by the vendor's own employees, is not a result — it is a statement that the three models are interchangeable on this workload. Read it as Hy4 reached the GLM 5.3 / Kimi K3 tier, which is a real achievement, and stop there.

Price, in context

This is the part that will decide whether Hy4 gets used.

Hy4 preview undercuts GLM-5.3 by roughly 40% on input and 43% on output, and undercuts Kimi K3 by 6x on output. Cached input at $0.042/M is aggressive — that is a 20x discount off the input rate, which matters enormously for agentic workloads that resend a large system prompt and a growing tool-call history on every turn.

DeepSeek V4 Pro remains cheaper, and no open-weight release this year has changed that. What Hy4 offers against it is the 1M context and the agentic benchmark profile. Chinese open weights now occupy nearly the whole price ladder below the Western frontier — the trend behind Asia-origin models crossing 60% of OpenRouter tokens.

Can you actually serve it?

Short answer: not on hardware you own, unless you own a lot of it.

770B parameters at one byte each is 770GB before you allocate a single token of KV cache. An 8xH100 node tops out at 640GB of VRAM, so the FP8 checkpoint does not fit — you need H200-class cards or multi-node tensor parallelism. This is the practical difference between Hy4 and the Hy3 GGUF quantizations that ran on a single GPU: Hy3 at 295B was squeezable, Hy4 at 770B is not, and no amount of llama.cpp cleverness makes a 770B model a laptop model.

For most teams that means the API, not the weights. The Apache 2.0 license still matters — it means no vendor can revoke your right to run it, and it means the model can be fine-tuned and redistributed — but "open weights" and "weights you can afford to serve" are different things at this scale.

Tencent recommends vLLM and SGLang for production, both with prebuilt Docker images that enable tensor parallelism and speculative decoding, and both exposing an OpenAI-compatible endpoint.

The caveats Tencent published themselves

This is the most credible part of the release, and it deserves quoting in full:

"This is an early version of Hy4. There is real headroom left in both pre-training and post-training, and we are shipping with known issues — among them, spending longer than necessary reasoning through complex tasks, and a tendency to over-verify its own work."

Vendors rarely name their model's failure mode on the model card. Take it seriously, because it compounds with a second number: 36 tokens/sec at P50 and 3.19s P50 latency as observed on OpenRouter, where Tencent Cloud is currently the sole provider. A model that over-reasons, at 36 tok/s, billed per output token, is a bad fit for a tight agent loop — the over-verification tax lands on your latency and your bill simultaneously.

Also worth knowing before you route production traffic: OpenRouter recorded roughly 86% availability over the model's first three days. That is a day-one capacity story, not necessarily a lasting one, but do not make Hy4 a single point of failure this week.

What people are asking

"Is it multimodal?" The release is a text model. Hunyuan ships separate image, video, and 3D models — including WorldClaw for agentic 3D world generation — but Hy4 preview is not a unified multimodal checkpoint.

"Why is nobody talking about it?" A fair question that appeared under the launch post itself. Partly release fatigue — this is the third major Chinese open-weight model in roughly six weeks. Partly the "preview" label, which signals to serious adopters that the checkpoint will be replaced.

"Should I switch from GLM 5.3?" Not on the benchmark claim, which is noise. Possibly on price, if your workload is output-heavy — the output rate is 43% lower. Definitely evaluate it if you need genuine long context, since GLM 5.3 does not offer 1M. Run your own eval on your own tasks; a vendor blind test with a 0.05 margin tells you nothing about your codebase.

"Is 1M context real or nominal?" Native, per the model card, supported by the sparse-attention plus IndexCache design rather than by rope-scaling a shorter model. That is a better sign than most 1M claims. It is still worth running a needle-in-haystack test at your actual working depth before trusting it.

Related reading on explainx.ai

• Tencent Hy3: 295B open-source MoE for agentic coding — the predecessor, 53 days earlier
• Hy3 GGUF at 1-bit and 4-bit on a single GPU — what Hy4's size takes off the table
• GLM 5.3 ties Kimi K3 on the Intelligence Index — the tier Hy4 just joined
• Kimi K3's open weights at 2.8 trillion parameters — the other trillion-scale MoE
• DeepSeek V4 Pro benchmarks and pricing — the price floor Hy4 did not undercut
• Asia-origin models cross 60% of OpenRouter tokens — the adoption trend this sits inside
• What is llama.cpp: running models locally — and why 770B is out of its reach
• Tencent Hunyuan WorldClaw: agentic 3D world generation — the rest of the Hunyuan lineup

Primary sources: the Hy4-preview model card on Hugging Face and Tencent's Hy4 preview research post.

Specs, pricing, and throughput figures are accurate as of August 28, 2026, the day of release. Preview checkpoints change; verify parameter counts, license text, and API rates on the model card before committing to a deployment.