Tencent Hy4 Preview Runs in vLLM From Day One: The 770B Open-Weight MoE You Can Actually Serve

When Tencent Hy4 Preview launched on August 28, 2026, it did something most frontier-class flagships skip on day one: it shipped runnable. Alongside the open weights, Tencent and the vLLM team published a prebuilt Docker image, an official serving recipe, and merged framework support in vLLM itself, so the largest open-weight model the Hunyuan line has produced — 770 billion total parameters, 49 billion active per token — was sitting behind an OpenAI-compatible endpoint on your own GPUs within hours of the announcement. That day-zero runtime story is the news this post is about, because "runs in vLLM" is not a footnote for a model this size. It is the difference between a press release and a thing you can put in production.

The rest of the launch is the usual preview-shaped package, and the caveats matter as much as the numbers. Tencent calls Tencent Hy4 Preview an early iteration, flags over-long reasoning and excessive self-verification as known behavior, and publishes a benchmark table that is entirely self-reported — no independent lab has scored it yet, and the model is two days old as of this writing. None of that changes the practical headline: the weights are Apache 2.0, the context window is 1 million tokens, and the day-one vLLM support means the self-hosting path is real rather than aspirational.

What Tencent Hy4 Preview actually is

Tencent positions Tencent Hy4 Preview as the successor to Hy3 (295B total, 256K context), roughly doubling active capacity and quadrupling the context window. The architecture is worth reading line by line, because each line changes what an open-weight model is allowed to do on your hardware.

• Architecture — Mixture-of-Experts with 770B total parameters and 49B active per token across 78 layers. The first layer is dense; the other 77 route each token through 256 routed experts plus one shared expert, activating the top 8. That roughly 16:1 sparsity ratio is what keeps inference cost in a range a serious team can actually run.

• Context window — 1,048,576 tokens native, one of the widest of any open-weight model. A full repository, a multi-file refactor, a long document batch: single-request problems.

• Attention — Gated Deep​Seek Sparse Attention (Gated DSA) with IndexCache, a sparse-attention design that computes a fresh sparse index on only 21 of the 78 layers and reuses it on the other 57. That is why serving it is not just "bigger vLLM" — it needs a backend that understands sparse attention.

• Built-in speculative decoding — an MTP (multi-token prediction) layer of about 10B total / 0.7B active parameters sits inside the model, so vLLM and SGLang can draft tokens without you hosting a separate draft model.

• Weights — Apache 2.0, in both BF16 and FP8 checkpoints, published to Hugging Face, ModelScope, GitCode, and CNB.

What "day-zero vLLM" actually means

Merged framework support on launch day is the concrete event behind the signal — the vLLM change that adds Tencent Hy4 Preview (PR #54160) landed in the same window as the release, and the official recipe targets vLLM 0.29.0 or newer. In practice that means the serving path is one command, not a week of kernel debugging.

docker run --gpus all -p 8000:8000 --ipc=host -v ~/.cache/huggingface:/root/.cache/huggingface vllm/vllm-openai:hy4-preview tencent/Hy4-preview-FP8 --tensor-parallel-size 8 --speculative-config '{"num_speculative_tokens":3,"method":"mtp"}' --attention-backend FLASHMLA_SPARSE --tool-call-parser hy_v4 --reasoning-parser hy_v4 --enable-auto-tool-choice --served-model-name hy4-preview

The flags matter, and each one maps to something in the model rather than being boilerplate:

• --attention-backend FLASHMLA_SPARSE — required, not optional. Tencent Hy4 Preview's Gated DSA sparse attention does not run correctly on the default dense backends, and this flag is what the official recipe sets.

• --speculative-config '{"num_speculative_tokens":3,"method":"mtp"}' — turns on the model's built-in MTP layer for speculative decoding, three draft tokens ahead. No separate draft model to deploy.

• --tool-call-parser hy_v4 and --reasoning-parser hy_v4 — the custom parsers that tell the server how Tencent Hy4 Preview emits tool calls and its chain-of-thought, with --enable-auto-tool-choice letting the model decide when to call tools.

Tencent recommends temperature 0.9 and top_p 1.0 for sampling. Reasoning defaults to a "high" effort chain-of-thought aimed at math, coding, and complex tasks; if you want a direct response without the long thinking, you pass a no_think reasoning effort in the request rather than swapping weights.

Day-zero support is not vLLM-only, which is itself notable for a release this fresh. SGLang shipped a prebuilt multi-architecture image (lmsysorg/sglang:hy4-preview) with NEXTN-style speculative decoding the same day, and the Ascend ecosystem landed 0-day support through vLLM-Ascend using W8A8-quantized weights across 16 Atlas 800I A3 NPUs. Open-weight releases often take weeks for the serving ecosystem to catch up; this one took hours.

The scores worth quoting

Every benchmark Tencent published for Tencent Hy4 Preview is vendor-reported — run on Tencent's own evaluation setup, unreproduced by any independent lab, and the model has been public for two days. Read them as the ceiling Tencent believes it hit, not as established fact. Independent verification will not exist until a third party runs it; nothing on the public leaderboards reflects this model yet.

The headline figure is Terminal Bench 2.1, a test of how well a model drives a real terminal while coding. Tencent reports Tencent Hy4 Preview at 85.4, which it claims ties Claude Opus 5 and passes DeepSeek V4 Pro, and which beats its own predecessor Hy3 by 14.6 points. That single number carries the release — it is the claim that puts an open-weight model at parity with the most expensive closed frontier models on a hard agentic-coding test — and it is the claim most in need of independent confirmation.

The rest of the internal table, all Tencent's own numbers: DeepSWE, a software-engineering benchmark, jumps from 28.0 on Hy3 to 64.3. SWE-bench Pro comes in at 65.7 and SWE-bench Multilingual at 82.9. On Toolathlon-Verified, a tool-calling test, Tencent reports 74.1, which it says surpasses Qwen 3.8 Max and GPT-5.6 Sol and approaches Kimi K3 and Claude Opus 5. On APEX-Agents pass@1 it lands at 37.1, a hair behind Kimi K3's 37.2. And in a separate internal blind evaluation, 163 Tencent-recruited experts scored 203 engineering tasks at 2.99 out of 4.00, edging GLM-5.3's 2.92 and Kimi K3's 2.94.

Two patterns are worth noting. Every strong number is on agentic engineering work — terminal driving, tool calling, repository-scale tasks — and none is on generic knowledge or reasoning, which fits the productivity positioning rather than being a coincidence. And Tencent describes DeepSWE and the others as narrowing, not closing, gaps; the one clean, marketable parity claim is the Terminal Bench 2.1 tie, and it is the claim most worth testing with your own workload.

What it actually costs to run

The self-hosting math is the first thing to be honest about. This is not a single-GPU model and it is not a desktop model. The FP8 checkpoint is close to a terabyte of weights, and the official recipe assumes tensor parallelism of 8 — eight high-bandwidth GPUs with fast interconnects (Tencent's reference hardware for FP8 is 8×B300, with full BF16 wanting 16×B200). If you do not have that footprint, you are not going to self-host it cheaply, and the sparse-attention backend means there is no shortcut around the memory for the KV cache at a 1M-token context.

One launch-week add-on changes part of that math for teams that want the open weights without the flagship footprint. Tencent's Hy team shipped a compressed GGUF build of Tencent Hy4 Preview, squeezing the ~1.5 TB BF16 release down to roughly 200 GiB with a per-layer mixed quantization scheme it calls MIX-STQ1_0 — bulk expert tensors drop to about 1.3 bits while residual-stream-critical layers keep higher precision. On Tencent's own evaluations the accuracy movement is small — SWE-bench Multilingual 82.9 to 81.3, MCP Atlas 83.7 to 83.2, IFBench 73.5 to 72.5 — a trade the vendor calls near-lossless. Those are Tencent's numbers on Tencent's setup, unreproduced so far. A 200 GiB model is still not a single-GPU model, but it is a meaningfully smaller bet than a near-terabyte FP8 checkpoint, and it puts the open-weights path within reach of a smaller multi-GPU node than the eight-B300 recipe assumes.

The API path is where the price gets interesting. On Tencent Cloud's TokenHub, Tencent Hy4 Preview lists at 6 yuan (~US$0.83) per million input tokens, 18 yuan (~US$2.50) per million output tokens, and 0.3 yuan (~US$0.04) per million tokens for cache hits. Output at roughly $2.50 per million is far below the $25-per-million output price of a top closed frontier model, and the cheap cache-hit rate makes long agentic loops — where the same system prompt and conversation prefixes get re-sent constantly — unusually affordable.

Tencent is also running two-week free access to Tencent Hy4 Preview inside WorkBuddy and CodeBuddy while the model is new, so the cheapest way to try it this week is free — and WorkBuddy is where the productivity positioning gets concrete. In any WorkBuddy conversation the model selector flips between Hy3 and Hy4 preview, so the split between office-productivity and analysis work on one side and coding on the other is a per-task choice rather than a deployment decision. That split matches where Tencent aimed the model, and hands-on tests in WorkBuddy have it one-shotting complex artifacts — a playable low-poly game prototype from a single prompt, a full quarterly business review assembled from ten attachments with zero manual interventions, and, in the tester demo that circulated this week, a black-hole simulation. These are community observations on Tencent's own app rather than vendor benchmarks — but they are the first hands-on signals of what the model does on real multi-step tasks, and they are free to reproduce before you spend anything.

The cost decision is exactly where a routing layer changes the math, and we will be honest about our own part in it: OrcaRouter does not route Tencent Hy4 Preview yet, because Tencent has not opened this model to third-party inference platforms — it runs on Tencent Cloud's own TokenHub endpoint and on self-hosted deployments of the open weights, and we do not claim to serve what we do not. What the routing layer brings is the decision framework around it. If you are choosing between an 8-GPU node and an API, the same pass-through discipline the rest of the catalog runs on applies the day Tencent opens this up: OrcaRouter passes provider list prices through at zero markup, so a vendor price cut is live on our side the same day rather than being resold and marked up. And for a two-day-old model whose only evidence is its vendor's benchmarks, automatic failover is the safe way to test it — put the proven model on your critical path and Tencent Hy4 Preview beside it, swapping in on tasks where its cost profile wins and failing back the moment it does not, all through one API and one key.

The limits that do not fit on a spec sheet

Tencent lists the known limitations plainly, and they should shape any deployment decision. Tencent Hy4 Preview is a text model, full stop — no vision or multimodal input — so anything image- or video-adjacent belongs on a different model. Tencent also flags slow-start behavior on complex tasks (long thinking before first output) and a tendency to over-self-verify, burning tokens double-checking work it has already done. That combination is exactly wrong for latency-sensitive chat and exactly tolerable for offline batch engineering work, which tells you where Tencent expects you to run it.

Two claims in the launch material deserve separate attention because they are about how the model was built, not what it scored. Tencent says Tencent Hy4 Preview participated in its own development — it helped optimize the training methods, data strategy, evaluation frameworks, and low-level operators that produced it, an early recursive self-improvement loop — and that it autonomously optimized its own inference system for a 31.8% end-to-end throughput gain over baseline. On the scientific side, Tencent reports it advanced the known lower bound of the Blaschke–Lebesgue problem in convex geometry from 0.380799 to 0.41104, and a 2.0x speedup on a 32,512-atom phospholipid bilayer simulation down to 54.9 milliseconds per step. Like everything else on day one, these are Tencent's own account.

And the most important absence is verification. There are no independent scores for Tencent Hy4 Preview anywhere yet — not on a public leaderboard, not from a third-party eval lab, not an Artificial Analysis entry. The model is too new for that. The internal expert blind test is a useful signal about how Tencent's own reviewers see it against GLM-5.3 and Kimi K3, but it is Tencent's reviewers on Tencent's tasks. Everything above the spec sheet is a claim awaiting a check.

Who should run Tencent Hy4 Preview this week

The honest answer splits into three groups this week, and one of them needs no hardware at all. If you just want to feel what Tencent Hy4 Preview does, the free WorkBuddy access is the fastest path — no GPU, no API key, open a conversation, switch the model selector to Hy4 preview, and hand it a Work or Coding task. If you have GPUs and run engineering workloads in batch — long-horizon agentic tasks, repository-scale analysis, offline evaluation — the model is worth a serious test right now: the weights are open, the context window is repository-scale, the vLLM path is a single command, and the launch-week GGUF build lowers the hardware bar for a trial run. If you are running latency-sensitive production chat, or anything multimodal, or anything where you cannot afford a model whose only evidence is its vendor's own benchmarks, wait — Tencent has iterated roughly every two months since February, and it has already said the next batch of Hy4 models is coming, so the preview is explicitly an early iteration.

For everyone else, the useful posture is a routing pattern: prove it beside a model you already trust, at a cost you can walk away from, and scale it only where its numbers hold up on your own workload. That is what a router exists for — the day Tencent opens this model to third-party inference, it joins the one API, 200-plus models, list-price pass-through catalog like any other, and the failover and composition tooling will already be in place. Until then, the weights are on Hugging Face, the API is on Tencent Cloud, and the free trial is in WorkBuddy — and the claim that an open-weight model reached the top tier of agentic coding finally has a specific number attached to it. The number is Tencent's. The test is yours.