Tencent Hy4 Makes Open-Weight AI Too Important to Ignore

TL;DR: Tencent released Hy4 preview on August 28 with a 770 billion-parameter sparse backbone, 49 billion active parameters per token, a separate 10 billion-parameter multi-token-prediction layer, a 1,048,576-position configuration, and weights under Apache-2.0.^[1][2][5] The real story isn't a claimed benchmark win. Hy4, arriving days after Z.ai's GLM-5.3-Flash release, shows that open-weight AI is becoming a serious product and infrastructure strategy rather than a charitable side project.

Here is the simple version. An AI model is the engine. The app, tools, cloud service, and evaluation system around it are the car. Tencent has made Hy4's engine downloadable, but it also put the easiest roads to that engine inside CodeBuddy, WorkBuddy, Tencent Cloud, and OpenRouter.^[1][6]

That matters because open weights give developers a real escape hatch. A team can inspect the artifact, run it with supported serving software, adapt it, or move it to another host. Tencent can still win by making its own managed route more convenient. Open weights and commercial cloud distribution are not opposites. They can be the same funnel.

The timing makes the release more important. Our earlier GLM-5.3 analysis argued that public weights, papers, and serving tools can compound across model families. Hy4 is the immediate test of that thesis: a second major Chinese platform company has shipped another long-context, tool-oriented model with inspectable weights and a permissive license.

Hy4 In Plain English: A Giant Model That Wakes Only Part Of Itself

Hy4 is a text-generation model built as a Mixture of Experts, or MoE. Imagine a newsroom containing hundreds of specialist desks. Each time the model processes a token, a router calls a small set of desks instead of waking the entire building.

Tencent reports a 770B-parameter backbone with 49B parameters active per token. The first of its 78 layers uses a dense feed-forward network. The remaining 77 layers each contain 256 routed experts plus one shared expert, with the router selecting eight routed experts for a token.^[2][4] That is roughly 6.4% of the backbone activated per token, but active-parameter count is not the same thing as latency, memory use, or cost.

There is a counting wrinkle. Tencent's 770B figure excludes a native multi-token-prediction, or MTP, layer with another 10B total and 0.7B active parameters. Hugging Face therefore displays the repository as a 780B model.^[2] Neither number is wrong. They count different boundaries, which is why model-size comparisons need footnotes.

Hy4's configuration permits 1,048,576 positions. That is the model's maximum-position setting, not a promise that every provider will accept a million-token prompt at every output length, batch size, or concurrency level. Tencent-linked SGLang examples document 131,072-token BF16 service on H200 hardware and 262,144-token recipes on specified Blackwell systems.^[8]

What Tencent Actually Opened: Useful Weights, Not A Reproducible Training Recipe

Tencent calls Hy4 open source. The precise description is open weights under Apache License 2.0, accompanied by an inspectable configuration, chat template, deployment recipes, and fine-tuning support.^[2][4][5] That is substantially more useful than an API-only release.

Apache-2.0 is a familiar permissive license. Subject to its terms, teams can use, modify, and redistribute the released artifact. The public materials do not include the complete pre-training corpus, raw data provenance, full training code, training logs, or an end-to-end recipe that lets an outsider reproduce the model from scratch.

This distinction is not pedantry. Public weights give a buyer portability and negotiating leverage. A reproducible training stack would provide a different level of scientific auditability. Hy4 clearly delivers the first. Tencent has not established the second.

The uncomfortable truth is that “open” now describes a spectrum. Hy4 is valuable because its weights, license, and deployment materials are concrete. Calling the entire development process open source would claim more than the evidence supports.

Under The Hood: Long Context Is A Systems Problem

Hy4 combines Gated DeepSeek Sparse Attention, IndexCache, identity Hyper-Connections, MoE routing, and the native MTP layer.^[2][4] In plain English, the model tries to avoid spending equal attention on every old token, reuse retrieval work across layers, keep information flowing through a deep network, activate only a subset of its experts, and draft more than one future token when the serving stack allows it.

That combination is designed for agent work: repositories, document collections, tool traces, retries, and multi-step tasks. It does not make one million tokens free. Memory for model weights, the key-value cache, runtime overhead, concurrent users, and output generation still has to fit somewhere.

The released files make the scale impossible to ignore. The BF16 repository is about 1.56TB. The FP8 repository is roughly 804GB before runtime and key-value-cache overhead.^[2][3] Tencent's model card shows an eight-way vLLM command, while hardware-specific SGLang examples use different tensor-parallel layouts. Those are serving recipes, not proof that any eight GPUs will fit the model or meet a production service level.^[8][9]

Let's be clear: “downloadable” does not mean “runs on a laptop.” Open weights remove a legal and platform boundary. They do not repeal memory physics.

The Benchmark Boundary: A Promising Internal Signal, Not A Crown

Tencent reports that 163 internal experts rated model outputs across 203 engineering tasks in a blind side-by-side study. Hy4 averaged 2.99 out of 4, compared with 2.92 for GLM-5.3 and 2.94 for Kimi K3. Against GLM-5.3, Tencent reports 46.8% wins, 12.8% ties, and 40.4% losses.^[1]

That is useful evidence about Tencent's own product environment. It is not independent evidence that Hy4 is universally better. Tencent selected the tasks, raters, model versions, interfaces, tools, resource budgets, and aggregation method. The raw judgments and complete task set are not public.

Tencent's broader benchmark appendix is similarly mixed. It reports Hy4 at 65.7 on SWE-bench Pro and 85.4 on Terminal-Bench 2.1, while GLM-5.3 is listed at 64.6 and 88.2 or 88.3, depending on the entry.^[10] The appendix also mixes harnesses, provider settings, internal tests, and model snapshots. These are deployment signals, not an apples-to-apples championship.

The real story isn't that Hy4 has already won. It is that Tencent has released enough of the model for outside teams to test whether its claims survive their own repositories, agents, documents, and hardware.

Hy4 Meets Z.ai: The Open-Model Flywheel Is Becoming Visible

Hy4 and Z.ai's GLM-5.3-Flash are not the same machine. Hy4 is text-only, has a 770B/49B sparse backbone plus its MTP layer, uses Gated DSA with IndexCache and iHC, and ships under Apache-2.0. GLM-5.3-Flash is a 320B/18B native-multimodal model whose public configuration combines KDA-style layers, layers named deepseek_sparse_attention, and mHC, with weights under MIT.^[11][12]

Tencent says Hy4 was inspired by DeepSeek and GLM, and separately compared it with GLM-5.3 in internal testing.^[1] That does not establish model lineage, shared training data, or a shared serving path. The connection is strategic: separate labs are combining public research ideas, releasing inspectable artifacts, and relying on overlapping serving ecosystems.

The provider count exposes the next battleground. At the August 30 snapshot, OpenRouter listed one Hy4 provider and 20 GLM-5.3-Flash providers.^[7][14] Hy4's current page showed real developer activity, including 2.34B tokens attributed to Command Code and 763M to Claude Code in its “apps using this model” panel. OpenRouter does not disclose the window or unique-user count for those figures, so they are adoption signals, not market share.

Here's the genius: public weights turn every compatible host, fine-tuning project, quantizer, evaluation harness, and agent framework into a possible distribution channel. The model lab no longer has to build every useful experiment itself. It still has to earn trust when developers turn those experiments into production.

Why It Is Worth Testing: Choice Is Becoming Part Of The Product

OpenRouter's August 30 Hy4 route listed $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cached input tokens.^[7] Z.ai's GLM-5.3-Flash promotional price was $0.075 input, $0.25 output, and $0.015 cached input through September 9, with stated list prices of $0.15, $0.50, and $0.03.^[13]

Those are hosted-price snapshots, not a normalized cost comparison. Provider hardware, precision, prompt length, output length, caching, concurrency, speculative-decoding acceptance, time to first token, tail latency, and tool overhead differ. Self-hosting adds engineering, observability, security, upgrades, and capacity planning.

Still, the strategic benefit is real. A downloadable model gives buyers an alternative when price, policy, region, privacy, or availability changes. A standard serving interface makes it easier to compare routes. Public configuration lets an engineering team inspect assumptions rather than accepting every marketing claim as a black box.

Tencent also labels Hy4 a preview and documents known problems: it may spend too long reasoning through complex tasks and over-verify its work.^[2] That candor is a reason to pilot the model, not a reason to promote it blindly.

The open-source AI space is worth watching because the useful unit is no longer one model. It is the growing network of weights, licenses, papers, serving engines, hosts, evaluators, and products around them. Hy4 does not prove that open-weight models have beaten the best closed systems. It proves that dismissing them now means ignoring one of the fastest-compounding parts of the AI market.

Last updated: August 30, 2026