Tencent Holdings pushed a new open-weight language model onto Hugging Face on Friday, and the specs alone explain why developers started downloading it within hours. Hy4 preview packs 770 billion total parameters into a mixture-of-experts design that only activates about 49 billion of them for any single request, according to Tencent’s own launch post. That’s a jump from the company’s last flagship release just seven weeks ago, and it lands the same week rivals Alibaba, Moonshot AI, and Zhipu are all fighting for the same slice of the open-source coding market.

The release matters beyond the spec sheet. Tencent is the world’s largest gaming company by revenue and one of China’s biggest cloud providers, and it just handed away the weights to a model built for software engineering, research, and financial analysis, three of the highest-value use cases in enterprise AI right now. This piece breaks down what Hy4 preview actually is, how it stacks up against the open-weight models it’s chasing, and what the release signals about where the AI coding-assistant race is headed next.

What Tencent Shipped on August 28

Tencent released and open-sourced a preview version of Hy4, its newest large language model, on Friday, August 28, 2026. The company said the model targets tasks like software engineering, scientific research, and financial analysis, and confirmed it uses a mixture-of-experts architecture rather than a single dense network. Tencent also published a lower-precision Hy4 preview-FP8 variant alongside the full-weight release, giving developers a smaller footprint option for local or edge deployment.

According to TechNode, the weights went live on Hugging Face the same day, with mirrors on ModelScope, GitCode, and CNB to cover developers who can’t easily reach Hugging Face from mainland China. Tencent’s Hunyuan team, which builds the Hy model family, wrote in its official announcement that the release is meant to demonstrate outstanding capabilities on real-world productivity tasks spanning coding, office work, and scientific research, a claim it’s now up to outside developers to test.

Inside the Mixture-of-Experts Architecture

Hy4 preview’s backbone runs 78 layers deep. The first layer uses a standard dense feed-forward network, while the remaining 77 swap in a mixture-of-experts block containing 256 routed experts plus one shared expert, per TechNode’s technical breakdown of the release. Every token that passes through activates the top eight routed experts along with the always-on shared expert, which is how the model keeps its active-parameter count down near 49 billion despite carrying 770 billion parameters in total.

That sparse-activation approach isn’t new (DeepSeek and Alibaba’s Qwen team both popularized similar designs over the past two years), but Tencent’s expert count and routing depth push the architecture further than its own prior release did. The tradeoff is the same one every MoE model makes: heavier storage and memory requirements to hold the full weight set, in exchange for inference costs that scale closer to the 49-billion active figure than the 770-billion total.

Context Window: Built for Long, Multi-Step Coding Sessions

Tencent said in its official launch post that Hy4 preview ships with a context window exceeding 1 million tokens, a spec confirmed independently by TechNode’s coverage of the release. That’s large enough to hold entire mid-sized codebases, long research papers with citations, or hours of agentic tool-call history in a single context, without the model losing track of earlier steps.

A window that size is aimed squarely at agentic coding workflows, where a model has to plan a task, call tools, read back the results, and keep adjusting its approach across dozens of turns. Tencent has said the model is designed for coding agents, complex tool-use workflows, and productivity tasks that require planning, context continuity, and sustained multi-step execution, rather than single-shot chat responses.

Hy4 Preview vs Hy3: What Changed in Four Months

Tencent’s last major open-weight release, Hy3, only reached full availability on July 6, 2026, after a preview build in late April that the company said was built in 90 days and gathered feedback from more than 50 products. Hy3 shipped as a 295-billion-parameter MoE model with 21 billion active parameters, a 256,000-token context window, and a hybrid fast/slow thinking mode that toggled between quick replies and deeper reasoning. It was also the first Hy release under a genuine Apache 2.0 license with no territorial restrictions, according to Tencent’s own announcement.

Hy4 preview roughly doubles Hy3’s active-parameter count and more than quadruples its total parameter count, while stretching the context window from 256,000 tokens to over 1 million. Less than two months separate Hy3’s full release and Hy4’s preview, a release cadence that mirrors what DeepSeek, Alibaba, and Moonshot AI have all been running this year as they trade the top spot on open-weight leaderboards.

Hy4 Preview vs Hy3: Spec Comparison

SpecHy3 (July 2026)Hy4 preview (Aug 2026)
Total parameters295 billion770 billion
Active parameters21 billion~49 billion
Context window256,000 tokens1,000,000+ tokens
LicenseApache 2.0, no territorial limitsOpen-weight release (license per Hugging Face card)
Target use caseAgentic coding, tool calling, long conversationsSoftware engineering, research, financial analysis
DistributionHugging Face, GitHubHugging Face, ModelScope, GitCode, CNB, OpenRouter

Benchmark Claims: Hy4 Preview vs GLM-5.3 and Kimi K3

Tencent ran an internal blind evaluation involving 163 outside experts scoring the model against 203 real engineering tasks, according to TechNode’s report on the release. Hy4 preview came out with an average score of 2.99 out of 4.00 in that evaluation, edging out Zhipu’s GLM-5.3 at 2.92 and Moonshot AI’s Kimi K3 at 2.94. The gap between all three scores is narrow enough that it’s more a statement of rough parity than a decisive win, and it’s worth flagging that Tencent designed and ran the evaluation itself.

None of the three companies have published a shared, third-party benchmark for this specific comparison yet, so treat Tencent’s numbers as a starting claim rather than an independently confirmed result. Public leaderboards on Hugging Face and OpenRouter, where developers can run their own head-to-head tests once the model is live, should fill that gap over the coming weeks.

Internal Benchmark: Hy4 Preview vs Rival Open-Weight Models

ModelDeveloperScore (163-expert eval, /4.00)
Hy4 previewTencent2.99
Kimi K3Moonshot AI2.94
GLM-5.3Zhipu2.92

Source: Tencent’s internal evaluation of 203 engineering tasks, as reported by TechNode.

Where Hy4 Preview Plugs Into Tencent’s Products

Tencent said it plans to integrate Hy4 preview with its own products, naming CodeBuddy and WorkBuddy specifically in its release notes. Both the domestic and international versions of CodeBuddy and WorkBuddy are expected to pick up the model, alongside Tencent’s Yuanbao and ima consumer AI apps, according to the company’s announcement. Developers who want to skip self-hosting can also reach Hy4 preview through Tencent Cloud’s Tokenhub API or through OpenRouter’s hosted listing, which went live alongside the open-weight release.

That distribution strategy, open weights for anyone who wants to self-host, plus a hosted API for anyone who doesn’t, is the same playbook DeepSeek and Alibaba have run with their own flagship releases. It lets Tencent capture usage from cost-sensitive enterprise teams running the weights on their own GPUs, while still selling inference through Tencent Cloud to developers who’d rather not manage the infrastructure.

Known Limitations: Slower Answers, Over-Verification

Tencent was direct about the rough edges. The company said the early-release model can sometimes take longer than necessary on complex questions and may over-verify its own answers, essentially double-checking work in ways that add latency without adding accuracy. That’s a common failure mode in reasoning-heavy models tuned for agentic tasks, where the training process rewards caution and tool re-checks, sometimes past the point of diminishing returns.

For teams evaluating Hy4 preview for production coding assistants, that tradeoff is worth testing directly against latency-sensitive workloads before committing. A model that second-guesses itself on straightforward requests can quietly inflate token costs and response times, even if its accuracy numbers look strong in a benchmark setting.

The Bigger Picture: China’s Open-Weight Models Are Winning Share

Hy4 preview didn’t land in a vacuum. Chinese AI companies held roughly 15% of global generative AI market share in November 2025, up from around 1% a year earlier, largely on the strength of DeepSeek’s open-source push, according to TrendForce. That number kept climbing: the South China Morning Post reported China’s open-source models now make up close to 30% of global AI usage, led by Alibaba’s Qwen and DeepSeek.

Qwen’s growth alone tells the story. Alibaba’s model family passed 700 million downloads on Hugging Face by January 2026 and hit 1 billion cumulative downloads by March, reaching that milestone faster than any open-source model family in history. In February 2026, Qwen generated 153.6 million downloads in a single month, more than its next eight competitors combined, and its ecosystem has spun off over 200,000 derivative models. By May 2026, Chinese open-weight models accounted for roughly 61% of all tokens consumed on OpenRouter, with four of the five most-used models on the platform built in China and Meta’s Llama family dropping out of the rankings entirely.

Market Share Snapshot: Chinese Open-Weight AI Growth

MetricFigureSource
Chinese AI global market share, Nov 2025~15% (up from ~1% a year prior)TrendForce
Chinese open-source share of global AI usage, 2026~30%South China Morning Post
Qwen cumulative Hugging Face downloads1 billion+ (by March 2026)Industry download trackers cited by SCMP/Forbes coverage
Chinese models’ share of OpenRouter tokens~61% (May 2026)OpenRouter usage data reported by industry press

Competitive Comparison: Where Hy4 Preview Sits in the Field

As of July 2026, the top tier of Chinese open-weight and hosted models includes DeepSeek, Qwen, Kimi, Doubao, GLM, and ERNIE, each backed by a different major tech company (Alibaba’s Qwen, Moonshot AI’s Kimi, ByteDance’s Doubao, Zhipu’s GLM, and Baidu’s ERNIE all compete for overlapping developer and enterprise audiences). Tencent’s Hy line has spent most of 2026 playing catch-up to Alibaba’s download numbers and DeepSeek’s benchmark reputation, and Hy4 preview is the company’s biggest swing yet at closing that gap on raw capability.

Outside China, Hy4 preview enters a field still dominated in enterprise deployments by closed models from OpenAI, Anthropic, and Google, but open-weight competition has been steadily eating into that lead for agentic coding tasks specifically, where cost per token and the ability to fine-tune or self-host matter more than they do in general chat use cases. A 770-billion-parameter model with a 1-million-token context window, released for free download, is a direct pitch to that exact audience: engineering teams that want frontier-adjacent capability without a per-token bill from a closed API provider.

What the Industry Is Saying

Tencent’s own materials do most of the talking on this release, since third-party technical review is still catching up to a model that shipped hours ago. In its official launch post, the company wrote that Hy4 preview “demonstrates outstanding capabilities on real-world productivity tasks spanning coding, office work, and scientific research” (Tencent).

On the research side, Tencent said the model “demonstrates stronger capabilities in understanding, reasoning through and solving complex research problems, with notable improvements across areas including AI research and development, molecular dynamics simulation, condensed-matter physics and fundamental mathematics” (Tencent). The Hunyuan team’s own model card confirms the distribution plan directly: “We open-source Hy4 preview and Hy4 preview-FP8 model weights on Hugging Face, ModelScope, GitCode, and CNB” (Tencent Hunyuan, Hugging Face model card).

Reaction outside Tencent’s own channels has moved fast on social media, where one widely shared X post relayed a claim that Hy4 “coordinated several Codex sessions in parallel, evaluated their results and adjusted the research direction, outperforming Codex working alone across eight benchmarks” (post by Chubby, X). That specific benchmark claim traces back to Tencent and hasn’t been independently verified, so it’s worth treating as a company talking point circulating online rather than a confirmed third-party result.

Historical Context: From Dense Models to a Sparse-Expert Arms Race

Tencent’s Hunyuan family didn’t start as a mixture-of-experts specialist. The broader Hunyuan ecosystem also includes image, 3D-asset, and video generation models, plus a set of compact 0.5-billion, 1.8-billion, 4-billion, and 7-billion parameter models Tencent released for low-power and edge deployments. The shift toward large, sparse MoE architectures for the flagship Hy line tracks a broader industry move that DeepSeek helped popularize in early 2025, when its V3 and R1 releases showed that MoE routing could deliver frontier-level reasoning at a fraction of the inference cost of dense models the same nominal size.

Alibaba, Zhipu, and Moonshot AI all followed with their own MoE flagships over the following year, and Tencent’s Hy3-to-Hy4 jump fits that same pattern: bigger total parameter counts, similar or lower active-parameter costs, and longer context windows aimed at agentic workflows rather than short-form chat. What’s changed since 2025 isn’t the architecture choice anymore, it’s the pace. Tencent shipped two major Hy generations in under five months.

Why This Matters for Software Engineering Teams

For engineering leaders evaluating AI coding tools, Hy4 preview adds one more credible open-weight option to a list that already includes Qwen, DeepSeek, Kimi K3, and GLM-5.3. The practical question isn’t whether Tencent’s internal benchmark numbers hold up exactly, it’s whether a 49-billion-active-parameter model with a 1-million-token context window can handle real repository-scale work at a lower cost than a closed API subscription.

Teams already running self-hosted inference for Qwen or DeepSeek models have the infrastructure to test Hy4 preview quickly, since it ships through the same Hugging Face and OpenRouter channels. Teams without that infrastructure will likely wait for Tencent Cloud Tokenhub pricing to stabilize, or for the model to show up in third-party benchmark aggregators like LMSYS or the Hugging Face Open LLM Leaderboard, before making a switch.

Predictions: What Comes Next

  • Expect a full, non-preview Hy4 release within two to three months, following the same preview-then-full-release pattern Tencent used with Hy3 (April preview, July full release).
  • Third-party benchmark sites will publish independent Hy4 preview scores within weeks, and those numbers will likely narrow or contradict Tencent’s internal 2.99-vs-2.92-vs-2.94 comparison against GLM-5.3 and Kimi K3.
  • Alibaba, Moonshot AI, or Zhipu will respond with their own updated releases before the end of 2026, continuing the sub-quarter release cadence the whole field has settled into.
  • Chinese open-weight models’ share of OpenRouter token volume, already around 61% as of May 2026, is likely to keep climbing as more agentic coding tools default to cost-competitive open models for high-volume workloads.
  • Enterprise adoption outside China will stay cautious in regulated industries, even as developer-level adoption on Hugging Face and OpenRouter grows, because procurement teams move slower than individual engineers experimenting with a new checkpoint.

How to Try Hy4 Preview

Developers who want to test the model without downloading 770 billion parameters worth of weights can query it through OpenRouter’s hosted API, which listed Hy4 preview the same day Tencent published the weights. A basic request looks like this:

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "tencent/hy4-preview",
    "messages": [
      {"role": "user", "content": "Summarize this pull request diff and flag any race conditions."}
    ]
  }'

Self-hosting is the other route, using the full-precision weights or the smaller Hy4 preview-FP8 build from Hugging Face, ModelScope, GitCode, or CNB. Given the parameter count, running the full model locally requires multi-GPU infrastructure well beyond a single consumer card, which is exactly why Tencent shipped the FP8 variant alongside it.

Frequently Asked Questions

What is Tencent Hy4 preview?

Hy4 preview is an open-source large language model Tencent released on August 28, 2026, built with a mixture-of-experts architecture and targeted at software engineering, scientific research, and financial analysis tasks.

How many parameters does Hy4 preview have?

Tencent said the model has 770 billion parameters in total, with about 49 billion active for any given request, thanks to its mixture-of-experts routing.

Is Hy4 preview free to use?

The model weights are open-source and free to download from Hugging Face, ModelScope, GitCode, and CNB. Hosted API access through Tencent Cloud Tokenhub or OpenRouter is billed per use, though exact pricing has not been independently confirmed at the time of writing.

How does Hy4 preview compare to GLM-5.3 and Kimi K3?

In Tencent’s own internal evaluation of 203 engineering tasks scored by 163 outside experts, Hy4 preview scored 2.99 out of 4.00, ahead of Kimi K3 at 2.94 and GLM-5.3 at 2.92. Independent third-party benchmarks have not yet confirmed these results.

What is the context window on Hy4 preview?

Tencent said the context window exceeds 1 million tokens, up from the 256,000-token window on its predecessor, Hy3.

Which Tencent products will use Hy4 preview?

Tencent said it plans to integrate the model with CodeBuddy and WorkBuddy, its coding and productivity assistants, along with the Yuanbao and ima consumer apps.

What are Hy4 preview’s known limitations?

Tencent acknowledged the early-release model can take longer than necessary on complex questions and may over-verify its own answers, adding latency without a proportional accuracy gain.

How does Hy4 preview fit into China’s open-source AI push?

It’s the latest entry in a wave of Chinese open-weight releases that has pushed China’s share of global AI usage to roughly 30%, according to the South China Morning Post, with Alibaba’s Qwen and DeepSeek leading download and usage numbers ahead of Tencent’s Hy line.