Tencent Hy4 Preview Explained: Inside the 770B Open-Weight Model That Helped Train Itself

Tencent open-sourced Hy4 preview on August 28, 2026, and buried inside the release notes is a detail that matters more than the parameter count: this is the first Hunyuan model that participated in optimizing its own training pipeline [1]. Not as a marketing line, but as a specific, described mechanism, Hy4 preview was used during its own development to help automate optimization of training methods, and it separately optimized its own inference infrastructure to a measured 31.8 percent throughput increase [1].

That is the headline worth writing about fast, and it is why this post exists the same day the weights landed on Hugging Face rather than a week later once the benchmark chart traffic has already peaked. The rest of the spec sheet is genuinely strong too: 770 billion total parameters with only 49 billion active per token, a context window past 1 million tokens, Apache 2.0 licensing, and day-one availability through OpenRouter, Tencent Cloud, and four of Tencent's own products [1] [2] [3].

This post walks through where Hy4 fits in the Hunyuan lineage, exactly how its architecture is built, what its self-optimizing training loop actually means in practice, how its benchmark numbers hold up against GLM-5.3, Kimi K3, and DeepSeek V4 Pro, and how to call or self-host it yourself. Every code block below is meant to be copied and run, not read as pseudocode.

Step 1: Where Hy4 Fits in the Hunyuan Lineage

Tencent did not surprise the market with Hy4. It told investors it was coming. During its Q2 2026 earnings presentation on August 12, 2026, Tencent confirmed it was training Hy4, the next generation of its flagship Hunyuan model, and said it would bring the model online later in the year, a promise it beat by more than three months [4]. That same earnings call disclosed that operating capital expenditure hit RMB 51.8 billion for the quarter, up 190 percent year over year, largely to support Hunyuan model training and serving infrastructure, and that usage of Hunyuan 3 tokens had grown roughly sixfold during the reporting period as the model moved from preview into general availability [4].

Hy3, the predecessor, was text-only, a meaningful limitation Tencent explicitly called out when confirming Hy4 was in training: the next generation would carry more parameters and, critically, would be multimodal, a capability Hy3 never had [4] [5]. Hy3 had already built a real reputation of its own. Independent comparisons against DeepSeek V4 Pro found Hy3 edging ahead on several coding and reasoning benchmarks despite entering the market as the less-hyped release, giving Tencent a genuine claim to open-weight competitiveness before Hy4 shipped at all [6].

Hy4 preview did not appear from nowhere on August 28 either. In the days before the open-source release, the model was spotted running gray tests inside Tencent's Yuanbao consumer AI app, the kind of soft, limited rollout labs commonly use to stress-test a model against real traffic before a public announcement [7]. That sequence, earnings-call confirmation in mid-August, quiet in-app testing in the following weeks, full open-source release by month's end, is a genuinely fast turnaround for a 770-billion-parameter frontier-class model, and it tracks with how aggressively Chinese labs have been shipping through the second half of 2026.

Hy4 Preview Did Not Ship in a Vacuum

Hy4 preview landed in one of the busiest stretches open-weight AI has seen all year. Two days earlier, on August 26, 2026, Alibaba released Qwen3.8-Flash-Next, a 125-billion-parameter multimodal MoE model with 6 billion active parameters, explicitly positioned as a preview of the coming Qwen4 architecture and delivering better results than Qwen3.7-Plus at roughly one-ninth the training cost [10]. Z.ai's GLM-5.3 shipped just a day before that. Three major Chinese labs put out frontier-class or near-frontier releases inside a single week, which says as much about the current pace of the open-weight race as any individual model's benchmark chart does.

What separates Hy4 preview from that cluster is scale and scope rather than timing alone. Qwen3.8-Flash-Next is explicitly a cost-efficiency play at a fraction of Hy4's parameter count, tuned for the workhorse tier rather than the top of the leaderboard. Hy4 preview, at 770 billion total parameters, is squarely aimed at the largest tier of open-weight models shipping in 2026, in the same conversation as GLM-5.3, Kimi K3, and DeepSeek V4 Pro rather than the smaller, cheaper models built to undercut them on cost per token alone. That positioning is exactly why the blind evaluation in Step 4 compares Hy4 preview against GLM-5.3 and Kimi K3 specifically, not against Qwen3.8-Flash-Next.

Step 2: Inside the 770B Mixture-of-Experts Architecture

Hy4 preview is a sparse Mixture-of-Experts transformer with 770 billion total parameters, of which only 49 billion activate for any given token, an activation ratio just above 6 percent [2] [3]. The network is 78 layers deep: a single dense feed-forward layer near the input, followed by 77 Mixture-of-Experts layers that carry the rest of the model's depth [2].

Expert Pool and Routing

Each MoE layer draws from a pool of 256 routed experts plus 1 shared expert that every token always passes through. For each token, the router selects the top 8 routed experts to activate alongside that always-on shared expert, meaning 9 of the pool's 257 total expert paths actually run per token, per layer [2]. The shared expert is worth pausing on specifically: rather than making every one of a token's expert assignments compete for capacity, a shared expert acts as a constant, general-purpose fallback the routing decision does not have to earn, which tends to stabilize training and give the router more freedom to specialize the remaining routed experts, since it no longer has to reserve one of its choices for generic, low-signal patterns every token needs handled.

A simplified illustration of what that activation ratio looks like at Hy4's actual scale:

python
/code
# Illustrates Hy4 preview's actual sparsity: 770B total parameters, but
# only the shared expert plus the top-8 routed experts activate per token,
# out of a pool of 256 routed experts across 77 MoE layers [2].
def hy4_active_params(total_params_b: float, num_moe_layers: int, routed_experts: int,
                       shared_experts: int, top_k: int) -> float:
    active_fraction = (top_k + shared_experts) / (routed_experts + shared_experts)
    # Rough approximation: MoE layers carry the bulk of total parameters,
    # so per-token active share tracks the per-layer expert activation ratio.
    return total_params_b * active_fraction

total = 770.0
estimate = hy4_active_params(total, num_moe_layers=77, routed_experts=256,
                              shared_experts=1, top_k=8)
actual_active = 49.0  # Hy4 preview's disclosed active parameter count [2]

print(f"Total parameters: {total}B")
print(f"Experts activated per token: 8 routed + 1 shared, out of 257 total paths")
print(f"Naive activation-ratio estimate: {estimate:.1f}B")
print(f"Actual disclosed active parameters: {actual_active}B")
print(f"Effective activation ratio: {actual_active / total:.1%}")

Running that script puts the sparsity in concrete terms: Hy4 preview carries 770 billion parameters of learned capacity but only ever computes with a small, token-specific slice of it, which is exactly what makes a model this large practical to serve at all. The tradeoff a Mixture-of-Experts design like this makes, more total capacity for the same per-token compute budget, is the same bet nearly every frontier-scale open-weight model released in 2026 has made, though the specific ratios differ. Our breakdowns of Nemotron 3 Ultra's hybrid Mamba-MoE design and DeepSeek V4-Pro's architecture cover two different labs' answers to the same underlying sparsity question, useful context for seeing where Hy4's roughly 6.4 percent activation ratio sits in that spectrum.

Attention: Gated DeepSeek Sparse Attention With IndexCache

Hy4's attention mechanism is what its model card calls Gated DeepSeek Sparse Attention with IndexCache, a sparse attention variant built on the sparse attention approach DeepSeek popularized, with an added gating mechanism and an indexed caching scheme layered on top [2]. Sparse attention in general works by having each token attend to a learned, data-dependent subset of prior tokens instead of the full sequence, cutting the quadratic cost of full self-attention down substantially while trying to preserve the ability to retrieve specific, important context from far back in a long sequence. The gating layer on top of that gives the model an additional learned signal for deciding how much weight to give the sparse attention pathway versus other signal at each position, and IndexCache is the serving-side mechanism that keeps the selected sparse index structure cached and cheap to reuse across a long generation, rather than recomputing which tokens matter on every single decoding step.

That combination, hidden size 6,144, vocabulary of 120,832 tokens, and a native context window past 1 million tokens, is what lets Hy4 preview serve genuinely long agentic sessions, entire codebases, long documents, multi-hour tool-use traces, without the attention cost exploding the way it would in a dense, full-attention transformer at the same context length [2] [3].

Step 3: The Self-Optimizing Training and Inference Loop

This is the part of the release that separates Hy4 from a routine parameter-count bump, and it is worth being precise about what Tencent actually claimed rather than rounding it up into a vaguer "AI trained itself" headline. Tencent's own release notes describe two distinct, separate applications of the model to its own development [1]:

1. Training method optimization. Hy4 preview participated, for the first time in a Hunyuan model, in the automated optimization of training methods, meaning the model itself was used as a component in a system that searched over or refined how later stages of its own training were configured, rather than a human engineer manually tuning every hyperparameter and data mixture by hand.
2. Inference infrastructure optimization. Separately, Hy4 preview autonomously optimized its own inference-serving infrastructure, and Tencent reports this produced a measured 31.8 percent throughput increase [1].

Neither claim means the model designed its own architecture from scratch or trained itself unsupervised end to end. What it describes is closer to using a capable model as an automated optimization agent inside two specific, bounded parts of the ML engineering pipeline, training configuration search and inference-serving tuning, tasks that traditionally consume significant senior engineering time and that a sufficiently capable coding-and-reasoning model is well suited to iterate on quickly. This is a meaningfully different and more grounded claim than the recursive self-improvement framing that circulates in AI safety discourse, and it is worth treating it exactly as specifically as Tencent described it, an agent automating a defined optimization task inside its own build pipeline, not a model rewriting its own weights.

A simplified illustration of the kind of search Tencent is describing, an agent iterating over inference-serving configuration choices to find a faster one:

python
/code
# Simplified illustration of the kind of automated configuration search
# Tencent describes Hy4 preview performing on its own inference-serving
# infrastructure, reportedly yielding a 31.8% throughput increase [1].
# Real optimization uses learned search over actual serving metrics, not
# random sampling; this only shows the iterative search-and-keep shape.
import random

def simulate_config_search(rounds: int, baseline_throughput: float) -> float:
    random.seed(3)
    best = baseline_throughput
    for round_num in range(1, rounds + 1):
        candidate = best * random.uniform(0.97, 1.06)
        if candidate > best:
            best = candidate
            print(f"round {round_num:>2}: found faster config, throughput={best:,.0f} tok/s")
    return best

baseline = 1000.0
final = simulate_config_search(rounds=25, baseline_throughput=baseline)
gain = (final / baseline) - 1
print(f"\nBaseline throughput: {baseline:,.0f} tok/s")
print(f"Final throughput after search: {final:,.0f} tok/s")
print(f"Relative gain: {gain:.1%} (Tencent reports 31.8% on the real system [1])")

The practical takeaway from Step 3 is not that Hy4 is unique in being used this way, plenty of labs use smaller internal models to help tune training runs, but that Tencent chose to disclose it specifically and attach a concrete number, 31.8 percent, to the inference-serving half of it. A specific, falsifiable number is a meaningfully stronger form of disclosure than a vague claim about AI-assisted development, and it is one more data point in a broader 2026 trend of frontier labs using their own models as engineering tools inside the next model's own development cycle.

Step 4: Benchmark Results, Blind Evaluation and Public Scores

Tencent evaluated Hy4 preview two different ways, and both are worth understanding on their own terms rather than collapsing into a single headline number.

The Blind Expert Evaluation

Tencent ran a blind, side-by-side evaluation using 163 internal domain experts rating model outputs across 203 real engineering tasks, comparing Hy4 preview head-to-head against GLM-5.3 and Kimi K3 without the raters knowing which model produced which response [4] [8]. The full win, tie, and loss breakdown:

Hy4 preview edged out both comparison models on average score, 2.99 out of 4.00 against GLM-5.3's 2.92 and Kimi K3's 2.94, and won more head-to-head matchups than it lost against each one [8]. It is worth flagging plainly that this is Tencent's own internal evaluation, run on Tencent's own selection of 203 tasks and rated by Tencent's own pool of experts, which is exactly the kind of self-reported benchmark this blog's own editorial standard treats as a claim to weigh alongside independent numbers, not a substitute for them.

Public Benchmark Scores

On the public benchmark suite Tencent published alongside the model card, Hy4 preview posts the following [2]:

GPQA Diamond is a graduate-level, multiple-choice science reasoning benchmark specifically constructed to resist simple lookup or memorization; a 92.3 score puts Hy4 preview solidly in frontier territory on pure scientific reasoning. DeepSWE and SWE-bench Pro are both real-world software engineering benchmarks built from actual GitHub issues and pull requests, testing whether a model can navigate an existing codebase and produce a correct, mergeable fix, not just generate plausible-looking code in isolation. SWE-bench Multilingual extends that same real-repository-fix format across codebases in multiple programming languages, and Hy4's 82.9 there suggests the model's coding ability is not narrowly tuned to English-heavy Python and JavaScript repositories the way some benchmarks quietly are.

The honest caveat belongs here, not buried in a footnote: independent reporting on Hy4's public benchmark standing notes it does not lead comprehensively, and it still trails GLM-5.3 specifically on some independently run coding and cybersecurity evaluations, including certain DeepSWE configurations and CyberGym, a benchmark focused on offensive and defensive security tasks [8]. Both facts are true at once: Hy4 preview won Tencent's own blind expert evaluation against GLM-5.3 head-to-head, and GLM-5.3 still leads on specific independently reported technical benchmarks. Neither number alone is the full picture, which is exactly why this post is showing both rather than picking whichever one flatters the headline more.

If you want the fuller landscape these numbers sit inside, our explainers on GLM-5.3's coding and cyber defense benchmarks and Kimi K3's 2.8 trillion parameter architecture cover both comparison models in the same depth this post covers Hy4.

Step 5: Pricing and Cost Efficiency

Hy4 preview's list pricing through OpenRouter and Tencent Cloud TokenHub is $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million tokens on cache hits [2] [3]. Converted from Tencent's native RMB 6-per-million-input, RMB 18-per-million-output pricing, that works out to meaningful savings against both comparison models from the blind evaluation: roughly 25 percent cheaper on input tokens and about 36 percent cheaper on output tokens than GLM-5.3, and roughly 70 percent cheaper on input and 82 percent cheaper on output than Kimi K3 [8].

A quick worked cost comparison for a realistic agentic workload, 500,000 input tokens and 2 million output tokens in a month, the kind of volume a mid-sized coding-agent product might see:

python
/code
# Worked monthly cost comparison for a mid-volume agentic workload,
# using Hy4 preview's list pricing and its disclosed savings vs GLM-5.3
# and Kimi K3 [2][3][8].
input_tokens_m = 0.5   # 500,000 input tokens
output_tokens_m = 2.0  # 2,000,000 output tokens

pricing = {
    "Hy4 preview": {"input": 0.834, "output": 2.501},
    "GLM-5.3 (approx, ~25%/36% higher)": {"input": 0.834 / 0.75, "output": 2.501 / 0.64},
    "Kimi K3 (approx, ~70%/82% higher)": {"input": 0.834 / 0.30, "output": 2.501 / 0.18},
}

for model, rates in pricing.items():
    cost = input_tokens_m * rates["input"] + output_tokens_m * rates["output"]
    print(f"{model:<38} monthly cost: ${cost:,.2f}")

That gap compounds fast at real production volume, and it is the specific reason cost-sensitive teams running high-throughput coding agents or document-processing pipelines are worth pointing at Hy4 preview even before comparing raw benchmark scores, since inference spend, not model capability alone, is very often the actual constraint on what an agentic product can afford to run at scale.

Step 6: Known Limitations, Read Directly From Tencent's Own Disclosure

A model card that only lists strengths is not a useful one, and Tencent's own documentation for Hy4 preview is specific about where the model still falls short, worth reading directly rather than skipping past [8]:

• It is an early preview with real headroom left in both pretraining and post-training. Tencent frames this explicitly as a preview, not a finished, fully post-trained release, which matters for anyone deciding whether to build production infrastructure around it today versus waiting for a stable release.
• It tends to spend longer than necessary reasoning through complex tasks. In practice this shows up as higher token consumption and latency on problems a more efficiently tuned model would resolve in fewer reasoning steps, a real cost consideration on top of the per-token pricing covered in Step 5.
• It has a tendency to over-verify its own work. Excessive self-checking is a known failure mode in reasoning models generally, and it directly compounds the previous point: more verification passes mean more tokens spent per completed task, even when the first answer was already correct.
• It does not lead comprehensively on public benchmarks, trailing GLM-5.3 specifically on some independently run coding and cybersecurity evaluations even while winning Tencent's own blind expert evaluation overall.

None of these are disqualifying for a preview release, and naming them plainly is a genuinely useful signal about how seriously to weigh the rest of the benchmark story in Steps 3 and 4.

Step 7: Calling Hy4 Preview Through OpenRouter

The fastest way to try Hy4 preview without standing up your own GPU infrastructure is through OpenRouter's OpenAI-compatible API, using the model slug tencent/hy4-preview [3]:

python
/code
# Hy4 preview is OpenAI-compatible through OpenRouter, using the model
# slug tencent/hy4-preview. It supports tools/tool_choice for function
# calling and response_format for structured JSON output [3].
from openai import OpenAI
import os

client = OpenAI(
    api_key=os.environ["OPENROUTER_API_KEY"],
    base_url="https://openrouter.ai/api/v1",
)

response = client.chat.completions.create(
    model="tencent/hy4-preview",
    messages=[
        {"role": "system", "content": "You are a precise coding agent working inside an existing repository."},
        {"role": "user", "content": "Find the bug in this function and return a patch: def merge(a, b): return a + b[::-1]"},
    ],
    temperature=0.2,
    max_tokens=1024,
    extra_headers={
        "HTTP-Referer": "https://your-app.example.com",
        "X-Title": "Hy4 preview test",
    },
)

print(response.choices[0].message.content)
print(f"Prompt tokens: {response.usage.prompt_tokens}, completion tokens: {response.usage.completion_tokens}")

Hy4 preview supports tools and tool_choice for function calling and response_format for structured JSON outputs [3], so agentic tool-use pipelines built around either the OpenAI function-calling convention or a JSON schema response contract can point at Hy4 preview with the same request shape they already use, no custom parsing layer required.

Step 8: Self-Hosting Hy4 Preview With vLLM

Because Hy4 preview ships under Apache 2.0 with open weights on Hugging Face, and has a published vLLM serving recipe, teams with data residency requirements or high enough request volume to justify their own infrastructure can pull the weights and serve them directly [2] [9]:

bash
/code
# Download Hy4 preview's open weights and serve them locally with vLLM.
# 770B total parameters requires a real multi-GPU node even quantized [2][9].
pip install "huggingface_hub[cli]"
huggingface-cli login

huggingface-cli download tencent/Hy4-preview \
  --local-dir ./hy4-preview \
  --local-dir-use-symlinks False

vllm serve tencent/Hy4-preview \
  --tensor-parallel-size 8 \
  --max-model-len 1048576 \
  --gpu-memory-utilization 0.92 \
  --trust-remote-code \
  --port 8000

At 770 billion total parameters, this genuinely requires a multi-GPU node even with quantization, the same capacity planning consideration that applies to any model in this size class. For most teams below very high inference volume, one of OpenRouter, Tencent Cloud TokenHub, or another hosted provider will be more cost-effective than the operational overhead of self-hosting, and self-hosting is the right call specifically once volume or data residency requirements justify it.

Step 9: Where Hy4 Preview Already Ships Inside Tencent's Own Products

Hy4 preview is not launching as a standalone API-only release. It is already integrated into four of Tencent's own products on day one: WorkBuddy and CodeBuddy, Tencent's productivity and coding-assistant tools, plus the consumer-facing Yuanbao and ima apps [1]. Tencent is also running a promotional access window, free usage on WorkBuddy and CodeBuddy for the first two weeks after launch, while extending free access to the outgoing Hy3 model through September 30, 2026, giving existing Hy3-dependent workflows a real runway to migrate rather than a hard cutover [1].

That pattern, shipping a new frontier-scale model directly inside existing consumer and productivity apps rather than only behind a developer API, is worth noting on its own. It means Hy4 preview's real-world usage signal will accumulate fast, across genuinely large existing user bases, well before most Western developers outside China have spent meaningful time with the model directly.

The same underlying idea, using a capable model as one component inside a larger automated content pipeline rather than a standalone chat window, is what powers tools like Text2Shorts in Miraflow AI, which turns a single topic into a script, scene visuals, and a finished video in one pass, or AI Clipping, which analyzes an entire long-form video upload to find and cut its strongest moments automatically. Neither end user ever needs to know which specific model is doing the work under the hood, the same way most WorkBuddy or CodeBuddy users interacting with Hy4 preview today likely have no idea a 770-billion-parameter model, one that helped tune its own training, is what is actually answering them.

Production Notes and Best Practices

A few practical points worth applying directly if you are evaluating Hy4 preview for a real workload rather than just testing it casually.

• Budget for verbose reasoning on complex tasks. Given Tencent's own disclosed tendency toward over-verification and longer-than-necessary reasoning traces, set conservative max_tokens limits and monitor actual token consumption on your specific task mix before committing to cost projections based on list pricing alone.
• Treat it as a coding-and-agentic specialist first. The benchmark profile, strong on DeepSWE, SWE-bench Pro, and SWE-bench Multilingual, is squarely aimed at software engineering and structured tool-use workloads, which is where the model card recommends deploying it.
• Use the free Hy3 migration window deliberately. If you are running production traffic on Hy3 today, the extended free access through September 30, 2026 is a real, time-boxed window to validate Hy4 preview against your own evaluation set before switching, not just a promotional gesture to ignore.
• Weigh the blind-eval win against the independent-benchmark gap honestly. Hy4 preview beat GLM-5.3 and Kimi K3 in Tencent's own expert evaluation while still trailing GLM-5.3 on some independently reported coding and security benchmarks. Pick whichever number matches your actual task distribution rather than quoting only the one that favors the model you already prefer.
• Plan multi-GPU capacity honestly if self-hosting. 770 billion total parameters is a genuinely large model to serve yourself. Confirm your node's memory budget against the vLLM recipe's requirements before committing to self-hosting over a hosted provider.

If you want to see the model's architecture rendered as motion rather than static diagrams, here is a video generation prompt built around the same tabletop metaphor used throughout this post's images, written for a Wan-style video model:

A wooden tabletop workshop scene shot from directly above. A single small brass token drops onto a wide grid of 257 numbered wooden compartments; most compartments stay dark, but a shared central compartment glows continuously and exactly eight scattered compartments light up briefly in sequence to receive and pass the token along, forming a short glowing chain before the token exits the frame at the bottom. Camera holds a steady overhead position, slowly zooming out at the end to reveal the full grid was mostly dark the entire time. Warm soft studio lighting, shallow depth of field, visible wood grain and brass reflections, no readable text, no logos, no people, smooth continuous motion throughout.

Common Mistakes to Avoid

• Treating 770 billion parameters and 49 billion active parameters as the same cost story. They describe total learned capacity versus actual per-token compute, and comparing Hy4 preview's inference cost against a dense 770-billion-parameter model badly overstates how expensive it is to run.
• Citing the blind expert evaluation as if it were an independent, third-party benchmark. It is Tencent's own internal evaluation, run on Tencent's own task selection. Useful signal, not a substitute for independently reported public benchmarks.
• Assuming the self-optimization claim means the model redesigned its own architecture. Tencent's own disclosure describes a bounded application to training-method configuration and inference-serving tuning, not autonomous architecture design or unsupervised self-training.
• Skipping the known-limitations section because the benchmark chart looks strong. Tencent's own disclosed weaknesses, over-verification and longer-than-necessary reasoning traces, directly affect real token spend and latency, independent of raw accuracy scores.
• Comparing Hy4 preview's RMB list price to a competitor's USD price without converting. The savings percentages in Step 5 are already converted; recomputing from mismatched currencies is an easy way to arrive at a misleading number.

Frequently Asked Questions

Is Hy4 preview free to use? The open weights are free to download and self-host under Apache 2.0. Hosted API access through OpenRouter and Tencent Cloud TokenHub is priced per token, though WorkBuddy and CodeBuddy offer free access for the first two weeks after launch [1] [2].

How is Hy4 preview different from Hy3? Hy3 was text-only. Hy4 preview is multimodal, roughly doubles the parameter count Tencent had previously disclosed for Hy3, and is the first Hunyuan model Tencent says participated in optimizing its own training and inference infrastructure [1] [4].

Does Hy4 preview really beat GLM-5.3 and Kimi K3? It won Tencent's own blind expert evaluation against both models head-to-head, 2.99 average score versus 2.92 and 2.94. On some independently reported public benchmarks, particularly certain coding and cybersecurity evaluations, GLM-5.3 still leads. Both are accurate, and neither alone is the complete comparison [8].

What does "helped train itself" actually mean here? Tencent describes two specific, bounded uses: Hy4 preview participating in automated optimization of training methods during its own development, and separately optimizing its own inference-serving infrastructure to a measured 31.8 percent throughput gain. It did not design its own architecture or train unsupervised end to end [1].

Can I run Hy4 preview on my own hardware? Yes. The weights are open on Hugging Face under Apache 2.0, and there is a published vLLM serving recipe. At 770 billion total parameters, this requires a real multi-GPU node even with quantization [2] [9].

How much cheaper is Hy4 preview than GLM-5.3 or Kimi K3? Roughly 25 percent cheaper on input tokens and 36 percent cheaper on output tokens than GLM-5.3, and roughly 70 percent cheaper on input and 82 percent cheaper on output than Kimi K3, based on Tencent's disclosed pricing comparison [8].

Conclusion

Hy4 preview is a genuinely significant release wrapped in a same-day launch that is easy to undersell if you only skim the parameter count. A 770-billion-parameter, 49-billion-active Mixture-of-Experts model with a context window past 1 million tokens, Apache 2.0 licensing, and day-one integration across four production apps would be a notable release on its own. What actually sets it apart is the specific, measured claim that it helped optimize both its own training configuration and its own inference-serving infrastructure, with a real number, 31.8 percent, attached to the second half of that claim. Tencent backed it with an honest accounting of where the model still falls short too, over-verification, longer reasoning traces than necessary, and real gaps against GLM-5.3 on specific independent benchmarks even after winning Tencent's own blind evaluation. That combination, genuine capability, a concrete efficiency story, and disclosed limitations instead of only a highlight reel, is what makes Hy4 preview worth a full technical read on the day it shipped rather than a headline skimmed a week later.

References and Sources

[1] Tencent. "Tencent Releases and Open-Sources Tencent Hy4 preview."

[2] Hugging Face. "tencent/Hy4-preview model card."

[3] OpenRouter. "Hy4 preview: API Pricing & Providers."

[4] BigGo Finance. "TENCENT HOLDINGS LTD Q2 FY2026 Earnings Call: CapEx Surges 190% to RMB 51.8B."

[5] Pasquale Pillitteri. "Tencent Confirms Hy4, a Bigger Multimodal Model, Is in Training."

[6] CodingFleet. "Hy3 vs DeepSeek V4 Pro: Open-Weight Showdown."

[7] KuCoin. "Tencent's Hunyuan Hy4 appears in the Yuanbao app, entering gray testing."

[8] KuCoin. "Tencent's Hy4 Outperforms GLM-5.3 and Kimi K3 in Internal Tests, Reduces Output Costs by 82%."

[9] vLLM Recipes. "tencent/Hy4-preview."

[10] MarkTechPost. "Alibaba's Qwen Team Releases Qwen3.8-Flash-Next: A 125B Multimodal MoE With 6B Active Parameters Previewing the Qwen4 Architecture."