Tencent Hy4 Preview vs GLM 5.3 and Kimi K3: Who Wins?
Tencent's blind expert evaluation shows Hy4 preview edging out GLM 5.3 and Kimi K3 on real engineering tasks. Here's how the numbers break down.
What is Hy4 preview and how does it compare to GLM 5.3 and Kimi K3?
Hy4 preview is Tencent’s newest open-weight flagship model, a 770 billion parameter Mixture-of-Experts system with 49 billion parameters active per token. In Tencent’s own blind evaluation, 163 internal experts scored it against GLM 5.3 and Kimi K3 on 203 real engineering tasks. Hy4 preview edged out GLM 5.3 (2.99 vs. 2.92 average score, winning 46.8% of head-to-head comparisons) and beat Kimi K3 more clearly (2.99 vs. 2.94, winning 51.2% of comparisons). Both margins are real but narrow, not a rout.
TL;DR
- Hy4 preview is a 770B-parameter MoE model from Tencent’s Hunyuan team with 49B parameters activated per token, released under Apache 2.0 with full weights on Hugging Face, ModelScope, GitCode, and CNB.
- The architecture borrows heavily from DeepSeek and GLM, using Gated DeepSeek Sparse Attention with IndexCache for cross-layer index reuse, plus identity Hyper-Connections to widen the residual stream to four parallel paths.
- The blind evaluation used 163 Tencent engineers and experts judging 203 real engineering tasks, not a synthetic benchmark suite, which makes the win margins harder to game but also harder to independently reproduce.
- Hy4 preview beat GLM 5.3 by a thin margin (46.8% wins, 12.8% ties, 40.4% losses) and beat Kimi K3 more decisively (51.2% wins, 7.9% ties, 40.9% losses).
- The model ships with a 1M token context window and a native multi-token-prediction layer for speculative decoding, aimed squarely at long-horizon coding and agentic workflows.
- Tencent is explicit about known weaknesses, including a tendency to over-reason on simple tasks and to over-verify its own outputs, issues it says it will iterate on quickly rather than delay release to fix.
- Deployment is production-oriented from day one, with official vLLM and SGLang Docker images, an FP8 quantized variant, and a finetuning pipeline already published.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
How was the Hy4 preview benchmark actually run?
Tencent didn’t rely purely on automated benchmark suites for its headline comparison. Instead, it ran a blind side-by-side evaluation using 163 internal experts, described as software engineers, game developers, finance analysts, and security specialists, rating model outputs on 203 engineering tasks pulled from the kind of work those teams ship day to day.
Each task was scored on an average rating scale, and also tracked as a win, tie, or loss against the competing model. Against GLM 5.3, Hy4 preview scored an average of 2.99 versus 2.92, translating to 46.8% wins, 12.8% ties, and 40.4% losses. Against Kimi K3, Hy4 preview again averaged 2.99 versus 2.94, with a cleaner spread of 51.2% wins, 7.9% ties, and 40.9% losses.
The methodology matters here. Blind human evaluation on real tasks is generally considered more resistant to benchmark contamination than public leaderboard scores, since the judges are rating actual work output rather than matching against a known answer key. But it’s also a closed, internal process. Tencent hasn’t published the task set or the judges’ rubric, so outside teams can’t rerun the comparison themselves. The numbers are best read as a strong signal of competitiveness, not a fully independent verdict.
What’s under the hood of Hy4 preview’s architecture?
Hy4 preview is a Mixture-of-Experts model with 770B total parameters, of which 49B activate per token. The backbone runs 78 layers: the first layer uses a standard dense feed-forward network, and the remaining 77 replace it with MoE blocks containing 256 routed experts plus 1 shared expert. Every token activates the top 8 routed experts alongside the shared expert, keeping compute manageable despite the model’s total size.
Two architectural choices stand out. First, the attention mechanism uses Gated DeepSeek Sparse Attention (Gated DSA) combined with IndexCache, a technique for reusing sparse attention indices across layers. This is explicitly inspired by DeepSeek’s and GLM’s own sparse attention work, a sign of how quickly architectural ideas are propagating across the open-source Chinese model labs. Second, the residual pathway uses identity Hyper-Connections (iHC), which expand the model’s residual stream into four parallel paths instead of the usual single stream, intended to widen information flow between layers.
The model also ships with a native multi-token-prediction (MTP) layer, a 10B-parameter add-on with 0.7B activated, built specifically for speculative decoding. That’s a meaningful inference-speed lever: MTP layers let the model draft multiple tokens ahead and verify them in a batch, cutting the number of full forward passes needed to generate a response. Context length tops out at 1 million tokens, which puts it in the same bracket as other frontier-class long-context models built for large codebases or document sets.
Is Hy4 preview worth using over GLM 5.3 or Kimi K3?
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
The honest answer is that the margins reported by Tencent are close enough that the “better” model probably depends on the specific task. A 46.8% to 40.4% win rate over GLM 5.3, with nearly 13% of comparisons ending in a tie, is a real edge but not a decisive one. The Kimi K3 comparison is a bit more favorable to Hy4 preview, with a cleaner majority of wins and fewer ties.
What might tip the decision in practice is less about raw win rate and more about deployment fit. Hy4 preview is released under Apache 2.0, a permissive license that allows commercial use and modification without the copyleft restrictions some other licenses carry. It ships with day-one support for vLLM and SGLang, official Docker images tagged for the release, and an FP8 quantized variant for teams that want to cut memory footprint without standing up their own quantization pipeline. Tencent also published a finetuning guide and points to its AngelSlim toolkit for further compression, including low-bit quantization and speculative sampling support.
Teams already invested in GLM or Kimi tooling may not find enough daylight in these win rates to justify a migration. Teams evaluating open-weight options fresh, particularly for coding-heavy or long-context agentic work, now have a third serious contender with transparent (if internally run) task-based evaluation data instead of only synthetic benchmark scores.
What tasks is Hy4 preview actually built for?
Tencent designed the post-training data around specific internal use cases rather than general chat performance. The stated focus areas are software engineering (understanding, planning, debugging, and verifying long development tasks, plus front-end visual quality), office and analysis work (turning scattered file context into documents, spreadsheets, and presentations, including financial modeling), game development (going from a single prompt to a playable prototype and iterating inside game engines across multiple turns), and scientific research (reasoning through problems in AI research, molecular dynamics, condensed matter physics, and pure mathematics).
This is consistent with the model’s co-design alongside Tencent’s own products, CodeBuddy and WorkBuddy, suggesting the benchmark tasks used in the blind evaluation likely skew toward engineering and office-productivity work rather than open-ended creative writing or casual conversation.
What are Hy4 preview’s known weaknesses?
Tencent is unusually direct about this in its own documentation. The model card states plainly that this is an early preview with real headroom left in both pre-training and post-training. Two specific issues are called out: the model tends to spend longer than necessary reasoning through complex tasks, and it has a tendency to over-verify its own work, essentially double-checking outputs more than needed. Both issues point toward inference costs that may run higher than the raw parameter-activation numbers suggest, since a model that reasons longer per query burns more compute per response regardless of its parameter efficiency.
Tencent frames this as a deliberate ship-early strategy, referencing its prior Hy3 preview release as evidence that shipping early and iterating on real feedback led to substantial improvement. Whether that pattern holds for Hy4 is something only time and further releases will confirm.
Frequently Asked Questions
How many parameters does Hy4 preview have?
Hy4 preview has 770 billion total parameters, with 49 billion active per token thanks to its Mixture-of-Experts design. It also includes a smaller 10 billion parameter multi-token-prediction module (0.7B activated) used for speculative decoding.
What license is Hy4 preview released under?
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Hy4 preview is released under the Apache 2.0 license, which permits commercial use, modification, and redistribution with minimal restrictions.
How was Hy4 preview compared against GLM 5.3 and Kimi K3?
Tencent ran a blind evaluation using 163 internal experts who rated model outputs on 203 real engineering tasks. Hy4 preview scored 2.99 on average versus 2.92 for GLM 5.3 and 2.94 for Kimi K3, with win rates of 46.8% and 51.2% respectively.
Does Hy4 preview support long context windows?
Yes, Hy4 preview supports up to 1 million tokens of context, positioning it for large codebases, long documents, and extended multi-turn agentic tasks.
What are the biggest limitations of Hy4 preview right now?
Tencent’s own documentation flags two main issues: the model often spends more time reasoning than necessary on complex tasks, and it tends to over-verify its own outputs. Both are expected to improve in future iterations.