Choosing between Hy3 preview and Hy4 preview comes down to one tradeoff: capability versus cost. Hy4 preview is Tencent's larger flagship, built for long-horizon coding and agent workflows. Hy3 preview is the lighter, cheaper option that still handles most day-to-day tasks. If you have been comparing the two Tencent Hunyuan models and cannot tell whether the jump is worth the price, this comparison lays out every difference that affects a real decision. We verified each figure against Tencent's own model cards, OpenRouter, and Artificial Analysis.
Hy4 preview is the flagship; Hy3 preview is the value pick
Hy4 preview and Hy3 preview are both open-weight Mixture-of-Experts models from the Tencent Hy (Hunyuan) team, but they sit at different points on the price-performance curve. Hy4 preview is the newer, far larger model aimed at the open-source frontier. Hy3 preview came first, runs at a fraction of the cost, and remains a strong fit for lighter production work.
The gap between them is generational. Hy4 preview activates more than twice the parameters per token, holds four times the context, and posts higher scores on every benchmark Tencent published. It also costs roughly four to five times as much per token and runs slower on early providers.
Suggested read: Kimi K3 benchmarks, which shows how much of a benchmark score comes from the harness rather than the model, with gaps of 17 to 26 points on the same model.
Specs: Hy4 preview roughly doubles the active parameters
Hy4 preview is a 770B-parameter model with 49B active per token, while Hy3 preview is a 295B model with 21B active. Both use a Mixture-of-Experts design, so only a slice of the full network runs on any given token. That slice is what drives inference cost and speed, which is why the active-parameter count matters more than the headline total.
The architectural leap is real. Hy4 preview runs a 78-layer backbone with 256 routed experts plus one shared expert per layer, activating the top eight routed experts on each token. It also adds a native multi-token prediction layer for speculative decoding and a 1M-token context window. Hy3 preview tops out at a 262K context.
Table 1: Hy3 preview and Hy4 preview core specifications, verified against Tencent and Hugging Face model cards (as of September 2026).
One difference gets missed in most write-ups: the reasoning controls narrowed. Hy3 preview exposes three effort levels, letting you dial thinking down for speed. Hy4 preview's published chat template accepts only two values, "high" (the default) and "no_think." For teams that relied on a middle setting to balance latency and depth, that is a step back, not forward.
The license change runs the other way. Hy3 preview shipped under the Tencent Hy Community License, which restricts commercial use with conditions. Hy4 preview is Apache 2.0, a far more permissive license for teams that want to self-host or build on the weights without legal review.
Benchmarks: Hy4 preview leads Tencent's numbers, but independent data lags
Hy4 preview outscores Hy3 preview on every benchmark Tencent published, though the most-cited result is an internal test rather than a third-party one. On Tencent's own reported figures, Hy4 preview posts gains across coding, agent, and reasoning suites.
The headline evidence is a blind evaluation Tencent ran in-house. According to Tencent's release page, 163 internal experts rated model outputs across 203 engineering tasks on a four-point scale. Hy4 preview averaged 2.99, edging out GLM 5.3 at 2.92 and Kimi K3 at 2.94. Against GLM 5.3 it recorded 46.8% wins, 12.8% ties, and 40.4% losses; against Kimi K3, 51.2% wins, 7.9% ties, and 40.9% losses.
Those margins are narrow. A 0.05 to 0.07 gap on a four-point scale, from an internal panel using Tencent's own task mix, is a directional signal rather than proof of broad superiority. It echoes other vendor-run scorecards from preview-stage launches. Treat it as a reason to run your own test, not a reason to skip one.
The head-to-head benchmark figures below are Tencent-reported, drawn from the Hy4 preview and Hy3 model cards. Several also appear as independently listed leaderboard results on Hugging Face for Hy4 preview.
Table 2: Tencent-reported benchmark scores, high-effort settings (as of September 2026). Figures sourced from Tencent and Hugging Face model cards, not independently reproduced.
The caveat that matters most for buyers is the absence of independent scoring. Artificial Analysis, the main independent benchmarking service, had not yet published an Intelligence Index score for Hy4 preview at the time of writing; its page noted an independent evaluation was forthcoming. For reference, the earlier Hy3-preview scored 34 on that index. Until independent numbers land, Hy4 preview's lead rests on vendor-reported data, the same labeling problem that shows up across recent frontier launches.
Suggested read: GLM 5.3 vs Kimi K3, a comparison where the benchmarks land near-tied and multimodality, pricing model, and weight availability decide it instead.
Pricing: Hy3 preview costs about one-fifth as much
Hy3 preview is dramatically cheaper to run, which is often the deciding factor. On OpenRouter, Hy3 preview lists at $0.18 per million input tokens and $0.60 per million output tokens. Hy4 preview lists at $0.834 input and $2.501 output, roughly four to five times higher.
That multiplier compounds in agent workflows. A multi-step agent resends large context on every turn, so a 1M-token window plus a higher per-token rate can turn into a serious bill. Estimate the cost of a full task loop, not a single prompt, before committing.
Table 3: OpenRouter list pricing, pricing as of September 2026. Provider rates change; reverify before publishing cost estimates.
Speed favors the smaller model too. OpenRouter measured Hy3 preview at about 79 tokens per second versus roughly 39 for Hy4 preview on early Tencent Cloud serving. Preview-stage throughput settles over time, so read these as launch-window figures rather than fixed performance.
Access: both are open weights, served through the same routes
Hy3 preview and Hy4 preview are both available through OpenRouter and Tencent Cloud, and both ship downloadable weights. Hy4 preview also released an FP8 quantized variant to reduce the hardware footprint, though a 770B backbone still demands serious multi-GPU capacity to self-host.
Neither is a lightweight local model. If you want the capability without operating inference infrastructure, a hosted API is the practical path for both. Self-hosting makes sense only when data placement, control, or sustained utilization justifies the operational work.
Which should you use?
Choose Hy4 preview when the task needs both long context and sustained multi-step execution, and choose Hy3 preview when cost, speed, or a controlled preview budget matters more. The decision is rarely about raw capability alone.
Hy4 preview is the stronger fit for repository-wide debugging, migration planning, document reconciliation, and research workflows that turn many files into a structured artifact. Its 1M context and higher benchmark scores are built for exactly that kind of long-horizon work.
Hy3 preview is the better default for high-volume, cost-sensitive tasks: short-form generation, classification, and lighter coding where a 262K window is plenty and the price gap is decisive. Its extra low-effort reasoning mode also gives you a speed lever Hy4 preview lacks.
One shared caveat applies to both. Tencent labels each a "preview" and flags two known issues: spending longer than necessary on reasoning, and over-verifying its own work. Both raise latency and token cost. Pin your configuration, define an acceptance test, and keep a fallback model.
Suggested read: Qwen 3.8 Max release date, on a model claimed to be second only to Fable 5 with no published benchmarks, model card, or methodology behind the ranking.
Build production apps with frontier models on Emergent
Hy3 preview and Hy4 preview show how fast open-weight models are climbing, but running a comparison is different from shipping a product. The right model for you depends on your actual workload, and the only reliable test is your own task set measured on cost, speed, and quality together.
Emergent lets you build and deploy full-stack applications without writing code, using frontier models from Anthropic, OpenAI, and Google through a single Universal LLM Key. You describe what you want to build, select your model per project, and Emergent handles the infrastructure end to end.
Most AI app builders stop at prototypes. Emergent creates production-ready apps you can actually launch.
- Production-ready apps
- Web & mobile apps
- Deploy in minutes
