AI Tools
What is Tencent Hy4 preview, and what does its API cost?
Tencent Hy4 preview API pricing is $0.834 per million input tokens and $2.501 per million output tokens, undercutting GLM-5.3 and Kimi K3 on both. The Apache-2.0 MoE model, 770B total parameters with 49B active and a 1M-token context window, launched August 28, 2026 and is live on OpenRouter and Tencent Cloud TokenHub.
What matters
- Hy4 preview costs $0.834 per 1M input tokens and $2.501 per 1M output tokens, with cache hits at $0.042 per 1M (verified August 31, 2026 on Tencent's announcement and OpenRouter).
- That undercuts GLM-5.3 (CNY 8/28, about $1.19/$4.16) and Kimi K3 ($3.00/$15.00) on both input and output.
- Weights are Apache-2.0 on Hugging Face: a 770B-parameter MoE with 49B active per token and a 1M-token context window.
- Tencent's internal blind evaluation scored Hy4 preview 2.99/4.00 versus GLM-5.3's 2.92 and Kimi K3's 2.94 (self-reported, 163 experts, 203 engineering tasks).
- It is a preview: Tencent lists over-reasoning and over-verification as known issues, so budget for longer outputs on simple tasks.
What is Tencent Hy4 preview?
Tencent Hy4 preview is a Mixture-of-Experts large language model from Tencent's Hunyuan team, released and open-sourced on August 28, 2026. It has 770B total parameters with 49B active per token, 78 layers, 256 routed experts plus one shared expert, and a context window of 1M tokens, per the official model card on Hugging Face.
Architecturally it follows the current open frontier: Gated DeepSeek Sparse Attention with IndexCache for long-context efficiency, an approach Tencent says was inspired by DeepSeek and GLM, plus a built-in MTP layer for speculative decoding. The weights ship under the Apache-2.0 license, so commercial use, modification, and self-hosting are all permitted.
Tencent positions the model for productivity work: software engineering, office tasks such as financial analysis and document-to-spreadsheet pipelines, game prototype generation, and scientific research. Outside the API, Hy4 preview is free for two weeks inside WorkBuddy and CodeBuddy, and reachable in Yuanbao and ima.
Hy4 preview API pricing: the exact numbers
Tencent sets Hy4 preview API pricing at USD 0.834 per million input tokens, USD 2.501 per million output tokens, and USD 0.042 per million tokens for cache hits, as stated in its August 28, 2026 announcement. OpenRouter lists identical figures for tencent/hy4-preview, so the same rates apply whether you call Tencent Cloud TokenHub or OpenRouter.
The cache-hit rate matters most for agentic workloads. Long agent sessions re-read large context blocks, and at $0.042 per 1M tokens a cached read costs roughly 20x less than a fresh input token. DeepSeek prices V4 Pro cache hits at $0.022 off-peak and $0.044 at peak, so Hy4 preview sits in the same bracket.
One cost caveat: Hy4 preview defaults to a high reasoning mode with deep chain-of-thought, per the Hugging Face model card. Output tokens accumulate quickly on simple requests unless you pass a lighter reasoning effort such as "no_think", so actual per-request spend depends on configuration.
How Hy4 preview pricing compares with GLM-5.3, Kimi K3, and DeepSeek V4
Against the current open-model frontier, Hy4 preview lands mid-pack on price, and the comparison depends on which rival and which time of day you pick. All four models below carry a 1M-token context window. GLM-5.3 is priced in CNY on Zhipu's official pricing page; at the August 31, 2026 exchange rate of about 6.74 CNY per USD, CNY 8/28 per 1M works out to roughly $1.19/$4.16.
Hy4 preview is cheaper than GLM-5.3 on both input and output, and dramatically cheaper than Kimi K3, whose $15.00 output rate is the outlier in this group. Against DeepSeek V4 Pro it splits the difference: V4 Pro off-peak pricing ($0.66/$1.98) undercuts Hy4, while peak pricing ($1.32/$3.96, during two weekday UTC windows) runs above it. DeepSeek peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday.
Tencent's own blind evaluation, run internally with 163 experts across 203 engineering tasks, scored Hy4 preview at 2.99 out of 4.00, ahead of GLM-5.3 (2.92) and Kimi K3 (2.94). That is a self-reported vendor result, not an independent benchmark, so treat it as directional.
Open weights, open license: what self-hosting Hy4 actually takes
Hy4 preview is genuinely open: Apache-2.0 weights on Hugging Face (tencent/Hy4-preview plus a pre-quantized Hy4-preview-FP8), ModelScope, GitCode, and CNB, with the repository at github.com/Tencent-Hunyuan/Hy4-preview. Tencent ships official vLLM and SGLang images, so serving it is a solved problem at the infrastructure level.
The honest catch is hardware. The official vLLM recipe runs the FP8 checkpoint with tensor-parallel size 8, meaning eight GPUs as the baseline for a usable deployment, and a 770B-parameter MoE is not something a single workstation hosts comfortably. Open weights remove licensing cost, not infrastructure cost. Teams without eight-GPU capacity should use the hosted API instead.
For fine-tuning, Tencent publishes a full pipeline in the model repository, and Apache-2.0 terms mean no gated license agreement or usage restrictions on the adapted model.
Who should use Hy4 preview, and who should wait
- Agent and tool-use builders: OpenRouter's listing highlights Hy4 preview for coding agents and complex tool-use workflows, and the 1M context with cheap cache hits suits long sessions.
- Teams that need open weights: the Apache-2.0 license and full fine-tuning pipeline remove the licensing friction some vendors attach to frontier models.
- Budget-conscious API users: $0.834/$2.501 per 1M is under GLM-5.3 and far under Kimi K3.
Hold off if you need production stability: this is a preview release with known rough edges. Tencent's model card lists over-reasoning on complex tasks and a tendency to over-verify its own work as known issues, which inflates output tokens on simple requests. If your workloads are bulk, non-urgent, and cache-heavy, DeepSeek V4 Pro's off-peak rates still win on price.
Known limitations of Hy4 preview
Tencent is explicit that this is an early version. The Hugging Face model card states there is real headroom left in both pre-training and post-training, and that the model ships with known issues, among them spending longer than necessary reasoning through complex tasks and a tendency to over-verify its own work.
The company also says the next batch of Hy4 series models is expected to roll out soon. Hy3 followed the same preview-first pattern, and Tencent credits that loop with making Hy3 substantially better, so the current pricing and capabilities are best read as a snapshot rather than a final state.
At a glance
| Model | Input | Output | Cache hit | Context |
|---|---|---|---|---|
| Tencent Hy4 preview | $0.834 | $2.501 | $0.042 | 1M tokens |
| DeepSeek V4 Pro | $0.66 off-peak / $1.32 peak | $1.98 off-peak / $3.96 peak | $0.022 off-peak / $0.044 peak | 1M tokens |
| GLM-5.3 (Zhipu) | CNY 8 (~$1.19) | CNY 28 (~$4.16) | CNY 2 (~$0.30) | 1M tokens |
| Kimi K3 (Moonshot) | $3.00 | $15.00 | $0.30 | 1M tokens |
FAQ
Is Tencent Hy4 preview free?
The weights are free: Apache-2.0 open source, downloadable from Hugging Face, ModelScope, GitCode, and CNB. The API is paid at $0.834 per 1M input tokens and $2.501 per 1M output tokens. WorkBuddy and CodeBuddy offer the model free for two weeks from the August 28 launch, and Hy3 access on both apps is free until September 30, 2026.
Where can I access Tencent Hy4 preview?
Through the API on OpenRouter (model tencent/hy4-preview) and Tencent Cloud TokenHub, or inside Tencent's apps: WorkBuddy, CodeBuddy, Yuanbao, and ima. For self-hosting, the Apache-2.0 weights are on Hugging Face, ModelScope, GitCode, and CNB.
How does Hy4 preview compare with DeepSeek V4 Pro and GLM-5.3?
On price, Hy4 preview ($0.834/$2.501 per 1M) undercuts GLM-5.3 (CNY 8/28, about $1.19/$4.16) and Kimi K3 ($3.00/$15.00), but DeepSeek V4 Pro is cheaper off-peak ($0.66/$1.98). All three offer 1M-token contexts. Tencent's internal blind evaluation put Hy4 preview ahead of both rivals, a self-reported result.