AI Tools
Tencent Hy4 preview vs DeepSeek V4: which Chinese frontier LLM should you build on?
Hy4 preview vs DeepSeek V4 is a real fork in the road for teams evaluating Chinese frontier models: Tencent's Hy4 preview is Apache-2.0 open weights you can self-host, priced at $0.834 per million input tokens on the API, while DeepSeek V4 is API-only with peak/off-peak rates from $0.22 per million cache-miss input tokens off-peak. Both models ship a 1M-token context window. Researched September 1, 2026.
What matters
- Tencent Hy4 preview launched and open-sourced August 28, 2026 under Apache-2.0: 770B total parameters, 49B active per token, 1M context.
- Hy4 preview API pricing is flat: $0.834 per 1M input tokens and $2.501 per 1M output tokens on Tencent Cloud, with a $0.042 cache-read rate.
- DeepSeek V4 is API-only: deepseek-v4-flash and deepseek-v4-pro, both 1M context, with off-peak prices 50% below peak; flash output is $0.66/M off-peak.
- The honest catch: Hy4's free weights need serious hardware, the official vLLM recipe runs the FP8 checkpoint with tensor-parallel-size 8, meaning 8 GPUs.
- Choose Hy4 preview for open weights, self-hosting, or license freedom; choose DeepSeek V4 when cheap cache-hit input and pay-as-you-go pricing matter more.
What is Tencent Hy4 preview?
Tencent Hy4 preview is a Mixture-of-Experts (MoE) flagship model from Tencent's Hunyuan team, released and open-sourced on August 28, 2026. The model card on Hugging Face lists 770B total parameters with 49B activated per token, 78 layers, 256 routed experts plus one shared expert, and a 1M-token context window. Architecture-wise it uses 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.
Two checkpoints are available: Hy4 preview (BF16) and Hy4 preview-FP8, both under the Apache-2.0 license, downloadable from Hugging Face, ModelScope, GitCode, and CNB. Tencent positions it as a productivity model, co-designed with its own CodeBuddy and WorkBuddy products.
What is DeepSeek V4?
DeepSeek V4 is DeepSeek's current generation of API models, offered as deepseek-v4-flash (version DeepSeek-V4-Flash-0731) and deepseek-v4-pro (DeepSeek-V4-Pro-0813). Both support a 1M-token context with a 384K maximum output, JSON output, tool calls, the Responses API, and the Anthropic API format, per DeepSeek's official pricing documentation. Unlike Hy4 preview, DeepSeek publishes no open weights for V4: there is no DeepSeek-V4 repo on Hugging Face, so self-hosting is not an option.
DeepSeek V4 pricing is time-based. Peak hours (01:00-04:00 and 06:00-10:00 UTC, Monday through Friday) cost double the off-peak rates for every line item, and context caching is built in, making cache-hit input dramatically cheaper than cache-miss input.
Hy4 preview vs DeepSeek V4: side-by-side comparison
The table below compares verified facts from the official Hugging Face model card, Tencent's release announcement, and DeepSeek's pricing documentation, all checked on September 1, 2026.
Pricing compared: flat rates vs peak/off-peak
Hy4 preview keeps things simple: one flat input price and one flat output price, no time-of-day discounting. DeepSeek V4 uses a peak/off-peak schedule where off-peak rates are exactly 50% of peak rates, plus a three-way split between cache-hit input, cache-miss input, and output.
The cheapest DeepSeek V4 path is deepseek-v4-flash with cache hits during off-peak hours: $0.007 per 1M input tokens. The same tokens cost $0.834 on Hy4 preview's API, which is why cache-heavy workloads with repeated system prompts can favor DeepSeek. But Hy4 preview's flat $0.834 input price has no hidden time-of-day penalty, and its $2.501 output rate sits between DeepSeek's flash ($0.66 off-peak, $1.32 peak) and pro ($1.98 off-peak, $3.96 peak) output prices.
The open-source reality check: what self-hosting Hy4 preview costs
Apache-2.0 means the weights are free to download and use commercially, but Hy4 preview is a 770B-parameter MoE. The official vLLM recipe in the model card runs the FP8 checkpoint with tensor-parallel-size 8, which means a minimum of 8 GPUs working together; SGLang's official prebuilt image ships the same configuration. That puts self-hosting out of reach for individual developers and small teams who do not already operate multi-GPU infrastructure.
The practical buyer's frame: Hy4 preview's open license matters for companies that need data control, fine-tuning, or license certainty. For everyone else, the API at $0.834/$2.501 per 1M tokens is the realistic way to use it, and DeepSeek V4's API remains the cheaper option for high-cache-hit traffic.
Which should you choose?
Pick Hy4 preview if you want Apache-2.0 weights you can fine-tune or self-host, if your workloads have low cache reuse, or if predictable flat pricing matters more than off-peak discounts. Tencent's own blind evaluation, run by 163 internal experts across 203 engineering tasks, scored Hy4 preview slightly ahead of GLM 5.3 (2.99 vs 2.92) and Kimi K3 (2.99 vs 2.94), though that is a vendor-run study.
Pick DeepSeek V4 if you are API-only, if repeated system prompts give you high cache-hit rates, or if you can schedule heavy jobs into off-peak hours. deepseek-v4-flash at $0.66/M output off-peak is the cheapest per-token path of the three models compared here. Both models are strong; the decision is about your infrastructure and traffic patterns, not raw capability.
At a glance
| Model | Open weights | Parameters | Context | Input price per 1M (cache miss) | Output price per 1M | API source |
|---|---|---|---|---|---|---|
| Tencent Hy4 preview | Yes, Apache-2.0 | 770B total / 49B active | 1M | $0.834 flat | $2.501 flat | Tencent Cloud via OpenRouter |
| DeepSeek V4 flash | No (API only) | Not published | 1M (384K max output) | $0.22 off-peak / $0.44 peak | $0.66 off-peak / $1.32 peak | DeepSeek official API |
| DeepSeek V4 pro | No (API only) | Not published | 1M (384K max output) | $0.66 off-peak / $1.32 peak | $1.98 off-peak / $3.96 peak | DeepSeek official API |
FAQ
Is Tencent Hy4 preview free and open source?
Yes. Hy4 preview is released under the Apache-2.0 license, and the weights are downloadable from Hugging Face, ModelScope, GitCode, and CNB. The API is paid: $0.834 per 1M input tokens and $2.501 per 1M output tokens. Free does not mean cheap to self-host: the official vLLM recipe requires 8 GPUs (tensor-parallel-size 8) for the FP8 checkpoint.
Is DeepSeek V4 open source?
No. DeepSeek V4 (deepseek-v4-flash and deepseek-v4-pro) is available only through DeepSeek's paid API. No open weights for V4 exist on Hugging Face, so self-hosting is not possible.
Which is cheaper, Hy4 preview or DeepSeek V4?
It depends on traffic. For cache-miss input and output tokens, Hy4 preview's flat $0.834/$2.501 per 1M sits between DeepSeek V4 flash ($0.22/$0.44 input, $0.66/$1.32 output off-peak) and pro ($0.66/$1.32 input, $1.98/$3.96 output off-peak). For cache-heavy workloads, DeepSeek V4 wins by a wide margin: cache-hit input is as low as $0.007 per 1M on flash off-peak, versus $0.042 cache-read on Hy4's API.
Related reading
Hy4 preview API pricing vs GLM-5.3 and Kimi K3, DeepSeek API pricing: peak and off-peak rates, All ToolBistro radar articles