Head-to-Head Specification Table
| Hy4 Preview | DeepSeek-V4-Pro | DeepSeek-V4-Flash | |
|---|---|---|---|
| Developer | Tencent Hy Team | DeepSeek | DeepSeek |
| Architecture | MoE | MoE + DSA sparse attention | MoE |
| Total parameters | 770B | 1.6T | 284B |
| Active parameters / token | 49B | 49B | 13B |
| Activation ratio | ≈6.4% | ≈3.1% | ≈4.6% |
| Layers | 78 | 61 | 61 |
| Experts per MoE layer | 256 routed + 1 shared | 384 routed + 1 shared | 384 routed + 1 shared |
| Experts activated per token | top-8 + shared | top-6 + shared | top-6 + shared |
| Max output | 64K | 384K | 384K |
| Input price (1M) | $0.834 (flat) | $0.66 off-peak / $1.32 peak | $0.22 off-peak / $0.44 peak |
| Output price (1M) | $2.501 (flat) | $1.98 off-peak / $3.96 peak | $0.66 off-peak / $1.32 peak |
| Open weights | No (hosted API only) | Yes | Yes |
| Release | 2026 (grey-testing) | 2026-04-24 (GA 2026-08-13) | 2026-04-24 |
Sources: Tencent Hy official research page and HuggingFace; DeepSeek API Docs, Models & Pricing, and V4 Preview release notes. DeepSeek prices shown are cache-miss rates; cache-hit rates may be lower.
The Most Interesting Detail: Same 49B Active, Very Different Total
Both Hy4 Preview and DeepSeek-V4-Pro activate 49B parameters per token. Hy4 reaches that active count with 770B total parameters, while DeepSeek-V4-Pro lists 1.6T total. Their activation ratios are therefore different: approximately 6.4% for Hy4 and 3.1% for DeepSeek-V4-Pro.
In practical terms, a higher activation ratio can mean capacity is distributed across fewer, larger experts, while a lower ratio can reflect a larger pool of specialists with a smaller share selected for each token. Neither ratio is a universal quality score. Hy4 documents its internals in unusual detail—78 layers, 256 routed experts plus one shared expert per layer, top-8 routing, and a native MTP layer—while equivalent layer-level detail for DeepSeek is not listed here.
Architecture Compared
Hy4 Preview
This is a Mixture-of-Experts model with a 78-layer backbone. Layer 1 is a dense FFN; layers 2–78 each contain 256 routed experts and one shared expert. Every token activates the top eight routed experts plus the shared expert. A native MTP layer, listed as 10B total and 0.7B active parameters, supports speculative decoding when enabled by the provider.
DeepSeek-V4-Pro
This model lists 1.6T total parameters and 49B active per token. Its headline structural features include token-wise compression and DSA (DeepSeek Sparse Attention). DeepSeek describes these techniques as a way to improve long-context efficiency while reducing compute and memory requirements.
Pricing Compared (per 1M tokens)
| Model | Input | Output | Pricing model |
|---|---|---|---|
| Hy4 Preview | $0.834 | $2.501 | Flat rate |
| DeepSeek-V4-Pro | $0.66 off-peak / $1.32 peak | $1.98 / $3.96 | Peak / off-peak |
| DeepSeek-V4-Flash | $0.22 / $0.44 | $0.66 / $1.32 | Peak / off-peak |
On the listed cache-miss rates, DeepSeek is cheaper: V4-Pro off-peak input is $0.66 versus Hy4's $0.834, and V4-Flash is lower still. The comparison is not entirely direct because DeepSeek splits peak and off-peak pricing while Hy4 uses one flat rate. During DeepSeek peak hours, V4-Pro input rises to $1.32, above Hy4's listed rate. Cache-hit pricing can also materially change the effective cost.
Note: verify current prices and peak-hour definitions on the provider's official pricing page before deploying a production workload.
Output Limit and Open-Source Status
Maximum output is DeepSeek's clear advantage: 384K tokens versus Hy4's 64K. For workflows that generate very long artifacts, that larger ceiling provides more headroom, although applications should still set practical output limits.
Open weights are another major difference: the supplied comparison lists both V4-Pro and V4-Flash as available on HuggingFace. Hy4 Preview is currently a hosted API with model ID tencent/hy4-preview. If self-hosting or fine-tuning is a requirement, confirm the latest release and license status before choosing a service.
Bottom Line: Which Should You Choose?
| Your situation | Relevant option |
|---|---|
| Need open weights, self-hosting, or fine-tuning | DeepSeek-V4-Pro / Flash |
| Need very long output above 64K | DeepSeek-V4-Pro / Flash |
| Lowest listed cost per token | DeepSeek-V4-Flash |
| Want Tencent ecosystem integration and flat pricing | Hy4 Preview |
| Want detailed expert architecture and native MTP documentation | Hy4 Preview |
Hy4 Preview is still in grey-testing, so its positioning and pricing may change. DeepSeek V4 is listed as generally available. The right choice depends on whether your priority is open weights, output length, cost controls, ecosystem integration, or a hosted API workflow.
Get Hy4 API Access on TokenRa →Frequently Asked Questions
Is Hy4 better than DeepSeek V4?
They target different strengths. Hy4 Preview has a more detailed expert architecture and a simpler flat pricing model, while DeepSeek-V4-Pro is open-sourced, supports a much larger maximum output, and is cheaper on input off-peak. Which is better depends on your workload and whether you need open weights.
How do Hy4 and DeepSeek V4 parameters compare?
Both activate 49B parameters per token. Hy4 Preview has 770B total parameters, while DeepSeek-V4-Pro has 1.6T total parameters. DeepSeek-V4-Flash is smaller at 284B total and 13B active.
Is Hy4 cheaper than DeepSeek?
No. Hy4 Preview is priced at $0.834 per 1M input and $2.501 per 1M output. DeepSeek-V4-Pro off-peak is $0.66 input and $1.98 output, while DeepSeek-V4-Flash is $0.22 input and $0.66 output. DeepSeek uses peak and off-peak pricing while Hy4 uses a single flat rate.
Is DeepSeek V4 open source?
Yes. DeepSeek-V4-Pro and V4-Flash are open-weight models available on HuggingFace. Hy4 Preview is currently served only as a hosted API with model ID tencent/hy4-preview and has not been open-sourced.