Tencent Hy3 vs DeepSeek V4: Benchmarks, Pricing, and Real Performance Compared
Tencent Hy3 vs DeepSeek V4 Flash and Pro — full benchmark breakdown, pricing, context windows, and licensing compared across both DeepSeek tiers. Which open-weight model actually wins?
Tencent rebuilt its training infrastructure from scratch and shipped Hy3 in under 90 days — and it's now beating DeepSeek's flagship V4 Pro on more benchmarks than it loses. But DeepSeek didn't just ship one model this cycle; it shipped two. Here's how Hy3 stacks up against both the budget-tier V4 Flash and the heavyweight V4 Pro.
Meet the Contenders
Tencent Hy3 is a 295-billion-parameter Mixture-of-Experts model with only 21 billion parameters active at inference time. It ships open-weight under Apache 2.0 licensing, supports configurable reasoning depth (disabled, low, and high chain-of-thought modes), and targets agentic workflows, coding, and long-horizon tasks with a 262K-token context window.
DeepSeek V4 Flash is DeepSeek's default, low-cost production model — fast, efficient, and built for everyday chat, extraction, classification, and lightweight coding, with a full 1M-token context window included at no extra charge.
DeepSeek V4 Pro is the heavyweight of the pair: a 1.6-trillion-parameter MoE model with roughly 49 billion active parameters per token, using a hybrid attention architecture (CSA + HCA) that makes million-token inference economically viable. It's MIT-licensed, meaning unrestricted fine-tuning, commercialization, and redistribution.
Architecture at a Glance
| Spec | Hy3 | DeepSeek V4 Flash | DeepSeek V4 Pro |
|---|---|---|---|
| Total Parameters | 295B | Smaller, undisclosed exact size | 1.6T |
| Active Parameters | 21B | — | 49B |
| Context Window | 262K | 1M | 1M |
| Max Output | 262K | 384K | 384K |
| License | Apache 2.0 | Proprietary | MIT |
| Reasoning Modes | Disabled / Low / High | Thinking / Non-thinking | Thinking / Non-thinking |
The parameter gap alone tells a story: DeepSeek V4 Pro is roughly 5.4x larger in total parameters and 2.3x larger in active parameters per token than Hy3 — yet Hy3 still keeps pace on more than half of their shared benchmarks. That's the entire narrative of this comparison in one sentence: Hy3 is the efficiency play, and both DeepSeek tiers are the scale play.
Round 1: Hy3 vs DeepSeek V4 Pro (The Heavyweight Matchup)
This is the more dramatic result of the two comparisons. Across 18 shared benchmarks, Hy3 won 12 — despite running on a fraction of the parameters.
| Benchmark | Hy3 | DeepSeek V4 Pro | Winner |
|---|---|---|---|
| SWE-bench Verified | 78.0% | 80.6% | V4 Pro (+2.6 pts) |
| SWE-bench Pro | 57.9% | 55.4% | Hy3 (+2.5 pts) |
| FrontierScience (Olympiad) | 74.8% | 74% (combined) | Roughly tied |
| Output Speed | 2.6x faster | Baseline | Hy3 |
| Repeated-context caching | Standard | $0.0036/M (16.5x cheaper) | V4 Pro |
The SWE-bench split is the most telling result here. On SWE-bench Verified — the standard real-world bug-fixing benchmark — DeepSeek V4 Pro takes a real but modest 2.6-point lead. On SWE-bench Pro, a harder and arguably more representative test of production coding ability, Hy3 actually pulls ahead. Neither model dominates the other on coding; they trade wins depending on which flavor of the benchmark you weight more heavily.
Where Hy3 pulls clearly ahead is efficiency: it's 2.6x faster at output generation while using less than half the active parameters of V4 Pro. Where V4 Pro pulls clearly ahead is long-context economics — its disk-based caching drops repeated-context costs to $0.0036 per million tokens, a rate Hy3 simply can't match, plus a full 4x larger context window (1M vs 262K).
Round 2: Hy3 vs DeepSeek V4 Flash (The Budget Matchup)
This comparison flips the framing: Flash isn't trying to be the smartest model, it's trying to be the cheapest capable one — and Hy3 is playing the exact same game.
| Metric | Hy3 | DeepSeek V4 Flash | Winner |
|---|---|---|---|
| BenchLM composite score | 55 | 40 | Hy3 (+15 pts) |
| Knowledge tasks (avg) | 33.8 | 56.5 | V4 Flash |
| Reasoning | Baseline | +0.6 pts | V4 Flash (marginal) |
| Output speed | 1.7x faster | Baseline | Hy3 |
| Context window | 262K | 1M | V4 Flash |
| Input pricing | Slightly cheaper | — | Hy3 (1.1x) |
Hy3 leads the general BenchLM composite score by a wide margin — 55 versus 40 — largely on the strength of being the reasoning-capable model in this pairing (Flash runs non-reasoning by default in this comparison set). But Flash claws back ground specifically on knowledge-intensive tasks, where it averages 56.5 versus Hy3's 33.8, and it edges ahead on general reasoning by a slim 0.6 points when both are pushed to their limits.
The practical takeaway: pick Hy3 if you want the stronger general benchmark profile and faster output; pick DeepSeek V4 Flash if knowledge-heavy tasks or a massive 1M-token context window are the priority.
Pricing: A Genuinely Complicated Picture
Pricing for both models has moved around a lot since launch, and figures vary by provider, so treat this as a snapshot rather than gospel:
| Model | Input (per 1M) | Output (per 1M) | Context |
|---|---|---|---|
| Hy3 (OpenRouter, standard) | $0.14 | $0.58 | 262K |
| Hy3 preview (OpenRouter) | $0.063 | $0.21 | 262K |
| DeepSeek V4 Flash | $0.14 | $0.28 | 1M |
| DeepSeek V4 Pro | $0.435 | $0.87 | 1M |
A few important notes:
- Tencent has run limited-time free access to Hy3 on OpenRouter, and its own Tencent Cloud TokenHub pricing (roughly 1 yuan input / 4 yuan output per million tokens) works out to figures broadly in the same range as the OpenRouter numbers above.
- DeepSeek automatically caches repeated prompt prefixes, dropping input costs by 98% on Flash and roughly 92% on Pro when you hit a cache — a meaningful advantage for production workloads with repeated system prompts or shared context.
- Both DeepSeek tiers run in "thinking mode" by default, which bills reasoning tokens at the output rate — a common source of unexpectedly high bills if you don't explicitly disable it for simple tasks.
On a pure sticker-price basis, DeepSeek V4 Flash and Hy3 land in a similar ballpark on input cost, while V4 Pro costs roughly 3x more than either on output — the price you pay for its much larger parameter count and superior long-context caching economics.
So Which One Should You Actually Use?
- Choose Hy3 if: you want the strongest all-around open-weight performance per dollar, need fast output generation, or want a genuinely open (Apache 2.0) model you can self-host and fine-tune without restriction.
- Choose DeepSeek V4 Flash if: your workload is knowledge-heavy, you need the full 1M context window on a budget, or you're running high-volume routine tasks like chat, extraction, and classification.
- Choose DeepSeek V4 Pro if: you're doing complex, long-context work with heavy repeated context (where its caching economics shine), or you specifically need MIT licensing for unrestricted commercial redistribution.
Bottom Line
Hy3 is the standout story of this comparison — not because it dominates outright, but because a 295B/21B-active model consistently trading blows with (and often beating) a 1.6T-parameter flagship is a genuinely rare result in this market. DeepSeek still wins on raw context window size and long-context caching economics across both tiers, and V4 Flash remains the stronger pick specifically for knowledge-dense workloads. But for teams optimizing for coding performance, output speed, and license freedom per dollar spent, Hy3 has quickly become one of the most credible open-weight options available right now.