Tencent’s pitch is that smaller comes close enough to flagship to matter. Hunyuan Hy3, open-sourced in early July 2026 after an April Preview, is a 295-billion-parameter Mixture-of-Experts model that activates only 21 billion parameters per token. Tencent’s Hy3 repository claims it “rivals flagship open-source models with 2-5x parameters.” Those benchmarks are vendor-reported and internal, so the direction is interesting and the magnitudes are not settled.
How is Hunyuan Hy3 architected?
Hy3 is a sparse Mixture-of-Experts model: 295 billion total parameters, 21 billion active per token, plus a 3.8-billion-parameter Multi-Token Prediction (MTP) layer for faster decoding, with a 256K-token context window. The figures come straight from the Hy3 GitHub repo.
MoE is the whole efficiency story in miniature. Every token is routed to a small subset of expert subnetworks rather than passing through the full network, so the model keeps a 295-billion-parameter knowledge reservoir but pays compute only for the 21-billion-parameter slice that fires. DeepSeek’s architecture popularized the same lever; Hy3 is another data point that sparse routing is now the default for large Chinese labs rather than a novelty.
The tradeoff matters operationally. With MoE you pay for total parameters in memory and for active parameters in compute: all 295B of expert weights have to sit in HBM, but each forward pass only does roughly 21B parameters’ worth of work. A serving team provisioning for Hy3 needs enough accelerator memory to hold the full model, then earns back the cost on every token through the lower active footprint. That is the concrete economics behind the “smaller” framing, and it is the number a deployment planner will care about more than the headline 295B.
The 3.8-billion-parameter MTP head is the speed play. Multi-Token Prediction trains an auxiliary network to propose several future tokens at once; at inference time the main model verifies those draft tokens in a single forward pass and accepts the ones that match. Speculative decoding layered on top of MTP cuts the number of sequential passes needed to emit a given response, lowering per-token latency without changing the output distribution. Tencent frames MTP as an efficiency feature; the repo does not publish acceptance rates or measured speedups, so treat the latency gains as architectural intent rather than a benchmarked result.
Does Hy3 actually match flagship performance?
On Tencent’s own blind evaluation, yes. On any independent benchmark, the question is open.
The Hy3 repo reports a blind evaluation in which 270 domain experts scored real-world tasks without knowing which model produced each output. Hy3 averaged 2.67 out of 4 against GLM-5.1 at 2.51 out of 4. Tencent also reports that during development Hy3’s hallucination rate dropped from 12.5% to 5.4% and commonsense error rates fell from 25.4% to 12.7%. Both sets of figures are vendor-reported and self-graded, and the repo does not publish the task prompts, the expert selection criteria, or the rubric the graders used.
The direction is the part worth taking seriously. Hallucination and commonsense-error rates falling over training is the expected outcome of more reinforcement learning and better post-training data; the magnitude (roughly halving in both cases) is plausible but unverifiable from the repo. What the numbers do not give you is a cross-comparable score against the models Hy3 is implicitly measured against. GLM-5.1 is named, but the flagship open-weight models with two-to-five times more parameters that Tencent says Hy3 “rivals” are not identified by name or by score in the published material.
How does Hy3 stack up against DeepSeek V4 and Kimi K3?
Structurally, there is a real gap. Measured head-to-head, there is nothing published to compare.
Hy3 sits at 295 billion total parameters and 21 billion active, so the trillion-scale models it positions against are larger by roughly an order of magnitude. DeepSeek’s V4 Preview, which appeared around July 2026, is marketed for stronger agent capabilities and reasoning; the 1.6-trillion-parameter figure associated with DeepSeek V4 appears in industry reporting but is not confirmed in the primary sources reviewed here [unverified]. Kimi K3, attributed to Moonshot at roughly 2.5 trillion parameters [unverified; no primary source in this brief], represents the upper end of the scale race. Neither has published a direct comparison against Hy3, and Tencent has not published scores against either.
| Model | Total params | Active params | Context | Source status |
|---|---|---|---|---|
| Hunyuan Hy3 | 295B | 21B/token | 256K | Vendor repo |
| DeepSeek V4 | ~1.6T [unverified] | Not disclosed | Not disclosed | Vendor site; param count unconfirmed |
| Kimi K3 | ~2.5T [unverified] | Not disclosed | Not disclosed | No primary source in brief |
The table makes the asymmetry obvious: Hy3’s specs are fully published, the others are not. That is itself a competitive argument. When a lab open-sources a 295B/21B model under Apache 2.0 with weights on Hugging Face, ModelScope, GitCode, and CNB, and the rivals it is chasing keep their internals opaque, the smaller model can win on deployability even where it trails on raw capability. You can run Hy3, audit its architecture, and build a serving stack around a known active-parameter budget. The rivals are, for now, numbers in a press cycle.
Why are Chinese labs betting on smaller, efficient models?
Efficiency is the constraint, not the preference. The structural incentives point the same way regardless of which lab you look at.
Compute is finite, advanced accelerators are subject to US export controls, and domestic cloud economics reward lower per-token serving cost. A model that activates 21 billion parameters per token costs less to serve than a dense or near-dense trillion-parameter model, and the gap shows up directly in GPU hours and in the number of concurrent users a fixed cluster can sustain. Inside China, where the addressable market is large and the GPU supply is tighter than in the US, the active-parameter count is arguably a more important number than the total.
The competitive framing matters here. As long as the headline metric was raw parameter count, DeepSeek V4 and Kimi K3 owned the narrative; bigger won by definition. Hy3’s release tries to move the conversation to size-to-performance ratio, where a 295B model that matches a 1T model on practical tasks is the more interesting result. If even one major Chinese lab starts publishing efficiency benchmarks alongside accuracy scores, the rest face pressure to follow, because “matches flagship at a fraction of the active parameters” is a harder claim to ignore than yet another parameter-count record.
Post-training is the other lever, and the one hardest to inspect from outside. Tencent’s Hy research blog frames Hy3 as a product of aggressive reinforcement learning and data curation rather than raw scale. If the thesis holds, capability parity with larger models comes from training quality, not parameter count, which is precisely the dimension competitors cannot read off a spec sheet. It is also the dimension most likely to be oversold.
The export-control pressure shows up second-hand. The harder it is to import top-tier accelerators, the more a lab gains from a model that fits on the hardware it can actually buy. Efficiency becomes a moat precisely when scale becomes expensive.
Should you trust the Hy3 benchmarks?
No, not as settled fact, and you should not deploy on the strength of them alone.
What is missing is the standard public benchmark ladder. The repo and the Tencent Hy research blog do not report MMLU, GPQA, C-Eval, or independent leaderboard placements alongside the internal eval. The blog discusses a CL-Bench suite for complex knowledge learning tied to the Hy3 Preview, but does not publish comparable numbers for rival models. Absent those, the only way to validate Hy3’s capability claims is to run it yourself, which the licensing makes straightforward: Apache 2.0, with permissive redistribution and weights available on the major open-weight hubs. There are no production-use warranties, which is standard for an open-weight release but worth stating plainly.
The steelman for Hy3 is narrow but real. If a 21B-active MoE model holds its own on genuine production traffic, the implication for inference cost is concrete and the implication for the scale race is uncomfortable. The caveat is that “if.” Tencent has published a model and a set of internal numbers; it has not published the evidence that would let anyone else confirm them. Until independent evaluations land, Hy3 is best read as a credible efficiency bet and an unproven benchmark champion.
Frequently Asked Questions
Which teams can actually download and use Hy3 today?
Anyone can, because Tencent released the weights under Apache 2.0 on Hugging Face, ModelScope, GitCode, and CNB. The CNB and GitCode mirrors matter most for mainland Chinese teams that may have slow or blocked access to Hugging Face, though the lack of production warranties still makes it an evaluation release.
How is Hy3 different from Tencent’s other Hunyuan models?
Hy3 is a text MoE; it is not the same product as Hunyuan Video, a 13-billion-parameter video-generation model, or Hunyuan 3D. Those are separate Tencent AI releases with their own weights and use cases, so benchmark numbers from one do not transfer to the other.
What infrastructure changes does Hy3 require compared with a dense model?
Serving teams must load the full 295B expert weights plus the 3.8B MTP head into accelerator memory, then integrate speculative decoding and MoE routing. If the serving stack cannot accept draft tokens or if all-to-all communication between expert shards becomes a bottleneck, the 21B active-parameter advantage can shrink or disappear.
What would make the Hy3 claims less impressive?
If independent evaluators run MMLU, GPQA, or C-Eval and find Hy3 trailing larger rivals, or if DeepSeek V4 publishes direct head-to-head comparisons, the internal 2.67/4 score loses much of its weight. Easing of US export controls on accelerators would also reduce the cost pressure that makes 21B active parameters attractive.
