Tencent's Hy3 Beats Flagship AI Models Without a Trillion Parameters

Tencent just shipped Hy3, the full release of its 295B Mixture-of-Experts language model, and it is making a pointed argument: you do not need a trillion-parameter model to get flagship-level results. The model is open-weight under Apache 2.0, available on Hugging Face right now, and free to call via API on OpenRouter for the next two weeks.

From preview to production

Hy3 is a 295B-parameter MoE model with 21B active parameters and a 3.8B MTP layer. Following the Hy3 Preview launch in late April, the team gathered feedback from 50+ product teams, fixed various issues in task execution and interaction, and improved both the quality and scale of the post-training pipeline. The full Hy3 release is the result of that iteration cycle.

Building on the Preview, Tencent improved post-training data quality and diversity while scaling up RL training. Hy3 shows solid gains across reasoning, agentic workflows, and long-context tasks, with performance close to leading flagship models both domestic and international.

The architecture doing the heavy lifting

MoE (Mixture-of-Experts) means the model has 295B total parameters but only activates a small fraction per token. Top-8 out of 192 experts are activated, routing routine queries to fast pattern-matching experts and complex problems to deeper reasoning chains. This is how Hy3 keeps compute costs low while maintaining quality.

Hy3 breaks the assumption that all experts cost the same compute when active. Some experts are wider and handle general token routing across domains; others are narrower and specialize in reasoning-dense subproblems like multi-step math or tool call sequencing. The router learns to send tokens to appropriately sized experts rather than uniformly sized ones, so 21B active parameters produce output quality normally associated with 40-70B active parameter models.

The model also ships with a dedicated MTP (Multi-Token Prediction) layer. Instead of predicting one token per forward pass, the 3.8B MTP head drafts several tokens simultaneously. This integrates directly with vLLM's speculative decoding: the MTP head acts as the draft model, the main model verifies the drafts, and accepted drafts skip redundant forward passes. On 8x H200 SXM5 hardware, Hy3 generates approximately 2,000-3,000 tokens per second, and with MTP speculative decoding enabled, throughput reaches roughly 3,000-4,500 tok/s.

The model also supports a configurable reasoning mode via the reasoning_effort parameter. The architecture fuses fast and slow thinking into a single model rather than splitting them across separate models as the previous generation did. Set it to "no_think" for direct responses, "low" for light chain-of-thought, or "high" for deep multi-step reasoning on hard math and coding problems.

from openai import OpenAI
client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="EMPTY")
response = client.chat.completions.create(
    model="hy3",
    messages=[{"role": "user", "content": "Refactor this function..."}],
    temperature=0.9,
    top_p=1.0,
    # "no_think" = direct, "low" = light CoT, "high" = deep reasoning
    extra_body={"chat_template_kwargs": {"reasoning_effort": "high"}},
)
print(response.choices[0].message.content)

Production reliability, not just benchmark scores

Tencent's framing here is deliberately practical. Rather than relying solely on public benchmarks, they ran a blind test with 270 experts from various disciplines working on real-world workflows, collecting 312 valid comparisons. Hy3 scored 2.67/4, outperforming GLM-5.1 at 2.51/4, with the clearest advantage in frontend development, CI/CD, and data and storage tasks.

Three specific reliability improvements stand out for anyone building agents:

• Tool call stability: Multiple baseline reliability issues were fixed, bringing the model to production-grade standards across tool configurations and output constraints. Tool-call success rates and error recovery improved, and invalid calls that trigger infinite loops dropped. Hy3 also generalizes across different agent scaffoldings; on SWE-Bench Verified, accuracy variance across CodeBuddy, Cline, and KiloCode remains within 4%.
• Hallucination reduction: Internal evaluations show Hy3's hallucination rate dropped from 12.5% to 5.4%, and commonsense error rates fell from 25.4% to 12.7%. The guiding principle is explicit: answer when grounded, flag when evidence is missing, never fabricate.
• Multi-turn coherence: On internal comprehensive multi-turn tests, the issue rate dropped from 17.4% to 7.9%, and on the open-source MRCR long-dialogue benchmark, scores jumped from 42.9% to 75.1%.

Where it shines and where to watch out

In productivity scenarios such as coding, office work, financial modeling, frontend design, and game development, Hy3 has made remarkable progress and can serve as a reliable, cost-effective model option. The Preview's SWE-bench Verified score of 74.4% already put it above DeepSeek-V3 on coding agents, and the full Hy3 release pushes further.

The main practical constraint is hardware. Hy3 has 295B parameters in total; to serve it on 8 GPUs, Tencent recommends using H20-3e or other GPUs with larger memory capacity. For teams without that infrastructure, the free OpenRouter API endpoint is the obvious starting point. Self-hosted Hy3 at spot pricing runs $0.90-$1.62 per million output tokens, roughly a 6-11x cost reduction compared to GPT-4o at $10/M.

How to get started

There are three practical paths depending on your setup:

1. Free API via OpenRouter: Hit tencent/hy3:free on OpenRouter with any OpenAI-compatible SDK. No setup required, free for two weeks.
2. Self-hosted with vLLM or SGLang: Pull weights from Hugging Face and serve with tensor parallelism across 8 GPUs. Both frameworks support the MTP speculative decoding layer natively.
3. Quantized variant: Tencent also open-sources Hy3-FP8 model weights, an FP8-quantized version that reduces memory footprint while preserving most of the capability.

The bigger picture

The open-weight frontier is getting crowded at the top. Kimi-K2, GLM-5, and now Hy3 are all MoE models in the 200-300B total parameter range that claim near-parity with proprietary flagships. What makes Hy3's release notable is the explicit focus on production reliability over raw benchmark numbers. Hy3 was already running inside Yuanbao, CodeBuddy, WorkBuddy, ima, Tencent Docs, and Peacekeeper Elite before public launch, meaning the reliability improvements were shaped by real product pressure, not synthetic evals.

Hy3 is released under the Apache License 2.0, a meaningful upgrade from the Preview's more restrictive community license. For teams evaluating open-weight models for commercial agentic pipelines, that combination of permissive licensing, production-tested reliability, and a free two-week trial makes Hy3 worth a serious look.