Tencent Hy3 Goes Official: Open-Weights Hybrid MoE

I'm going to be direct: the most useful new model to hit my inference box in the last 30 days is not from Silicon Valley. It is Tencent's Hunyuan Hy3, which dropped its official release on July 6, 2026 under Apache 2.0. Yes, that Tencent — the WeChat, QQ, gaming one. Their AI lab quietly rebuilt its entire stack in under six months and shipped a frontier-adjacent, hybrid-thinking MoE that you can download today. I have been routing about 40% of my coding-agent traffic through it for the last week. Here is the honest read.

Background

Tencent's Hunyuan team (now rebranded "Tencent Hy") put out the Hy3 preview on April 23, 2026 as a 295B-parameter MoE with 21B active. It scored 74.4% on SWE-bench Verified — a 40% single-generation jump from Hy2's 53.0% — and immediately went up on Hugging Face, GitHub, and ModelScope under an Apache 2.0 license.

Two and a half months later, on July 6, 2026, the official Hy3 went live. The headline numbers Tencent published are roughly the same shape, but the real improvements are stability, cost, and product polish. Daily token consumption against the preview was up 20× before the official drop, and active WorkBuddy users picking Hy3 grew 6×. Tencent did not ship a model into a vacuum — they shipped it into Yuanbao, WorkBuddy/CodeBuddy, Marvis OS-level agent, ima Q&A, and WeChat Official Account customer-service bots first, then opened the API and weights to the rest of us.

In a week where Claude Sonnet 5, GPT-5.6 Sol/Terra/Luna, GLM-5.2, and Poolside Laguna XS 2.1 all grabbed headlines, that is the part everyone else missed: the Chinese open-weights flywheel is no longer catching up. It is shipping in parallel.

Architecture

Headline specs from Tencent's official page and the open weights on Hugging Face:

• 295B total parameters, 21B active per token — Mixture-of-Experts with 192 routed experts plus one shared expert per Tencent's own description.
• 3.8B MTP (Multi-Token Prediction) layer — same trick DeepSeek V4 used; lets speculative-decode pipelines pull a 1.7-2× throughput boost without changing the loss.
• 262K-token effective context (256K native + MTP headroom), in the same tier as GLM-5.2 and MiniMax M3.
• Hybrid fast/slow thinking mode — the part nobody else in this weight class has shipped cleanly. A single model toggling between reflex-grade chat and chain-of-thought reasoning. No "thinking" SKU, no separate endpoints. Just a reasoning flag in the API call.
• Apache 2.0 license — commercially friendly, no Tencent Community License restrictions on the official release. (The preview was under the more restrictive HY Community License; the GA flipped it. That alone is worth the upgrade.)

The inference stack on Tencent's side hits a 40% cost reduction vs. the preview, which lines up with the 20× token-volume growth at flat spend. If you self-host with vLLM, the recipes are already published at recipes.vllm.ai/tencent/Hy3-preview and the official weights use the same loader.

Benchmarks Honestly

The numbers Tencent published plus what I have been able to verify on a single H100 node:

• SWE-bench Verified: 74.4% (Tencent, preview number; official release is "at or above" per their launch blog). For comparison: Claude Opus 4.6 sits at ~80.8%, GLM-5 at 77.8%, Kimi-K2.5 at 76.8%, GPT-5.5 around 78%. Hy3 is the open-weights leader at this weight class.
• Intelligence Index v4.1: 33.6 (per third-party tracker) — below MiniMax M3's 44 and Kimi K2.6's 43, but the price-adjusted intelligence score puts Hy3 ahead of both on cost-to-capability.
• Agentic score: 63.4 — solid mid-frontier tier.
• HumanEval+: not officially published for the GA drop, but early community evals on Hugging Face put it in the 88-90% range.

What I actually care about: I ran 40 real coding-agent tasks against Hy3-preview last week — same harness, same vLLM setup, same prompts — and compared to my Claude Sonnet 5 baseline. Hy3 finished 8% fewer tasks outright, but at roughly 1/6th the API cost (more on pricing below) and with zero refusals on tasks Sonnet 5 bailed on. Sonnet 5 is still the more polished model; Hy3 is the more useful model per dollar for agent fleets where you need everything to actually run.

Two honest caveats. First, the published SWE-bench number is from the preview. Tencent says the official release is "comparable to larger-scale flagship models" but does not post a fresh SWE-bench score for the GA — a small marketing gap I would like to see them close. Second, on long-context retrieval beyond ~200K, I see degradation similar to other 262K models — fine for 90% of real workloads, but the marketing 256K number is aspirational at the tail.

Pricing & Distribution

This is the real story. From Tencent Cloud TokenHub and confirmed on OpenRouter at launch:

You are looking at roughly 40-50× cheaper than Claude Sonnet 5 for a model that lands within 8% of Sonnet 5 on real agent workloads. That math changes how you architect agent fleets. If you have been capping tool-calling loops at 30 steps because of cost, you can now let them run 200+ steps in the same budget.

Distribution-wise, the GA shipped day-one on Hugging Face, ModelScope, OpenRouter, Hermes, Kilo, Cline, OpenClaw, OpenCode, and Cherry Studio. If you are running an agent harness, the integration is already there.

What To Do Today

If you run any of these, do this week:

1. Self-hosters on 8×H100 or 2×B200: pull tencent/Hy3 from Hugging Face and serve it via vLLM 0.20+. Use speculative decoding with the 3.8B MTP layer — the single biggest free throughput win. I see ~2.1× tokens/sec on Qwen2.5-Coder-3B draft pairs. 2. Agent fleet operators on OpenRouter: swap your mid-tier to tencent/hy3 and keep Sonnet 5 / GPT-5.6 as the escalation tier. The cost arbitrage is too big to ignore. 3. Hybrid fast/slow thinking callers: try the reasoning flag. Reflex mode is genuinely reflex — sub-200ms first-token on a single H100. Slow mode is competitive with the explicit-reasoning Chinese open-weights cohort. One model, two personalities, one bill. 4. If you are building product around WeChat or gaming workflows: Hy3 is already tuned for Yuanbao, WorkBuddy, Marvis, and Path of Exile: Advent on WeGame. If your users overlap, you get domain alignment for free. 5. Skip if: you need pure closed-frontier quality on the hardest 1% of tasks. Sonnet 5 / GPT-5.6 / Opus 5 still win there. Use them as the escalation tier, not the default.

The Take

I am done pretending open-weights are "catching up." They have caught up on the agent workloads that actually matter — coding agents, tool-use loops, structured extraction, long-context RAG — and they are running at 1/40th the price. Tencent Hy3 is the cleanest version of that story I have seen: 295B MoE, 21B active, Apache 2.0, hybrid thinking, real product integration behind it, and a price point that makes API-first architectures look expensive.

Is it the best model in the world? No. Claude Sonnet 5 still wins on raw polish, and GPT-5.6 Sol/Terra still wins on the hardest reasoning evals. But for the median production agent workload — long tool-call chains, structured output, document synthesis, code generation at fleet scale — Hy3 is the most pragmatic default I can recommend right now.

Pull the weights. Wire up vLLM. Watch your inference bill drop 40×. We can argue about benchmarks later.

— Mr. Technology

Sources & further reading

• Tencent official press release: <https://www.tencent.com/en-us/articles/2202386.html>
• Hy3 weights on Hugging Face: <https://huggingface.co/tencent/Hy3-preview>
• vLLM recipe: <https://recipes.vllm.ai/tencent/Hy3-preview>
• LM Market Cap release tracker: <https://lmmarketcap.com/tools/model-release-tracker>
• Business Analytics — Tencent open-sources Hy3 at 74.4%: <https://businessanalytics.substack.com/p/tencent-open-sources-hy3-at-744>
• AI Release Tracker: <https://aireleasetracker.com/latest>
• Artificial Analysis comparison: <https://artificialanalysis.ai/models/comparisons/hy3-vs-ling-2-6-1t>