Tencent Hunyuan Hy3 Ships as a 295B Apache 2.0 Open-Weight Model

Photo by Kier in Sight Archives on Unsplash

Tencent's Hy3 open source model landed on July 6, 2026 as a 295-billion-parameter mixture-of-experts release under the Apache 2.0 license, a licensing shift that matters more to most developers than the benchmark numbers attached to it. The April 2026 preview shipped under restricted licensing; the July 6 full release reversed that, publishing weights under a permissive license that researchers on X called the real headline of the launch.

That detail reframes the entire release. A 295B MoE with competitive frontier-class scores is interesting. A 295B MoE with competitive frontier-class scores that any company can fine-tune, redistribute, and ship inside a commercial product without a bespoke license negotiation is a different category of event.

What the Hunyuan Hy3 Apache 2.0 Release Actually Includes

The architecture is engineered for inference economics rather than raw parameter bragging rights. Hy3 is a 295B-parameter MoE with 21B active parameters per forward pass, routed via top-8 selection across 192 experts, plus a 3.8B multi-token prediction (MTP) layer for speculative decoding, and a 256K context window. Only 21B parameters fire per token, which is why serving costs collapse relative to dense models of similar quality.

i-scoop's technical write-up fills in the transformer detail: 64 layers, grouped-query attention, 192 experts with top-8 routing, a 3.8B MTP layer on top to accelerate generation, and a reasoning_effort setting that configures reasoning depth per request. The per-request reasoning switch is the piece most teams will care about in production. Long chain-of-thought is expensive. Being able to dial it down per call, rather than routing between two model endpoints, cuts inference latency for the 80 percent of queries that do not need deep reasoning.

Hy3 is available under the Apache 2.0 license, which permits commercial use, modification, and redistribution with only attribution and patent-grant obligations. That is the same license umbrella under which most enterprise-friendly open source projects ship. For teams that had to route around Meta's Llama community license or Alibaba's Qwen terms, this removes a legal review step.

Tencent 295B MoE Model Benchmarks Against GPT-5.5 and GLM-5.2

Hy3 leads open weights on agentic search and long-context retrieval and loses to a bigger Chinese competitor on coding. According to VentureBeat's benchmark tabulation, on agentic search benchmarks Hy3 scores 84.2 on BrowseComp and 91.0 on DeepSearchQA, ahead of every open model in Tencent's evaluation table and competitive with Claude Opus 4.8 and GPT-5.5. The same source reports Hy3 leads open models on tool orchestration (79.1 on MCP-Atlas) and long-context retrieval (73.4 on AA-LCR), but concedes repository-scale coding to GLM-5.2.

The size framing matters. GLM-5.2 is a roughly 744B MoE with about 40B active parameters versus Hy3's 295B total and 21B active, and Hy3 trails on SWE-bench Pro at 57.9 despite that efficiency advantage. Hy3 is smaller and cheaper to serve. It is also worse at large-codebase agentic tasks than the model it most directly competes with in the Chinese open-weight tier.

STEM numbers put Hy3 in reach of the closed frontier without matching it outright. BiggoNews' launch coverage lists a GPQA Diamond score of 90.4 against GPT-5.5's 93.6, HLE with tools at 53.2 (exceeding DeepSeek-V4-Pro's 48.2), and a FrontierScience-Olympiad score above GPT-5.5. i-scoop's version puts the FrontierScience-Olympiad numbers at 74.8 for Hy3, 72.5 for GLM-5.2, and 73.8 for GPT-5.5.

The advanced math gap is the hard one to explain away. GPT-5.5 leads MathArena Apex at 85.4 against Hy3's 38.7, and Claude Opus 4.8 leads SWE-bench Pro at 69.2 against Hy3's 57.9. Coding and math are the two workloads where a two-week free API window will not paper over the deficit.

The Blind Eval That Reads Better Than the Benchmarks

Benchmark contamination is now assumed. What is harder to fake is a rater study. i-scoop reports Tencent ran a blind evaluation with 270 domain experts in which Hy3 scored 2.67 out of 4 versus GLM-5.1's 2.51, and internal hallucination rate dropped from 12.5% at preview to 5.4% at general availability. That hallucination reduction is the number worth watching in production.

The H20-3e Question Nobody at Tencent Wants to Answer Directly

Here is the part of the release that competitor coverage keeps missing. Tencent's recommended self-hosting configuration targets Nvidia's H20-3e, the GPU Nvidia designed specifically to comply with U.S. export restrictions on China. An Apache-licensed frontier-adjacent Chinese model whose reference deployment guide targets the export-compliant Nvidia SKU is a policy artifact as much as a technical one. It tells you where the training was done, which hardware Tencent expects Chinese enterprises to buy, and how the cost curve for the rest of 2026 gets drawn.

Western deployers get a different tradeoff. Ayautomate's operational writeup notes that self-hosting requires multi-GPU nodes because the 295B MoE keeps all parameters resident in VRAM, and the FP8 build carries a footprint under 300GB, which is the practical choice. i-scoop confirms the FP8 checkpoint roughly halves memory footprint versus BF16, making self-hosting on a single 8x H20-3e node feasible. On H100s or MI300Xs the math still works, but the reference config was not written for that hardware.

For teams already navigating hardware constraints, the parallel with what Intel is attempting with Crescent Island's 480GB LPDDR5X design is worth noting. Memory capacity, not compute, is now the binding constraint for serving models this size.

How to Actually Run It This Week

For two weeks the meter is off. Ayautomate reports free API access runs through July 21, 2026 via the OpenRouter tencent/hy3:free endpoint, after which standard OpenRouter pricing applies, and Tencent had not announced post-window pricing publicly as of July 7, 2026. Context is 256K tokens per the HuggingFace model card, though OpenRouter displays 262K due to rounding convention differences.

The self-hosting path is unusually clean for a model this size. Hy3 works out of the box with vLLM and SGLang, exposes an OpenAI-compatible API, and supports fine-tuning via LLaMA-Factory with DeepSpeed ZeRO. If your stack already runs Qwen or DeepSeek, the swap is a config change rather than a rebuild. That matters for anyone building agentic workflows with tool use, because the OpenAI-compatible surface means existing agent frameworks work without rewrites.

For teams that would rather rent than rack, Tencent Cloud published its own rate card. BiggoNews lists 1 yuan per million input tokens, 4 yuan per million output tokens, and 0.25 yuan per million cache-hit input tokens, which works out to roughly $0.14 per million input and $0.56 per million output at the July 2026 exchange rate. That undercuts every Western frontier API by an order of magnitude, which is consistent with the broader pattern of Chinese lab pricing running 5 to 30 times lower than GPT and Claude.

GMI Cloud published the first independent throughput numbers. Artificial Analysis measured 171.4 tokens per second output throughput for Hy3 preview on GMI Cloud versus 175.5 tokens per second on SiliconFlow, which is the first independent third-party data point for the model's serving characteristics.

Open Weight LLM Agentic Coding 2026: Where Hy3 Fits

Usage patterns tell a clearer story than headline benchmarks. GMI Cloud reports that within two weeks of the preview launch, Hy3 processed 3.66 trillion tokens on OpenRouter, a 298% week-over-week increase, with highest usage coming from Hermes Agent, Claude Code, Kilo Code, OpenClaw, and Cline. Four of those five are coding agents. Developers are routing coding traffic to an open model that Tencent itself concedes is not the best open model for repository-scale coding.

The reason is price. In a GMI Cloud frontend code generation evaluation, Hy3 operated at roughly half the cost of the GLM 5.x series, matching DeepSeek V4's pricing efficiency. For iterative agent loops that spin thousands of tokens per task, halving the per-token cost matters more than a 12-point SWE-bench Pro spread. BiggoNews reports that at Tencent's own general availability launch, daily token consumption surged 20-fold versus the preview version, with coding and agent scenarios growing over 16.5 times.

Inside Tencent's own product surface, the deployment numbers are specific. WeChat Official Account AI avatar intent accuracy rose to 98.94%, QQ Browser programming task success rate increased 37.6%, and ima knowledge base Agent task system stability reached 95.1%. Whether those numbers replicate outside Tencent's fine-tuned in-house pipelines is the open question.

How Hy3 Got Built This Fast

The timeline is the part most launch coverage omits. i-scoop reports that in February 2026 Tencent tore down its pre-training and RL infrastructure and rebuilt both from scratch, training on the new framework started six weeks later, and the Hy3 preview went live ten weeks after that. Roughly four months from infrastructure teardown to shipping a competitive frontier-adjacent model. BiggoNews credits the release to Chief AI Scientist Yao Shunyu, six months into his tenure.

A rebuild that fast is only possible on hardware and tooling that were already there. The H20-3e reference config is the visible tip of what looks like a very deliberate stack decision made months earlier. For a longer view of how these releases fit together, the Open-Weight AI archive tracks the pattern across Qwen, DeepSeek, GLM, and now Hy3.

What to Watch Before July 21

Three things will decide whether Hy3 sticks after the free window closes. First, whether OpenRouter's post-window pricing lands closer to Tencent Cloud's yuan-denominated rate or drifts up toward GLM-5.2 economics. Second, whether independent labs reproduce the 5.4% hallucination figure on out-of-distribution prompts, because Tencent's own 270-expert blind eval is not the same as a public rater study. Third, whether Anthropic or OpenAI move on pricing before enterprise procurement teams cite Hy3 in their next contract renewal.

The date to circle is July 21, 2026. That is when the free OpenRouter endpoint closes and the market gets to price a 295B Apache-2.0 model without a subsidy underneath it.