Hunyuan Hy3: 295B Open Weights Under Apache 2.0

Tencent released the full Hunyuan Hy3 open-weight reasoning model on July 6, 2026 — a 295-billion-parameter Mixture-of-Experts that activates just 21 billion parameters per forward pass, shipped under Apache 2.0 with no field-of-use or geographic restrictions. The headline for most coverage was a benchmark race. The headline that actually matters for teams evaluating self-hosting is that Hy3 runs in under half the memory of the rival it is measured against.

The comparison target is GLM-5.2, the roughly 744B-parameter open model from Z.ai that has owned the open-weight coding benchmarks since mid-June. On Tencent’s own appendix, GLM-5.2 still wins the agentic coding suite. What Hy3 offers instead is a specific trade: leadership on agentic search, tool orchestration and long-context retrieval among open models, at a sub-300GB FP8 footprint that fits a single serving node rather than an eight-GPU cluster.

This guide covers what shipped, the license reversal that makes Hy3 usable at all in Europe, the architecture and deployment math, an honest benchmark scorecard against GLM-5.2, and where the numbers deserve skepticism. Every performance figure below is Tencent-reported unless stated otherwise — independent third-party verification had not been published as of writing, and we label the vendor claims plainly throughout.

Tencent’s Hunyuan team published the complete Hy3 model on July 6, 2026, framing it in the company newsroom as the culmination of a “full model-development loop in less than six months” that ran from a February 2026 architecture rebuild through the July release. The weights went up on Hugging Face as tencent/Hy3, with distribution promised across third-party developer platforms including OpenRouter, Cline, OpenCode and Cherry Studio.

Hy3 is a pure reasoning and agentic language model. It is worth stating plainly because the name invites confusion: there is no image or video product called Hy3, and the earlier April preview was also a text LLM, not a media model. Everything discussed here is a single 295B MoE optimized for coding, tool-calling, reasoning and 256K long-context work.

For a European agency’s audience, the license is the detail that changes whether Hy3 is on the table at all. The April 23, 2026 Hy3 preview shipped under a restrictive “Tencent Hy Community License Agreement” that explicitly excluded the EU, UK and South Korea from its usage terms. The July full release drops that entirely: Apache 2.0, no field-of-use clause, no geographic carve-out.

That reversal reframes the whole release. When developers reacted on launch day, the license change — not the benchmark scores — was widely treated as the real headline; some argued Tencent had joined the top tier of open-source labs, while others held that GLM-5.2 remained the best currently usable open-weight model in practice. Both can be true. A model you can legally deploy without a regional exclusion is categorically more useful to a Bratislava or London team than one you cannot, whatever the leaderboard says.

This is the pattern worth watching across the Chinese open-weight cluster in 2026. Permissive licensing has become a competitive lever in its own right, not an afterthought — a way to buy developer adoption and platform distribution that a restrictive license forecloses. For the practical mechanics of running these models in-house, our open-weight self-hosting decision guide walks through the license-then-hardware sequence Hy3 now fits cleanly into.

Hy3 is a sparsely activated MoE. Of the 295 billion total parameters, only 21 billion fire on any given token — the model routes each token through 8 of 192 experts (top-8 routing) across an 80-layer transformer with a 4096 hidden size and grouped-query attention using 64 attention heads to 8 key-value heads. That sparsity is the whole economic argument: frontier-scale capacity, a fraction of the per-token compute.

A separate 3.8B-parameter Multi-Token Prediction layer predicts several tokens at once for faster decoding and is compatible with vLLM and SGLang speculative decoding, so the throughput gains are available in the mainstream serving stacks rather than a bespoke runtime. Native context is 256K tokens; note that OpenRouter lists the model at 262K, a platform-side rounding and overhead figure rather than a separate capability.

Here is the concrete own-your-stack argument. GLM-5.2’s FP8 weights consume roughly 744GB, which makes an 8x H200 node the practical minimum for production serving. Hy3’s FP8 footprint is under 300GB — less than half the memory, with roughly half the active-parameter compute per request. That is the difference between provisioning one serving node and standing up an eight-GPU cluster, and it compounds across every replica you run.

Tencent’s recommended serving configuration targets Nvidia’s H20-3e — the memory-boosted, export-compliant variant of the H20 built for the China market — with no mention of Huawei or Ascend chips in the deployment guide, unlike GLM-5.2’s. For teams outside China the specific SKU matters less than the headline: the model was engineered to fit commodity single-node memory budgets, which is exactly the constraint that governs whether self-hosting pencils out.

The framing to hold onto is that Hy3 does not win the benchmark race outright and does not need to. If self-hosting economics are your binding constraint — and for most brands evaluating on-prem or sovereign deployment, they are — a model that delivers competitive agentic performance at 40% of the memory changes the build-versus-buy math more than a few points of coding benchmark ever would. For the hardware side of that same calculation, our breakdown of the hardware reality of self-hosting a 744B model is the direct counterpoint to Hy3’s lighter footprint.

Be precise about the headline. The shorthand circulating on launch day — that Hy3 “wins everywhere except coding” — overstates it. On Tencent’s own benchmark appendix, GLM-5.2 leads essentially the entire agentic coding suite, and by wide margins in places. Hy3’s strength is not a clean sweep of everything non-coding; it is a specific lead on agentic search, tool-use and long-context workloads, which we cover in the next section. The scorecard below pairs the deployment math with the head-to-head coding deltas so both sit in one glance.

The DeepSWE gap in particular — 46.2 to 28.0 — is not close, and it is worth reading before anyone routes a coding agent to Hy3 on the strength of the launch buzz. If competitive coding performance is your primary requirement, GLM-5.2 remains the open model to beat, and our roundup of matching open-weight coding models to hardware is the better starting point than Hy3. Read the full head-to-head in GLM-5.2’s own benchmark numbers.

This is where the trade pays off. On Tencent’s tables Hy3 leads every open model on agentic-search and tool-orchestration workloads, and is described as competitive with — not necessarily beating — Claude Opus 4.8 and GPT-5.5 on some of them. The bars below are the Tencent-reported scores that define Hy3’s actual strength profile. Treat them as vendor claims pending independent confirmation, but note that they map cleanly onto the model’s design emphasis: cost-effective agentic use.

The design intent shows in the training principles too. Tencent describes Hy3’s reliability objective in terms an agentic builder should appreciate: answer only from evidence, and say so when there is none. That is exactly the failure mode — confident fabrication mid-tool-chain — that breaks long-horizon agent runs in production.

Tencent reports large reliability improvements from the April preview to the July release. On its internal evaluations the hallucination rate fell from 12.5% to 5.4%, the commonsense error rate from 25.4% to 12.7%, and the multi-turn issue rate from 17.4% to 7.9%, while the score on the open MRCR long-dialogue benchmark rose from 42.9% to 75.1%. If those hold up, they describe a materially more dependable model for multi-step agent work.

The caveat is unavoidable and Tencent’s own coverage flags it: these are internal, self-reported measurements. Between the April preview and the July release, Tencent says it gathered feedback from more than 50 product teams and scaled up its post-training pipeline with higher-quality data — a plausible mechanism for the gains, but not independent evidence of them. We repeat the numbers because they are the vendor’s stated claims, not because they have been confirmed.

The open weights are on Hugging Face as tencent/Hy3— 598GB at full precision, 300GB quantized to FP8 — for on-prem, fine-tuning or sovereign deployment. For hosted access, Tencent said Hy3 would be progressively rolled out across global developer platforms including OpenRouter, Cline, OpenCode and Cherry Studio. Nous Research made it free on its Nous Portal for two weeks from July 6 for “cost-effective agentic use,” and OpenCode added free access the same week.

Pricing needs a caveat. OpenRouter routes tencent/hy3across multiple hosting providers at different rates, and the figures shifted as a free-tier promotion phased in and out. Listed routes ranged around $0.06–$0.20 per million input tokens and $0.21–$0.80 per million output at launch coverage, with individual routes listing figures such as $0.14 in / $0.58 out per million and a 262K context window. A free tier (tencent/hy3:free) was also available on OpenRouter, scheduled to end July 21, 2026. Do not treat any single figure as definitive — re-check the live listing at the moment you provision.

Hy3 does not create a new category — it lands in an already crowded Chinese open-weight cluster alongside GLM-5.2, DeepSeek-V4 and Kimi K2. What it adds is a specific combination: Apache 2.0 licensing plus a sub-300GB FP8 footprint plus a genuine lead on agentic search and tool use. The decision tree below maps that profile to real workload classes rather than a single verdict.

Looking forward, the signal in this release is not Hy3’s benchmark line — it is the direction of travel. Two of the strongest open-weight labs are now competing on deployment economics and permissive licensing as hard as on raw capability, which is precisely the axis that determines whether a brand can own its AI stack rather than rent it. Expect the next wave of releases to keep pushing footprint down and license terms open, because that is what wins adoption when the top of the leaderboard is a crowded field. Deciding which of these models to actually run is the kind of comparative evaluation our AI transformation engagements start with.