Hunyuan Hy3: When to Route Work to Open Weights

Hunyuan Hy3 is the first Chinese open-weight reasoning model cheap enough to self-host casually — Apache 2.0, a sub-300GB FP8 footprint, and a 256K context window. That combination turns a question most agencies never seriously asked into a live one: for a given piece of work, do you route it to Hy3 or keep it on a frontier closed model?

The answer is almost never a blanket yes or no. It is a per-task decision that turns on three variables at once — the type of work, the token economics, and how sensitive the data is. Hy3 is genuinely competitive on agentic search, tool orchestration and long-context retrieval; it trails a frontier model on repo-scale coding by a real margin. It costs one to two orders of magnitude less per token. And because it is a Tencent model, the API-versus-self-host choice is a data-jurisdiction decision, not just a compute-budget one.

This guide is the companion to our launch coverage, not a repeat of it. It assumes you already know Hy3 shipped; its job is to give you a decision-ready routing framework — a task matrix, a cost read that stays honest about a moving price, and a self-host fork grounded in the actual legal exposure — that an agency ops lead can screenshot and use.

The April 2026 Hy3 preview shipped under Tencent’s restrictive “Hy Community License Agreement,” which explicitly excluded the EU, the UK and South Korea from its usage terms. For a European agency that made the model a non-starter — you could read about it, not build on it. The July 6 full release reversed that: Hy3 now ships under Apache 2.0, which carries no field-of-use or geographic restriction and no royalty obligation, and permits commercial deployment, fine-tuning and redistribution of the weights without Tencent’s involvement.

That licence change is the whole reason this is a decision guide rather than a news recap. When a capable reasoning model becomes freely deployable and an order of magnitude cheaper than the frontier per token, “which model do we use” stops being a default and becomes a per-workload call. Hy3 is a 295B-parameter Mixture-of-Experts model with 21B active parameters per token and a native 256K context window — capable enough to matter, small enough to run. The question is no longer whether to consider it, but where.

Before going further, one disambiguation: Hy3 is a text reasoning model. It is not an image or video product, and it is a separate release from the more restrictively licensed Hy3 preview of April. Everything below concerns the July open-weight reasoning model and how an agency should slot it into an existing agent stack.

Most Hy3 coverage presents its benchmarks and its self-hosting economics as two separate stories. In practice they are inputs to a single decision, alongside a third variable the benchmark coverage rarely connects to a routing choice: data sensitivity. Weigh all three together and the right destination for a given task usually falls out cleanly.

The chart below sets Hy3’s agentic and long-context scores next to its coding results, drawn from Tencent’s own benchmark appendix as reported by VentureBeat. Orange marks the tasks where Hy3 leads the open-weight field; blue marks the coding suite, where GLM-5.2 — a larger model using roughly double Hy3’s active-parameter compute per token — is ahead. One caveat applies to every bar: these are vendor-reported numbers. As of early July 2026, no independent evaluator had published a third-party verification, so treat them as directional and benchmark on your own workloads before switching a default.

The pattern is consistent, and it maps directly onto agency work. Client and market research that walks the web, multi-step tool loops that call MCP servers, and question-answering over long briefs or document sets are exactly where Hy3 is strongest — and those are also some of the most token-hungry, repeatable tasks an agency runs. The coding gap is equally clear-eyed: on repo-scale software work Hy3 is a capable assistant but not a frontier one, and the DeepSWE result in particular shows the distance is not marginal. If your agent writes production code across a real repository, that work belongs on a frontier model. For the precise coding-vs-hardware trade, we walk through matching open-weight coding models to hardware separately.

Hy3’s model card documents a reasoning_effort parameter with three settings, selectable per request via extra_body. This is more than a quality dial — inside an agent loop it is a genuine cost lever, and using it well is one of the biggest practical savings Hy3 offers over a flat frontier call.

The pattern most sources miss is what this enables in an agent loop. A single frontier call reasons at one fixed effort for the whole task. With Hy3 you can decompose: keep the expensive high setting for the one or two planning steps that actually need it, and drop every trivial sub-step to no_think. On a loop with a dozen tool calls, the majority of which are parsing and formatting, that is a large output- token saving on top of the already low per-token price — and it costs nothing to implement beyond setting a field per request.

The headline is straightforward: Hy3 is one to two orders of magnitude cheaper per token than any frontier model. The important discipline is not to pretend the exact number is fixed. OpenRouter routes tencent/hy3 across multiple hosting providers at different price points, and a promotional free tier is phasing in and out, so any single figure is a time-variant snapshot rather than a standing rate.

The strategic read matters more than the arithmetic. A cost gap this wide changes which work is economically worth automating at all. Pipelines you would never run against a frontier model because the token bill would swamp the value — exhaustive web research across hundreds of sources, large-scale document triage, high-frequency agent loops — become viable when the per-token cost falls by one to two orders of magnitude. That is the deeper reason to care about Hy3: not that it makes existing workloads cheaper, but that it moves the line on what is worth building. It also strengthens the general case for keeping a second-source strategy against vendor lock-in.

Here is the framework as a single reference. For each common category of agency work, it gives a route, the reason, how much weight cost should carry in that specific call, and the data-sensitivity flag to watch. It is the artifact to screenshot — the rest of this guide is the reasoning behind each cell.

The single most common mistake is to read only the first two columns. An agentic-research task on public data and a regulated-data task can both be capability-and-cost wins for Hy3, yet route to completely different infrastructure — the public one to any convenient API, the regulated one to self-hosted weights only. The right-hand column is what keeps the matrix from producing a compliant-looking decision that is quietly a data-governance problem. This is the same model-selection discipline we bring to our AI transformation engagements.

Most coverage frames self-hosting purely as a hardware-and-cost trade. For any agency touching sensitive data it is first a data-residency question, and it turns on a distinction the phrase “using Hy3” blurs: routing prompts through Tencent Cloud directly, routing them through an OpenRouter third-party host, and self-hosting the weights are three different data-exposure profiles. Only the last removes the China-jurisdiction question rather than relocating it.

The self-hosting path is now genuinely reachable, which is what makes the fork a live choice rather than a theoretical one. The FP8 weights sit under 300GB, and one independent hardware analysis (ComputeLeap) sizes a working deployment at roughly 8× NVIDIA H200 for full BF16 precision at 256K context, or 4× H200 for the FP8 build at a reduced 32–64K context, with a KV cache of roughly 80–120GB per concurrent long-context sequence on top. The same analysis estimates spot cost at around $14.56 per hour for the 8-GPU configuration and $7.28 for the 4-GPU one — an independent estimate, not a Tencent figure, and spot pricing is inherently volatile, so treat it as an order-of-magnitude guide rather than a quote. Consumer hardware is ruled out: even a 512GB Mac Studio would have almost nothing left for KV cache once the weights load.

Crucially, GLM-5.2 — the model that beats Hy3 on coding — needs roughly 744GB in FP8 and an 8× H200 node as a floor, so it is far less practical for an agency-sized team to self-host. Hy3’s sub-300GB footprint is the whole reason it, specifically, makes self-hosting a realistic option. We ran the full version of this arithmetic in the same hardware math we ran for GLM-5.2, and the routing conclusions here feed our general self-hosting decision guide.

The reason data sensitivity is its own routing axis is that a China-domiciled API carries distinct, dated, citable legal exposure that a frontier US or EU endpoint does not. None of this makes Hy3 unusable — self-hosting the open weights sidesteps all of it — but it is exactly why the self-host fork matters for regulated work. The key facts, each from its own established source:

• National Intelligence Law (2017, amended 2018). Article 7 requires all organisations and citizens to support and cooperate with national intelligence work. Legal scholarship debates how the proactive-data-sharing obligation is enforced in practice, but the statutory obligation itself is not in question.
• Cybersecurity Law, AI amendment. An amendment adopted 28 October 2025 took effect 1 January 2026 — six months before Hy3 launched — adding an AI-specific article on state support and strengthening security-oversight language for AI systems.
• DSL and PIPL cross-border controls. China’s Data Security Law and Personal Information Protection Law impose separate transfer controls, with a regulatory security assessment required above defined thresholds; PIPL Article 66 sets penalties up to RMB 50 million or 5% of prior-year turnover for serious violations.
• GDPR Chapter V. China holds no EU adequacy decision, so any EU personal data sent to a China-hosted endpoint needs Standard Contractual Clauses or another approved transfer mechanism first.
• Section 1260H designation. On 7 January 2025 the US Department of Defense added Tencent to its list of “Chinese military companies” under Section 1260H of the FY2021 NDAA. Tencent called the listing a mistake and says it is not a military company or supplier. The designation imposes no sanction or export control, but it is a recognised due-diligence flag for government-adjacent and defence-sector procurement.

None of this is a novel angle invented for this post. We applied the identical framing to Zhipu AI when we wrote up GLM-5.2’s benchmark numbers, and it sits inside the wider question of the broader open-versus-closed tradeoff. Treating China-domiciled model APIs as a live data-residency consideration — not a reflexive ban, and not a non-issue — is the consistent house position.