Union Alpha vs Tencent Hy3 (Free): Same Shape, Only One of Them Has a Name

Put Union Alpha and Tencent Hy3 (Free) next to each other on a spec sheet and they look like the same offer: a free, roughly 256K-context, mixture-of-experts model pitched at agentic and coding work, rate-limited, at $0. The difference is not in the specs. It is that one of them ships with a named lab, a published licence, a documented architecture, a benchmark table, and a stated price for when free ends — and the other one is a stranger with a good test score and a one-week window.

What is genuinely identical

Both models run on OrcaRouter's free tier at $0, both return HTTP 429 once you cross your plan's cap, and neither charge is ever applied to your balance. Both are marketed for coding and agentic workflows. Both expose tool calling. Both carry a context window of 262,144 tokens.

That is a lot of overlap for two models with so little else in common, and it is why the comparison is worth making properly rather than on price.

Side by side

• Context window — 262,144 tokens on both. A genuine tie, and the reason these two get compared at all.

• Max output — Union Alpha 131,072 tokens vs Tencent Hy3 (Free) listing up to 262,144 tokens.

• Inputs — text and image on Union Alpha vs text only on Tencent Hy3 (Free), which does not accept images.

• Reasoning controls — none on Union Alpha vs three modes on Tencent Hy3 (Free): no_think, low, and high, selected via reasoning_effort.

• Architecture — undisclosed on Union Alpha vs a documented 295B-parameter dense-MoE hybrid activating roughly 21B per token, with 192 routed experts plus one shared expert, top-8 sigmoid routing, and a 3.8B multi-token-prediction layer for speculative decoding.

• Weights — none published on Union Alpha vs open-source releases of Hy3.

• Pricing after free — unannounced on Union Alpha vs an official paid rate around ¥1 per million input tokens and ¥4 per million output tokens, with cached input at ¥0.25.

• Time to first token — Union Alpha 10.00 s p50 vs Tencent Hy3 (Free) 3.01 s p50, both over the same seven-day window on our endpoint. Output speed is closer than you would expect: 170 tokens per second against 155.

What Tencent Hy3 (Free) gives you that Union Alpha cannot

The list is long, and it is mostly about being able to plan.

Hy3 has a published benchmark table. SWE-bench Verified lands at 74.4, up from 53.0 for Hy2 — a roughly 40% relative jump. SWE-bench Pro is 57.9, SWE-bench Multilingual 75.8, BrowseComp 84.2, MCP Atlas 79.1, GPQA Diamond 87.2 on Tencent's own reporting (our model page lists 89.7), MMLU 87.42, and Humanity's Last Exam 33.5 with 53.2 when tools are allowed. Those are overwhelmingly vendor-reported and should be labelled as such, but they exist, they are dated, and they are attached to a company with a reputation to lose.

Our own page adds an independently measured layer: AA Coding 58.8, AA Intelligence 25.8, long-context recall 79.0, SciCode 48.6, ClawEval 68.5, NL2repo 45.6. There is also a blind evaluation in which 270 domain experts scored Hy3 at 2.67 out of 4 against GLM5.1's 2.51, with the strongest categories in front-end, data and storage, and CI/CD.

It has a licence, with a caveat worth knowing. Reporting is inconsistent about which one applies: some sources describe the Tencent Hy Community License, which permits commercial use with usage-policy and attribution clauses, while others report Apache 2.0 for the full Hy3 release. The preview's community licence originally excluded the EU, UK, and South Korea. If you intend to self-host, read the actual licence file rather than the blog post — and note that self-hosting needs the correct tool-call parser configured (hy_v3, or hunyuan for SGLang) or your structured calls will misbehave in ways that look like model failures.

And it starts fast. Hy3's p50 time-to-first-token is 3.01 seconds, it streams at 155 output tokens per second, and its error rate over the same seven-day window is 0.08% — roughly one failure in twelve hundred requests. Union Alpha, measured the same way, needs ten seconds or more to emit a first token on at least half of all requests and fails 7.7% of them. Raw decode speed is close between the two; readiness is not, and readiness is what decides whether a model can sit in front of a human.

What Union Alpha gives you that Hy3 cannot

Exactly one structural advantage, and it is not nothing: vision.

Union Alpha accepts text and images. Hy3 does not accept images at all. For screenshot-driven agent work, UI automation, chart reading, or document layout tasks, that is a hard capability gate rather than a preference — and it is the single strongest reason to keep Union Alpha in the rotation while it lasts.

Its second advantage is the output ceiling: 131,072 tokens is a large single response. Its third is price, in the narrow sense that free is free.

Everything else is a blank. Union Alpha appeared on third-party catalogues on 16 September 2026 with an operator listed only as "Stealth". No lab, no weights, no licence, no parameter count, no architecture note, no knowledge cutoff. No reasoning controls, and tool choice effectively only supports "auto". Its performance claim — frontier-level across general-purpose tasks — is operator-reported and unaudited.

The one independent measurement is SMF Clearinghouse's Official A run: 157 tests, reasoning off, 136 correct (86.6%), zero errors, tenth of twenty-six. Reasoning 30/30, tools 2/2, coding 26/30, math 24/30, writing 2/5.

The fine print of free

Free endpoints are where the weakest data terms live, and these two are not equivalent there either.

On Union Alpha, the operator states that prompts and completions may be retained by the provider but are not used for training — while one distribution channel has separately promised zero retention. Those are two different promises, and the discrepancy is unresolved. Treat retention and training as separate questions and assume the more conservative answer.

On Tencent Hy3 (Free), you are dealing with a company that publishes a privacy policy and has a commercial reputation attached to it. That is not a guarantee either, but it is a named party you can hold to something.

Both share the operational caveats that come with any free tier: shared capacity, deprioritisation under load, and rate limits that flex with demand rather than being published. Neither is a sensible place for a production path without a fallback behind it.

The one thing that makes this easy

OrcaRouter passes provider list price through at 0% markup, so Hy3's official paid rate — around ¥1 per million input and ¥4 per million output — is the rate you step onto when the free tier is not enough, with no repricing layer in between. And because both models sit behind a single key, "which of these two should I use" stops being a procurement decision. There is no second contract, no separate SDK, and no code change: you move a model name.

That matters most for exactly the case this article is about. The rational play is not to pick one. It is to run Hy3 as the default because it is fast, documented and permanent, and route image-bearing requests to Union Alpha while its preview is open — then let the routing rule absorb the day the preview closes.

Verdict

Tencent Hy3 (Free) is the better model to depend on, and it is not especially close. Documented architecture, published benchmarks, open weights, three reasoning modes, a three-second time to first token against ten, a licence you can read, and a known price after free. Its one real weakness — no image input — is a genuine gap, but it is a gap in capability rather than in trustworthiness.

Union Alpha is the better model to experiment with, and the vision input makes it the only option of the two for a whole class of screenshot and document work. But an anonymous operator, one independent test, a fifth of the speed, and an unannounced end date are not a foundation. They are an opportunity with a deadline.

Take the opportunity. Keep the foundation.

The named half of the pair is documented: Tencent Hy3 (Free) model page carries the 295B/21B MoE architecture, the 262K window, the three reasoning modes and the paid rate you step onto after free.