Wygenerowana karta tytułowa z napisem „Tencent Hy4 kontra Kimi K3.1 Pro” z dwiema kolumnami statystyk: Hy4 z indeksem inteligencji 34,6, 770 mld łącznych i 49 mld aktywnych parametrów, 0,834 USD/2,501 USD za milion tokenów; Kimi K3.1 Pro z indeksem inteligencji 46,5, 2,8 bln łącznych i 104 mld aktywnych, 2,00 USD/10,00 USD; oraz stopka z napisem „Indeks według Artificial Analysis v5.0.1, odczyt z 20 listopada 2026 r.; wagi obu modeli są publicznie dostępne”.
Guides & Insights

Hy4 vs Kimi K3.1 Pro: oba do pobrania, jeden do samodzielnego hostowania

Autor

Alistair Wren

Data publikacji

Najnowsze modele · 20Zobacz wszystkie modele →
Benchmarki: Artificial Analysis · aktualizowane codziennie
Powrót do wszystkich wpisów

Kimi K3.1 Pro scores 46.5 on the Artificial Analysis Intelligence Index. T​encent Hy4 scores 34.6. That is a 12-point gap, and it is the largest deficit Hy4 runs in any matchup this week — larger than the one against Qwen4-Max, more than ten times the one against Grok 4.7. On the board, Kimi K3.1 Pro is the best open-weight model listed, first of 32 in its class.

Then you look at what the two weigh. Kimi K3.1 Pro is a 2.78-trillion-parameter model with 104 billion active per token. T​encent Hy4 is 774 billion total with 49 billion active. Both were published as downloadable weights — K​imi's on 2 November 2026, Hy4's on 18 November, three weeks apart. Only one of them fits on a machine a normal company owns, and that single fact reorganises the entire comparison. This is not a story about which model is smarter. It is a story about which model you can actually run, and what the smarter one costs you per completed task if you rent it instead.

Jedyny benchmark, który o tym decyduje

Minimalny możliwy do wdrożenia rozmiar, przy zalecanej przez samego dostawcę precyzji:

• T​encent Hy4 — 774 billion parameters in BF16, published as 245 shards of roughly 12.9 GB and 6.4 GB, about 1.45 TB before any quantisation. T​encent's own serving recipe uses tensor parallelism of 8 on a single node's worth of GPUs with the FP8 build, which is roughly 786 GB of weights.

• Kimi K3.1 Pro — 2,78 biliona parametrów, 2,72 TB skwantyzowanych wag U8 w repozytorium w postaci dostarczonej, a tensory BF16 zajmują kolejne 57 GB. Nie istnieje konfiguracja jednomaszynowa, w której ten model byłby projektem na weekend.

Both vendors ship the software to make serving tractable, and both are unusually good about it. T​encent publishes a vLLM image tag (vllm/vllm-openai:hy4) and an SGLang image (lmsysorg/sglang:hy4) with the tensor-parallel and speculative-decoding flags written out; K​imi's repository ships a compressed-tensors MXFP4 configuration and a 4-bit path that brings the weight file down to the 2.72 TB figure above. Neither of those is a small-model story. The difference is that an 8-GPU node is a purchase most engineering organisations can authorise, and a 2.8-trillion-parameter deployment is a data-centre decision.

Jeśli samodzielne hostowanie było powodem, dla którego interesowały Cię otwarte wagi, to zdanie jest całym argumentem i nie zależy od benchmarku.

Generated two-column scoreboard comparing Tencent Hy4 and Kimi K3.1 Pro across six dimensions: Intelligence Index 34.6 versus 46.5, total parameters 770B versus 2.8T, active parameters 49B versus 104B, price $0.834/$2.501 versus $2.00/$10.00 per million tokens, Terminal-Bench 4.0 37.4% versus 67.2%, and cost per index task $1.03 versus $3.68, with a footer reading 'Index figures per Artificial Analysis v5.0.1; parameter counts from vendor model cards.'

Gdzie znajduje się dwanaście punktów

12-punktowa różnica w indeksie brzmi jednolicie. Nie jest — Hy4 trzyma poziom w kilku ocenach, a w jednej zostaje zmiażdżony.

• Agentic coding, Terminal-Bench 4.0 — Kimi K3.1 Pro 67.2%, T​encent Hy4 37.4%. A 30-point gap, and the single biggest contributor to the index difference.

• Agentic science in a terminal, Terminal-Bench-Science — K​imi 34.3%, Hy4 12.9%.

• Software engineering in real repositories, FrontierCode-AA 2.0 — Hy4 31.4%, K​imi 25.5%. Hy4 is the better model here, by six points, and it is the one evaluation where a terminal agent's real-repository work favours T​encent.

• Hard reasoning, HLE-Diamond — K​imi 37.3%, Hy4 32.7%. Closer than the headline suggests.

• Physics, CritPt v2 hard subset — K​imi 49.5%, Hy4 39.0%.

• SaaS workflow execution, AutomationBench-AA — K​imi 73.3%, Hy4 65.6%.

• Agentic knowledge work, AA-Briefcase Elo — K​imi 1,673.6, Hy4 1,445.9. A 228-Elo gap with non-overlapping intervals.

The FrontierCode result is the one worth pausing on, because it contradicts the obvious reading of the other six. On a benchmark of changing real open-source repositories, Hy4 leads. On a benchmark of a clean terminal environment, K​imi leads by 30 points. T​encent's model card explains why that could be: Hy4 was trained specifically against T​encent's own game and frontend pipelines — Unreal, Unity, Godot, UI state management — and reports a 19.1% gain on its internal GameDev suite over Hy4 preview. That is a model tuned for building things, evaluated on a test of repairing them. Take both rows and the honest conclusion is that the coding story is workload-dependent, not that one model codes better.

Screenshot of the Artificial Analysis page for Tencent Hy4, captured 20 November 2026, showing its Intelligence, Performance, Cost and Verbosity rank tiles, 93 tokens per second output, 14.0 minutes per task, a 118.9-second time to first token and a 1M-token context window, with the header still labelling the model proprietary.

Cena 3,6 razy niższa za wykonane zadanie

To tutaj porównanie staje się arytmetyką, a nie argumentem. Artificial Analysis podaje zmierzony koszt na zadanie Intelligence Index dla obu modeli, który uwzględnia ceny tokenów, długość rozumowania i zachowanie pamięci podręcznej:

• T​encent Hy4 — $1.03 per task, at $0.834 input and $2.501 output per million tokens, 95% cache discount at $0.042.

• Kimi K3.1 Pro — 3,68 USD za zadanie, przy 2,00 USD za dane wejściowe i 10,00 USD za dane wyjściowe za milion tokenów, 90% zniżki na pamięć podręczną przy 0,20 USD.

So the 12-point index gap costs 3.6 times as much per task. Per index point, K​imi is roughly 4.4x the price. That is not a scandal — a model that is genuinely better at hard agentic work is supposed to cost more — but it reframes the decision. If your workload needs 67% on Terminal-Bench, there is no substitute and you should pay. If your workload needs 37%, paying three and a half times for the other 30 points is money spent on capability you will not invoke.

There is a second-order cost that the per-task figure hides. K​imi takes 39.7 minutes per index task against Hy4's 14.0, and returns its first answer token after 162.8 seconds against 115.9. At 30 output tokens per second against Hy4's 93, the slowest model in this comparison by a factor of three is also the most expensive. In an agent loop that runs hundreds of steps, latency and cost compound together rather than trading off.

Dwie licencje z dwiema różnymi pułapkami

Oba modele są „open weight” w tym sensie, że można je pobrać i używać. Żaden nie jest objęty licencją Apache 2.0, a różnica między ich klauzulami komercyjnymi nie jest kwestią akademicką:

• T​encent Hy4 — the T​encent Hy License 1.0, derived from Apache 2.0. Section 10 requires a separate written agreement before offering the model as a service if your worldwide gross revenue exceeds $50 million in any 12-month period, or if the product carrying it exceeds 100 million monthly active users. Internal use is carved out.

• Kimi K3.1 Pro — licencja Kimi K3.1-Pro. Ten sam rodzaj ograniczenia, ale przy progu przychodów wynoszącym 20 mln USD, a nie 50 mln USD.

Both carve out end-user products where the model is embedded in specific features rather than exposed as general-purpose model access. Both include mere relaying to a third-party model service as permitted. In plain terms: a small product team building a feature on either model is fine. A company running an inference business on either needs a phone call, and K​imi's phone call comes at less than half the revenue.

Dla pełności obrazu poprzednia generacja jest tu wyjątkiem: Hy4 preview, wydany 28 sierpnia 2026 r., był objęty licencją Apache 2.0 bez jakiegokolwiek progu komercyjnego. Każdy, kto budował na licencji wersji preview i teraz przechodzi na Hy4, przechodzi na inne warunki.

Uruchamianie obu bez dokonania wyboru

Powód, dla którego tego porównania nie trzeba rozstrzygać na papierze, jest taki, że oba modele są dostępne z jednego punktu końcowego z przekazaną stawką samego dostawcy: tencent/hy4 po 0,834 i 2,501 USD za milion tokenów, kimi/kimi-k3.1-pro po 2,00 i 10,00 USD, z odczytami z pamięci podręcznej po 0,20 USD. To są właśnie te liczby, bez narzutów, a zmiana ceny dostawcy trafia tutaj tego samego dnia, a nie dopiero przy odnowieniu umowy.

For a pair this far apart on capability and cost, the useful pattern is not a split — it is a ceiling. Route to kimi/kimi-k3.1-pro through the routing DSL, set tencent/hy4 as the automatic failover target, and watch how often the failover fires. If it fires rarely, you are paying for K​imi's capability and getting it. If it fires constantly because of K​imi's latency, you have just discovered something a benchmark table cannot tell you, at Hy4's prices rather than K​imi's. Model fusion is the other option: a panel of the two would let the stronger model's reasoning carry the weaker model's cheaper token budget on the evaluations where they converge.

Screenshot of the OrcaRouter model page for Kimi K3.1 Pro, captured 20 November 2026, showing the kimi/kimi-k3.1-pro endpoint with a 1,048,576-token context window, 943,718-token maximum output, text, image and video input, reasoning effort from none to max, a 143-second median time to first token and $2.00/$10.00 per million tokens.

Co zmieniłoby tę odpowiedź?

Three things, in order of likelihood. First, a Hy4 fine-tune: the weights have been public for two days, and the 30-point Terminal-Bench gap is exactly the kind of margin a targeted fine-tune on a company's own terminal traces erodes. Second, an independent evaluation of Kimi K3.1 Pro at a higher reasoning effort than the default max the index uses — the 943,718-token output ceiling in the model's record has not been exercised by anyone publishing results. Third, a Hy4 Flash-class sibling that inherits the agentic training at a quarter of the price, which would make the cost argument against K​imi even harder to answer.

Until one of those lands, the split is clean. Kimi K3.1 Pro is the better model and the more expensive one, by a factor that holds up whether you measure per token, per task or per index point. T​encent Hy4 is the one you can put on your own hardware, at a licence threshold twice as generous and a price a third as steep. Neither of those is a reason to regret the other.

Porównane w tym artykule2

Wykryto na podstawie tego artykułu · Benchmarki: Artificial Analysis · aktualizowane codziennie