Models · Tencent Hunyuan · Out sinceReleased 6 Jul 2026

Hy3

Available on OpenRouterOn OpenRouterAvailable nowGenerally availableCan run on your own serversOpen weights

Hy3 is made by Tencent Hunyuan. We don't have enough test results yet to rank it. It's cheap to use. Its makers have published it, so you can run it on your own servers.

295B total / 21B active MoE (+3.8B MTP layer), 256K context, plain Apache 2.0 (no territory carve-out in LICENSE). GA after the April Hy3 preview. Official finetuning pipeline and verl RL recipe in repo. OpenRouter ~$0.13/$0.53.

Writing & creativity: —
Research & analysis: —
Coding: —
Price: Cheap$0.13 / $0.53
Price per 1M (blended)Blended / 1M: $0.23
MemoryContext: 262K
Longest answerMax output: —
Test resultsResults: 8 (8 independent8 indep.)
Out sinceReleased: 6 Jul 2026
Made byVendor: Tencent Hunyuan
UnderstandsInputs: text

Scores are out of 100 for each area, compared with every model we track; “#” is its rank. Context is how much text it can read at once (1M is roughly 700,000 words). Blended price mixes the cost of what you send and what it writes back.

Thinking levels

Reasoning settings

You can't choose how long this one thinks.

No adjustable reasoning setting listed.

Links

Using it through OpenRouter

On OpenRouter

OpenRouter is a service that gives access to many AI models in one place. This model is listed there as tencent/hy3.

Route it as tencent/hy3 at $0.13 in / $0.53 out per 1M tokens, 262K context. Listed since 6 Jul 2026.

frequency_penaltyinclude_reasoninglogit_biasmax_completion_tokensmax_tokensmin_ppresence_penaltyreasoningreasoning_effortrepetition_penaltyresponse_formatseedstopstructured_outputstemperaturetool_choicetoolstop_ktop_p

Run it yourself

Self-hosting

On your own servers,Licence, sizeunder your own control.and hardware.

Compare all open models →All open-weights models →

LicenceLicence: Apache-2.0
OK for business useCommercial use: Yes
SizeParams (total / active): 295B / 21B
Can be trained furtherBase model: No

Hardware you'd need

Hardware tier & memory

Needs one full AI server.

≤ 1.1 TB at 8-bit (one 8×H200 node). Weights ≈ 310 GB at 8-bit, 162 GB at 4-bit (+10–30% for KV cache). MoE: 21B active per token.

Languages: Languages: Not stated.

Training it further: Fine-tuning: Official finetuning pipeline (finetune/README.md) and GRPO RL recipe with verl + Megatron-LM (rl/README.md).

Quantisations: bf16, fp8 (tencent/Hy3-FP8), gguf (community: AngelSlim/Hy3-GGUF). Engines: vLLM, SGLang.

Download (Hugging Face)Weights on Hugging Face ↗huggingface.co ↗huggingface.co ↗

Every test result

Every result

All the numbers,with where they came from.

Every result we've found for this model, with who measured it and a link to where we read it.

All raw rows (vendor and independent kept separate), with setting label, source, date, harness and notes.

All results for this model
						NotesNotes
AA-LCR	79.0%	DefaultHy3	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug hy3
AA-Omniscience Index	-18.5	DefaultHy3	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug hy3; AA-Omniscience Index (-100..100)
Artificial Analysis Intelligence Index	25.3	DefaultHy3	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug hy3
Artificial Analysis output speed	87 tok/s	DefaultHy3	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug hy3; median output tokens/sec across AA-tracked API providers (not a self-hosted measurement)
GPQA Diamond	89.7%	DefaultHy3	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug hy3
Humanity's Last Exam	33.5%	DefaultHy3	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug hy3
SciCode	48.6%	DefaultHy3	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug hy3
Terminal-Bench 2.1	64.4%	DefaultHy3	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug hy3