Hunyuan 3
TencentReleased Jul 6, 2026Open weights · Apache 2.0hy3
- Reasoning
- yes, with effort levels
- Context
- 262K tokens
- Input
- text
- API features
- tools, structured outputs, seed
Read Nov 3, 14:34 UTC from the hosts' published offers and OpenRouter, spec sheet from Tencent
The cheapest host, DeepInfra, serves fp4 weights (maker: bf16). Weights and context per host
| Provider | Input | Output | Cache ↑ | Checked |
|---|---|---|---|---|
| DeepInfrafp4 (bf16) · 256K ctx | $0.13 | $0.53 | $0.033 | Oct 8 |
| Novita AIprecision not stated · 256K ctx | $0.14 | $0.58 | $0.035 | Nov 3 |
2 of 2 providers · prices per million tokens, read Nov 3, 14:34 UTC · under each host: weights precision, context and max output as served, the maker's value in brackets when the host's differs
| Provider | Weights | Context | Max output | Tools | Structured outputs | Reasoning |
|---|---|---|---|---|---|---|
| DeepInfra | fp4 | 256K | not stated by the host | — | — | — |
| Novita AI | not stated by the host | 256K | not stated by the host | yes | yes | yes |
Weights, context, max output: from the host where its terms allow, else as declared to OpenRouter (hover for source and date) · features as declared to OpenRouter, not verified by us
Hunyuan 3 spec sheet, field by field, with sources
Hunyuan 3 spec sheetmaker-declared
What Tencent declares, from its own weights and docs. What each host serves is in the offers: both can be true.
- Context window
- 262,144 tokens
- Parameters
- 298,786,155,776≈ 298.8B
- Active experts
- 8 of 192 active per token
- Weight precision
- BF16
- License
- Apache 2.0 — open weights, permissivecurated · read 2026-09-16Scope: the license of the published weights only — not the maker's acceptable-use policy, nor the terms of the host you call.
- Knowledge cutoff
- to verify
- Modalities
- text
- Max output (official)
- to verify
- Release date
- to verify≈ July 6, 2026 third-party catalogue — not the maker's figure
- Maker lifecycle status
- to verify
Also declared in the weights
- Architecture class
- HYV3ForCausalLM
- Layers
- 80
- Vocabulary
- 120,832 tokens
- Weights first published
- July 2, 2026
- Weights last updated
- July 20, 2026
tencent/Hy3 · config.json @ a960ebc · read 2026-10-08 · 6 of 10 fields published by the maker
| Date | Provider | Input | Output | Change |
|---|---|---|---|---|
| Sep 22 | DeepInfra | $0.14 → $0.13 | $0.58 → $0.53 | −8% |
| Sep 18 | DeepInfra | $0.10 → $0.14 | $0.43 → $0.58 | +33% |
Every published change of input or output price, newest first · change on input + output · full series: pricing-history.csv.
| Provider | Chatbot 1,000 in / 500 out | RAG / long context 8,000 in / 500 out | Batch summarize 4,000 in / 1,000 out |
|---|---|---|---|
| DeepInfra | $0.40 | $1.30 | $1.05 |
| Novita AI | $0.43 | $1.41 | $1.14 |
What each host bills for 1,000 requests of a representative job
Your own workload
Run it yourself
open weights, Apache 2.0 · fits 2 of the 14 workstations in our guideSelf-host guide ›- Parameters
- 295B
- 21B active per token (MoE)
- Memory needed
- 177 GB
- in INT4 · 354 GB in INT8
- Smallest cloud setup
- 4× H100
- 320 GB · $10.56/h, about $7,709 a month
- Worth it above
- 12B tokens
- a month, against the cheapest API at $0.66 per M
| Quantization | Memory | Smallest cloud GPU | Quality |
|---|---|---|---|
| INT4 (Q4) | 177 GB | 4× H100 · $10.56/h | smallest footprint, minor quality loss |
| INT8 (Q8) | 354 GB | 8× H100 · $21.12/h | near-lossless |
| FP16 (full) | 708 GB | multi-node, over 640 GB | reference quality |
At home: in INT4 it fits the Mac Studio M5 Ultra (256 GB unified) — about $12,299 once, or $0.468/h over 3 years, electricity aside, at 1200 GB/s of memory bandwidth — that, not FLOPS, sets tokens per second; it pays for itself after about 19B tokens against the cheapest API.
Memory ≈ parameters × bytes × 1.2 · cloud: on-demand, 24/7 · builds: GGUF, AWQ, MLX, GPTQ, Ollama · break-even calculator
Other Tencent models · all its models, hosts and history ›
- Hunyuan 4 Preview1 provider · from $3.33