Hy4 preview
- Input
- $0.751/M tok
- Output
- $2.251/M tok
- Blended
- $1.126/M tok · 3:1
- Speed
- ~78tok/sest.
- Context
- 1Mtokens
- Cost per task
- ~$0.95/taskest.
Open weights (Apache 2.0), 770B total / 49B active MoE. Price per https://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/ (Tencent Cloud TokenHub). Not on AA or LMArena yet; its GDPval-AA V2 and AutomationBench v1.0.6 figures use a different version than our AA benchmarks, so they are left out.
Rankings
Self-hostingEstimated
How we estimate →What it would cost to run these open weights yourself on rented GPUs, compared with the API price above.
- Hardware
- 16× H200 141GB
- FP8 weights
- Throughput
- ~25,961 tok/s
- many requests batched
- Price
- $0.33–2.71 /M tok
- busy → light use
- On your own machine
- Mac Studio M5 Ultra 512GB (int4, ~34 tok/s single-stream)
- single consumer GPU or Mac
Assumptions (5)
- FP8 weights (770B params, 49B active per token) + 40% KV-cache headroom ≈ 1078 GB
- 16× H200 141GB at $3.59–$7.91/GPU-hour on-demand (2026-09-24)
- ~25,961 output tok/s aggregate at batch 1098 (bandwidth-bound); MoE compute scales with active params
- Blended 3:1 input:output; prefill ~129,766 tok/s
- Low = 75% utilization at the low GPU price; high = 20% at the high price
Benchmark scores
~ italic, dashed = no published score yet, estimated from related benchmarks.
Coding
SWE-bench VerifiedEstimated
~#2 of 21 (estimated)~86.1%
No published score — estimated from related benchmarks
SWE-bench Pro (Public)Lab-reported
#8 of 2365.7%
- #1Claude Fable 5.293.0%
- #2Claude Opus 5.589.9%
- #7Beam 1.0 501B A23B66.1%
- #8Hy4 preview65.7%
- #9GPT-5.6 Terra63.4%
huggingface.co · 2026-08-28
DeepSWE v1.1Lab-reported
#22 of 4564.3%
- #1Gemini 4 Argon78.8%
- #2DeepSeek V4.1 Flash 101777.0%
- #22Claude Fable 5.164.3%
- #22Hy4 preview64.3%
- #24GPT-6 Luna63.7%
huggingface.co · 2026-08-28
Terminal-Bench 4.0Estimated
~#24 of 102 (estimated)~28.9%
- #1Kimi K3.1 Pro67.2%
- #2Claude Sonnet 5.563.6%
- #22GLM-5.3 Flash32.8%
- ~#24Hy4 previewEstimated28.9%
- #24DeepSeek V4.1 Flash26.8%
No published score — estimated from related benchmarks
LiveCodeBenchEstimated
~#10 of 28 (estimated)~85.2%
- #1Qwen3.8-Omni-Flash92.6%
- #2Qwen3.8 Flash-Next91.9%
- #9Nemotron 3.5 Super VL NVFP487.4%
- ~#10Hy4 previewEstimated85.2%
- #10Ternary Bonsai 27B82.8%
No published score — estimated from related benchmarks
SciCodeEstimated
~#29 of 119 (estimated)~54.6%
- #1Claude Fable 5.267.7%
- #2Claude Opus 5.566.9%
- #27GPT-5.6 Terra55.0%
- ~#29Hy4 previewEstimated54.6%
- #29GPT-6 Luna54.6%
No published score — estimated from related benchmarks
AA Coding Agent IndexEstimated
~#13 of 26 (estimated)~48.1
- #1Claude Opus 5.561.7
- #2Claude Opus 559.7
- #12Mistral Large 448.3
- ~#13Hy4 previewEstimated48.1
- #13DeepSeek V4.1 Flash47.3
No published score — estimated from related benchmarks
Hard reasoning
GPQA DiamondLab-reported
#17 of 19292.3%
- #1GPT-6 Astra96.1%
- #2Gemini 3.8 Flash95.3%
- #16GPT-5.6 Terra92.5%
- #17Hy4 preview92.3%
- #17Qwen3.8 Flash-Next92.3%
huggingface.co · 2026-08-28
Humanity's Last ExamLab-reported
#24 of 19843.4%
- #1Claude Fable 5.274.3%
- #2Claude Opus 5.561.4%
- #23Muse Spark Open43.9%
- #24Hy4 preview43.4%
- #25DeepSeek V4.1 Flash 101743.3%
huggingface.co · 2026-08-28
AIME (latest)Estimated
~#2 of 34 (estimated)~98.7%
No published score — estimated from related benchmarks
AA Intelligence IndexEstimated
~#36 of 103 (estimated)~37.5
- #1Gemini 4 Carbon59.4
- #2Claude Opus 5.557.6
- #35GPT-6 Luna38.1
- ~#36Hy4 previewEstimated37.5
- #36DeepSeek V4 Pro (0813)36.0
No published score — estimated from related benchmarks