Hy4 preview hy4-preview

tencent · hunyuan

unranked — no quality signals yet

params
780B
arch
moe
context
1024k
license
apache-2.0 open weights
released
Aug 2026
downloads/30d
16.6k

# architecture

attn shared GQA 64q/8kv router top-8 of 256 +1 shared expert ×256 expert out ×78 layers ctx 1,048,576 780B pool
layers
78
d_model
6144
heads
64q / 8kv
head dim
64
experts
top-8 of 256 + 1 shared
vocab
120832
family
hy_v4

signal path of one layer, generated from the registry's structured fields — dimension lines quote the model's real numbers; an MoE trace forks at the router, a dense trace runs straight through. Geometry fields come from the repo's config.json.

# trend

collecting history — trends appear after a few daily runs.

# benchmarks

benchmarkscoresourcedate
Agents-last-exam 0.2 / 1 LLM Stats —
Apex-agents 0.4 / 1 LLM Stats —
Arxivmath 0.7 / 1 LLM Stats —
Automationbench 0.3 / 1 LLM Stats —
Bankertoolbench 0.8 / 1 LLM Stats —
Biomysterybench 0.7 / 1 LLM Stats —
Gdpval-aa 1678.0 / 3000 LLM Stats —
LMArena Elo (Code) 1632.0 LMArena Oct 2026

# where to run

providerquantctx$/M in$/M out$/M cacheprice srctpsuptime✓
Cohereunknown 256k $0.83$2.50$0.04 via openrouter — 99.4%
DeepInfrafp4 1024k $0.83$2.50— deepinfra — 100.0% ✓
Novitafp8 1000k $0.83$2.50— novita — 100.0% ✓
Opencode Go 1000k $0.83$2.50— via litellm — —
SiliconFlowfp8 1024k $0.83$2.50$0.04 openrouter — 100.0% ✓
Tencentfp8 1024k $0.83$2.50$0.04 via openrouter — 100.0%
Aihubmix 1024k $0.84$2.53— via litellm — —
Venicefp4 1024k $1.04$3.13— venice — — ✓

sorted by blended price ((3·input + output) / 4 per 1M) · ✓ = the provider's own catalog confirms the offer · "via …" prices are what the aggregator routing the offer charges, not the provider's own list price

source aliases
openrouter
tencent/hy4-preview