Hy3 hy3

tencent · hunyuan

unranked — no quality signals yet

params
298.8B
arch
moe
context
256k
license
apache-2.0 open weights
released
Jul 2026
downloads/30d
1.3M

# architecture

attn shared GQA 64q/8kv router top-8 of 192 +1 shared expert ×192 expert out ×80 layers ctx 262,144 298.8B pool
layers
80
d_model
4096
heads
64q / 8kv
head dim
128
experts
top-8 of 192 + 1 shared
vocab
120832
family
hy_v3

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.

# why ranked

Overall: provisional — score 69.8 ● low — a single signal; treat with caution

provisional: too few independent signals for a numbered position — the score is shown, but this model sorts after every ranked model.

signalweightinput (0–100)
aa_intelligence 0.70 —
bench_composite 0.30 69.8

benchmark panel evidence:

complete panelscore (0–100)
browsecomp-v169.8

missing signals are dropped and the remaining weights renormalized — never imputed.

full methodology

# trend

OGM score · last 108 days
OGM score over 108 days
overall rank (up is better)
Overall rank over 108 days

methodology changed during this history window; score movement across that boundary is not model movement. See methodology v6.

# benchmarks

benchmarkscoresourcedate
Aa-lcr 0.7 / 1 LLM Stats —
Apex-agents 0.3 / 1 LLM Stats —
Arxivmath 0.5 / 1 LLM Stats —
Browsecomp 0.8 / 1 LLM Stats —
Cl-bench 0.2 / 1 LLM Stats —
Claw-eval 0.7 / 1 LLM Stats —
Usamo-2026 30.2 / 42 LLM Stats —

# where to run

providerquantctx$/M in$/M out$/M cacheprice srctpsuptime✓
Featherless — ——— via hfrouter — —
SiliconFlow — ——— siliconflow — — ✓
DeepInfra 256k $0.13$0.53— via hfrouter 42 —
Novita 256k $0.14$0.58— novita — — ✓
Opencode Go 192k $0.14$0.58— via litellm — —
Phala 256k $0.15$0.64— phala — — ✓
Aihubmix 250k $0.16$0.62— via litellm — —
Kyma 256k $0.16$0.66— via litellm — —
AtlasCloudfp8 256k $0.20$0.80— atlascloud — — ✓

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
hf
tencent/Hy3-FP8