HyperCLOVA X SEED Think 32B

NAVER · naver/hyperclovax-seed-think-32b

ProviderNAVER
Familyhyperclova
Typellm-reasoning
Statusactive
Released2025-12-23
Updated2026-07-14
Parameters33.3B
Open weightsyes

What it fits on

Ordered by predicted speed. Bandwidth sets decode rate; memory decides whether it runs at all.

Every figure here is computed, not measured. Fit is weights at each quantisation against device memory, with a 25% allowance for the KV cache, activations and the OS. Decode rate is the memory-bandwidth roofline at 70% efficiency. Nobody has run this model on these devices.
DeviceMemoryBandwidthBest qualityWeightstok/sFastest
Cerebras WSE-344.0 GB21000000.0 GB/sq624.99 GB~588351.7~882527.5 q4
NVIDIA GB200 Grace Blackwell Superchip372.0 GB16000.0 GB/sbf1666.63 GB~168.1~672.4 q4
AMD Instinct MI355X288.0 GB8000.0 GB/sbf1666.63 GB~84.1~336.2 q4
NVIDIA B200180.0 GB7700.0 GB/sbf1666.63 GB~80.9~323.6 q4
Google TPU7x (Ironwood)192.0 GB7380.0 GB/sbf1666.63 GB~77.5~310.1 q4
AMD Instinct MI325X256.0 GB6000.0 GB/sbf1666.63 GB~63.0~252.2 q4
AMD Instinct MI300X192.0 GB5300.0 GB/sbf1666.63 GB~55.7~222.7 q4
NVIDIA H200 SXM141.0 GB4800.0 GB/sbf1666.63 GB~50.4~201.7 q4
NVIDIA H100 NVL94.0 GB3900.0 GB/sbf1666.63 GB~41.0~163.9 q4
NVIDIA H100 SXM80.0 GB3350.0 GB/sfp833.31 GB~70.4~140.8 q4
AMD Instinct MI250X128.0 GB3200.0 GB/sbf1666.63 GB~33.6~134.5 q4
Google TPU v5p95.0 GB2765.0 GB/sbf1666.63 GB~29.0~116.2 q4
NVIDIA A100 80GB SXM80.0 GB2039.0 GB/sfp833.31 GB~42.8~85.7 q4
NVIDIA H100 PCIe80.0 GB2000.0 GB/sfp833.31 GB~42.0~84.1 q4
NVIDIA GeForce RTX 509032.0 GB1792.0 GB/sq520.82 GB~60.2~75.3 q4
Google TPU v6e (Trillium)32.0 GB1638.0 GB/sq520.82 GB~55.1~68.8 q4
AMD Instinct MI21064.0 GB1600.0 GB/sfp833.31 GB~33.6~67.2 q4
NVIDIA A100 40GB SXM40.0 GB1555.0 GB/sq624.99 GB~43.6~65.3 q4
Google TPU v432.0 GB1200.0 GB/sq520.82 GB~40.3~50.4 q4
NVIDIA GeForce RTX 409024.0 GB1008.0 GB/sq416.66 GB~42.4~42.4 q4
AMD Radeon RX 7900 XTX24.0 GB960.0 GB/sq416.66 GB~40.3~40.3 q4
NVIDIA GeForce RTX 309024.0 GB936.0 GB/sq416.66 GB~39.3~39.3 q4
NVIDIA L40S48.0 GB864.0 GB/sfp833.31 GB~18.2~36.3 q4
Apple M3 Ultra512.0 GB819.0 GB/sbf1666.63 GB~8.6~34.4 q4
Apple M2 Ultra192.0 GB800.0 GB/sbf1666.63 GB~8.4~33.6 q4
Apple M4 Max128.0 GB546.0 GB/sbf1666.63 GB~5.7~22.9 q4
Apple M1 Max64.0 GB400.0 GB/sfp833.31 GB~8.4~16.8 q4
Apple M2 Max96.0 GB400.0 GB/sbf1666.63 GB~4.2~16.8 q4
Apple M3 Max128.0 GB400.0 GB/sbf1666.63 GB~4.2~16.8 q4
NVIDIA L424.0 GB300.0 GB/sq416.66 GB~12.6~12.6 q4
Apple M4 Pro64.0 GB273.0 GB/sfp833.31 GB~5.7~11.5 q4
Apple M432.0 GB120.0 GB/sq520.82 GB~4.0~5.0 q4

Where it runs

Taken from the card's availability section.

PlatformId on that platform
Hugging Facenaver-hyperclovax/HyperCLOVAX-SEED-Think-32B

Verified benchmark evidence

Each row was checked against its source by a reviewer, and carries the model identifier as actually evaluated — which is not always the same as this card's.
BenchmarkEvaluated asScoreEvidence dateSource kind
aa_lcrHyperCLOVA X SEED Think (32B)13.67%2026-09-10 evaluatedindependent evaluatorsource
critptHyperCLOVA X SEED Think (32B)0.0%2026-09-10 evaluatedindependent evaluatorsource
gpqa_diamondHyperCLOVA X SEED Think (32B)61.52%2026-09-10 evaluatedindependent evaluatorsource

Reported benchmark scores

This card reports no benchmark scores yet.

BenchmarkCatalogue standingScore

Data

This card as JSON · See it in the graph · Edit on GitHub