EXAONE 4.5 33B

LG AI Research · lgai-exaone/exaone-4-5-33b

ProviderLG AI Research
Familyexaone
Typevlm
Statusactive
Released2026-04-04
Updated2026-07-06
Parameters34.4B
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/sq625.76 GB~570595.2~855892.8 q4
NVIDIA GB200 Grace Blackwell Superchip372.0 GB16000.0 GB/sbf1668.7 GB~163.0~652.1 q4
AMD Instinct MI355X288.0 GB8000.0 GB/sbf1668.7 GB~81.5~326.1 q4
NVIDIA B200180.0 GB7700.0 GB/sbf1668.7 GB~78.5~313.8 q4
Google TPU7x (Ironwood)192.0 GB7380.0 GB/sbf1668.7 GB~75.2~300.8 q4
AMD Instinct MI325X256.0 GB6000.0 GB/sbf1668.7 GB~61.1~244.5 q4
AMD Instinct MI300X192.0 GB5300.0 GB/sbf1668.7 GB~54.0~216.0 q4
NVIDIA H200 SXM141.0 GB4800.0 GB/sbf1668.7 GB~48.9~195.6 q4
NVIDIA H100 NVL94.0 GB3900.0 GB/sbf1668.7 GB~39.7~159.0 q4
NVIDIA H100 SXM80.0 GB3350.0 GB/sfp834.35 GB~68.3~136.5 q4
AMD Instinct MI250X128.0 GB3200.0 GB/sbf1668.7 GB~32.6~130.4 q4
Google TPU v5p95.0 GB2765.0 GB/sbf1668.7 GB~28.2~112.7 q4
NVIDIA A100 80GB SXM80.0 GB2039.0 GB/sfp834.35 GB~41.6~83.1 q4
NVIDIA H100 PCIe80.0 GB2000.0 GB/sfp834.35 GB~40.8~81.5 q4
NVIDIA GeForce RTX 509032.0 GB1792.0 GB/sq521.47 GB~58.4~73.0 q4
Google TPU v6e (Trillium)32.0 GB1638.0 GB/sq521.47 GB~53.4~66.8 q4
AMD Instinct MI21064.0 GB1600.0 GB/sfp834.35 GB~32.6~65.2 q4
NVIDIA A100 40GB SXM40.0 GB1555.0 GB/sq625.76 GB~42.3~63.4 q4
Google TPU v432.0 GB1200.0 GB/sq521.47 GB~39.1~48.9 q4
NVIDIA GeForce RTX 409024.0 GB1008.0 GB/sq417.18 GB~41.1~41.1 q4
AMD Radeon RX 7900 XTX24.0 GB960.0 GB/sq417.18 GB~39.1~39.1 q4
NVIDIA GeForce RTX 309024.0 GB936.0 GB/sq417.18 GB~38.1~38.1 q4
NVIDIA L40S48.0 GB864.0 GB/sfp834.35 GB~17.6~35.2 q4
Apple M3 Ultra512.0 GB819.0 GB/sbf1668.7 GB~8.3~33.4 q4
Apple M2 Ultra192.0 GB800.0 GB/sbf1668.7 GB~8.2~32.6 q4
Apple M4 Max128.0 GB546.0 GB/sbf1668.7 GB~5.6~22.3 q4
Apple M1 Max64.0 GB400.0 GB/sfp834.35 GB~8.2~16.3 q4
Apple M2 Max96.0 GB400.0 GB/sbf1668.7 GB~4.1~16.3 q4
Apple M3 Max128.0 GB400.0 GB/sbf1668.7 GB~4.1~16.3 q4
NVIDIA L424.0 GB300.0 GB/sq417.18 GB~12.2~12.2 q4
Apple M4 Pro64.0 GB273.0 GB/sfp834.35 GB~5.6~11.1 q4
Apple M432.0 GB120.0 GB/sq521.47 GB~3.9~4.9 q4

Where it runs

Taken from the card's availability section.

PlatformId on that platform
Hugging FaceLGAI-EXAONE/EXAONE-4.5-33B

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_lcrEXAONE 4.5 33B54.67%2026-09-10 evaluatedindependent evaluatorsource
critptEXAONE 4.5 33B0.29%2026-09-10 evaluatedindependent evaluatorsource
gdpval_aaEXAONE 4.5 33B6.38%2026-09-10 evaluatedindependent evaluatorsource
gpqa_diamondEXAONE 4.5 33B79.39%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