Olmo 3 32B Think

Allen Institute for AI · allen-ai/olmo-3-32b-think

ProviderAllen Institute for AI
Familyolmo
Typellm-reasoning
Statusactive
Released2025-11-19
Updated2026-06-25
Parameters32.2B
Open weightsyes

Lineage

RelationshipModel
Is a finetune ofallenai/Olmo-3-32B-Think-DPO

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/sfp832.23 GB~456047.0~912093.9 q4
NVIDIA GB200 Grace Blackwell Superchip372.0 GB16000.0 GB/sbf1664.47 GB~173.7~694.9 q4
AMD Instinct MI355X288.0 GB8000.0 GB/sbf1664.47 GB~86.9~347.5 q4
NVIDIA B200180.0 GB7700.0 GB/sbf1664.47 GB~83.6~334.4 q4
Google TPU7x (Ironwood)192.0 GB7380.0 GB/sbf1664.47 GB~80.1~320.5 q4
AMD Instinct MI325X256.0 GB6000.0 GB/sbf1664.47 GB~65.1~260.6 q4
AMD Instinct MI300X192.0 GB5300.0 GB/sbf1664.47 GB~57.5~230.2 q4
NVIDIA H200 SXM141.0 GB4800.0 GB/sbf1664.47 GB~52.1~208.5 q4
NVIDIA H100 NVL94.0 GB3900.0 GB/sbf1664.47 GB~42.3~169.4 q4
NVIDIA H100 SXM80.0 GB3350.0 GB/sfp832.23 GB~72.8~145.5 q4
AMD Instinct MI250X128.0 GB3200.0 GB/sbf1664.47 GB~34.7~139.0 q4
Google TPU v5p95.0 GB2765.0 GB/sbf1664.47 GB~30.0~120.1 q4
NVIDIA A100 80GB SXM80.0 GB2039.0 GB/sfp832.23 GB~44.3~88.6 q4
NVIDIA H100 PCIe80.0 GB2000.0 GB/sfp832.23 GB~43.4~86.9 q4
NVIDIA GeForce RTX 509032.0 GB1792.0 GB/sq520.15 GB~62.3~77.8 q4
Google TPU v6e (Trillium)32.0 GB1638.0 GB/sq520.15 GB~56.9~71.1 q4
AMD Instinct MI21064.0 GB1600.0 GB/sfp832.23 GB~34.7~69.5 q4
NVIDIA A100 40GB SXM40.0 GB1555.0 GB/sq624.18 GB~45.0~67.5 q4
Google TPU v432.0 GB1200.0 GB/sq520.15 GB~41.7~52.1 q4
NVIDIA GeForce RTX 409024.0 GB1008.0 GB/sq416.12 GB~43.8~43.8 q4
AMD Radeon RX 7900 XTX24.0 GB960.0 GB/sq416.12 GB~41.7~41.7 q4
NVIDIA GeForce RTX 309024.0 GB936.0 GB/sq416.12 GB~40.7~40.7 q4
NVIDIA L40S48.0 GB864.0 GB/sfp832.23 GB~18.8~37.5 q4
Apple M3 Ultra512.0 GB819.0 GB/sbf1664.47 GB~8.9~35.6 q4
Apple M2 Ultra192.0 GB800.0 GB/sbf1664.47 GB~8.7~34.7 q4
Apple M4 Max128.0 GB546.0 GB/sbf1664.47 GB~5.9~23.7 q4
Apple M1 Max64.0 GB400.0 GB/sfp832.23 GB~8.7~17.4 q4
Apple M2 Max96.0 GB400.0 GB/sbf1664.47 GB~4.3~17.4 q4
Apple M3 Max128.0 GB400.0 GB/sbf1664.47 GB~4.3~17.4 q4
NVIDIA L424.0 GB300.0 GB/sq416.12 GB~13.0~13.0 q4
Apple M4 Pro64.0 GB273.0 GB/sfp832.23 GB~5.9~11.9 q4
Apple M432.0 GB120.0 GB/sq520.15 GB~4.2~5.2 q4

Where it runs

Taken from the card's availability section.

PlatformId on that platform
Hugging Faceallenai/Olmo-3-32B-Think

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_lcrOlmo 3 32B Think0.0%2026-09-10 evaluatedindependent evaluatorsource
critptOlmo 3 32B Think0.0%2026-09-10 evaluatedindependent evaluatorsource
gpqa_diamondOlmo 3 32B Think61.01%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