Tri-21B-Think

Trillion Labs · trillionlabs/tri-21b-think

ProviderTrillion Labs
Familytri
Typellm-reasoning
Statusactive
Released2026-02-18
Updated2026-02-19
Parameters20.7B
Open weightsyes

Lineage

RelationshipModel
Is a finetune oftrillionlabs/Tri-21B

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/sfp820.73 GB~709270.5~1418541.1 q4
NVIDIA GB200 Grace Blackwell Superchip372.0 GB16000.0 GB/sbf1641.45 GB~270.2~1080.8 q4
AMD Instinct MI355X288.0 GB8000.0 GB/sbf1641.45 GB~135.1~540.4 q4
NVIDIA B200180.0 GB7700.0 GB/sbf1641.45 GB~130.0~520.1 q4
Google TPU7x (Ironwood)192.0 GB7380.0 GB/sbf1641.45 GB~124.6~498.5 q4
AMD Instinct MI325X256.0 GB6000.0 GB/sbf1641.45 GB~101.3~405.3 q4
AMD Instinct MI300X192.0 GB5300.0 GB/sbf1641.45 GB~89.5~358.0 q4
NVIDIA H200 SXM141.0 GB4800.0 GB/sbf1641.45 GB~81.1~324.2 q4
NVIDIA H100 NVL94.0 GB3900.0 GB/sbf1641.45 GB~65.9~263.4 q4
NVIDIA H100 SXM80.0 GB3350.0 GB/sbf1641.45 GB~56.6~226.3 q4
AMD Instinct MI250X128.0 GB3200.0 GB/sbf1641.45 GB~54.0~216.2 q4
Google TPU v5p95.0 GB2765.0 GB/sbf1641.45 GB~46.7~186.8 q4
NVIDIA A100 80GB SXM80.0 GB2039.0 GB/sbf1641.45 GB~34.4~137.7 q4
NVIDIA H100 PCIe80.0 GB2000.0 GB/sbf1641.45 GB~33.8~135.1 q4
NVIDIA GeForce RTX 509032.0 GB1792.0 GB/sfp820.73 GB~60.5~121.0 q4
Google TPU v6e (Trillium)32.0 GB1638.0 GB/sfp820.73 GB~55.3~110.6 q4
AMD Instinct MI21064.0 GB1600.0 GB/sbf1641.45 GB~27.0~108.1 q4
NVIDIA A100 40GB SXM40.0 GB1555.0 GB/sfp820.73 GB~52.5~105.0 q4
Google TPU v432.0 GB1200.0 GB/sfp820.73 GB~40.5~81.1 q4
NVIDIA GeForce RTX 409024.0 GB1008.0 GB/sq615.54 GB~45.4~68.1 q4
AMD Radeon RX 7900 XTX24.0 GB960.0 GB/sq615.54 GB~43.2~64.8 q4
NVIDIA GeForce RTX 508016.0 GB960.0 GB/sq410.36 GB~64.8~64.8 q4
NVIDIA GeForce RTX 309024.0 GB936.0 GB/sq615.54 GB~42.2~63.2 q4
NVIDIA L40S48.0 GB864.0 GB/sfp820.73 GB~29.2~58.4 q4
Apple M3 Ultra512.0 GB819.0 GB/sbf1641.45 GB~13.8~55.3 q4
AMD Radeon RX 7900 XT20.0 GB800.0 GB/sq512.95 GB~43.2~54.0 q4
Apple M2 Ultra192.0 GB800.0 GB/sbf1641.45 GB~13.5~54.0 q4
Google TPU v5e16.0 GB800.0 GB/sq410.36 GB~54.0~54.0 q4
NVIDIA GeForce RTX 4080 SUPER16.0 GB736.0 GB/sq410.36 GB~49.7~49.7 q4
NVIDIA GeForce RTX 4070 Ti SUPER16.0 GB672.0 GB/sq410.36 GB~45.4~45.4 q4
AMD Radeon RX 9070 XT16.0 GB640.0 GB/sq410.36 GB~43.2~43.2 q4
Apple M4 Max128.0 GB546.0 GB/sbf1641.45 GB~9.2~36.9 q4
Apple M1 Max64.0 GB400.0 GB/sbf1641.45 GB~6.8~27.0 q4
Apple M2 Max96.0 GB400.0 GB/sbf1641.45 GB~6.8~27.0 q4
Apple M3 Max128.0 GB400.0 GB/sbf1641.45 GB~6.8~27.0 q4
NVIDIA L424.0 GB300.0 GB/sq615.54 GB~13.5~20.3 q4
NVIDIA GeForce RTX 4060 Ti 16GB16.0 GB288.0 GB/sq410.36 GB~19.5~19.5 q4
Apple M4 Pro64.0 GB273.0 GB/sbf1641.45 GB~4.6~18.4 q4
Apple M432.0 GB120.0 GB/sfp820.73 GB~4.1~8.1 q4

Where it runs

Taken from the card's availability section.

PlatformId on that platform
Hugging Facetrillionlabs/Tri-21B-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_lcrTri-21B-Think15.67%2026-09-10 evaluatedindependent evaluatorsource
critptTri-21B-Think0.29%2026-09-10 evaluatedindependent evaluatorsource
gpqa_diamondTri-21B-Think60.1%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