Sarvam M

Sarvam AI · sarvam/sarvam-m

ProviderSarvam AI
Familysarvam
Typellm-chat
Statusactive
Released2025-05-20
Updated2025-05-28
Parameters23.6B
Open weightsyes

Lineage

RelationshipModel
Is a finetune ofmistralai/Mistral-Small-3.1-24B-Base-2503

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/sfp823.57 GB~623610.6~1247221.2 q4
NVIDIA GB200 Grace Blackwell Superchip372.0 GB16000.0 GB/sbf1647.14 GB~237.6~950.3 q4
AMD Instinct MI355X288.0 GB8000.0 GB/sbf1647.14 GB~118.8~475.1 q4
NVIDIA B200180.0 GB7700.0 GB/sbf1647.14 GB~114.3~457.3 q4
Google TPU7x (Ironwood)192.0 GB7380.0 GB/sbf1647.14 GB~109.6~438.3 q4
AMD Instinct MI325X256.0 GB6000.0 GB/sbf1647.14 GB~89.1~356.3 q4
AMD Instinct MI300X192.0 GB5300.0 GB/sbf1647.14 GB~78.7~314.8 q4
NVIDIA H200 SXM141.0 GB4800.0 GB/sbf1647.14 GB~71.3~285.1 q4
NVIDIA H100 NVL94.0 GB3900.0 GB/sbf1647.14 GB~57.9~231.6 q4
NVIDIA H100 SXM80.0 GB3350.0 GB/sbf1647.14 GB~49.7~199.0 q4
AMD Instinct MI250X128.0 GB3200.0 GB/sbf1647.14 GB~47.5~190.1 q4
Google TPU v5p95.0 GB2765.0 GB/sbf1647.14 GB~41.1~164.2 q4
NVIDIA A100 80GB SXM80.0 GB2039.0 GB/sbf1647.14 GB~30.3~121.1 q4
NVIDIA H100 PCIe80.0 GB2000.0 GB/sbf1647.14 GB~29.7~118.8 q4
NVIDIA GeForce RTX 509032.0 GB1792.0 GB/sfp823.57 GB~53.2~106.4 q4
Google TPU v6e (Trillium)32.0 GB1638.0 GB/sfp823.57 GB~48.6~97.3 q4
AMD Instinct MI21064.0 GB1600.0 GB/sbf1647.14 GB~23.8~95.0 q4
NVIDIA A100 40GB SXM40.0 GB1555.0 GB/sfp823.57 GB~46.2~92.4 q4
Google TPU v432.0 GB1200.0 GB/sfp823.57 GB~35.6~71.3 q4
NVIDIA GeForce RTX 409024.0 GB1008.0 GB/sq617.68 GB~39.9~59.9 q4
AMD Radeon RX 7900 XTX24.0 GB960.0 GB/sq617.68 GB~38.0~57.0 q4
NVIDIA GeForce RTX 508016.0 GB960.0 GB/sq411.79 GB~57.0~57.0 q4
NVIDIA GeForce RTX 309024.0 GB936.0 GB/sq617.68 GB~37.1~55.6 q4
NVIDIA L40S48.0 GB864.0 GB/sfp823.57 GB~25.7~51.3 q4
Apple M3 Ultra512.0 GB819.0 GB/sbf1647.14 GB~12.2~48.6 q4
AMD Radeon RX 7900 XT20.0 GB800.0 GB/sq514.73 GB~38.0~47.5 q4
Apple M2 Ultra192.0 GB800.0 GB/sbf1647.14 GB~11.9~47.5 q4
Google TPU v5e16.0 GB800.0 GB/sq411.79 GB~47.5~47.5 q4
NVIDIA GeForce RTX 4080 SUPER16.0 GB736.0 GB/sq411.79 GB~43.7~43.7 q4
NVIDIA GeForce RTX 4070 Ti SUPER16.0 GB672.0 GB/sq411.79 GB~39.9~39.9 q4
AMD Radeon RX 9070 XT16.0 GB640.0 GB/sq411.79 GB~38.0~38.0 q4
Apple M4 Max128.0 GB546.0 GB/sbf1647.14 GB~8.1~32.4 q4
Apple M1 Max64.0 GB400.0 GB/sbf1647.14 GB~5.9~23.8 q4
Apple M2 Max96.0 GB400.0 GB/sbf1647.14 GB~5.9~23.8 q4
Apple M3 Max128.0 GB400.0 GB/sbf1647.14 GB~5.9~23.8 q4
NVIDIA L424.0 GB300.0 GB/sq617.68 GB~11.9~17.8 q4
NVIDIA GeForce RTX 4060 Ti 16GB16.0 GB288.0 GB/sq411.79 GB~17.1~17.1 q4
Apple M4 Pro64.0 GB273.0 GB/sbf1647.14 GB~4.1~16.2 q4
Apple M432.0 GB120.0 GB/sfp823.57 GB~3.6~7.1 q4

Where it runs

Taken from the card's availability section.

PlatformId on that platform
Hugging Facesarvamai/sarvam-m

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_lcrSarvam M0.0%2026-09-10 evaluatedindependent evaluatorsource
critptSarvam M0.0%2026-09-10 evaluatedindependent evaluatorsource
gpqa_diamondSarvam M41.62%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