Kimi Linear 48B A3B Instruct

Moonshot AI · moonshot/kimi-linear-48b-a3b-instruct

ProviderMoonshot AI
Familykimi
Typellm-chat
Statusactive
Released2025-10-30
Updated2025-12-16
Parameters49.1B
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. This is a mixture-of-experts model and no card carries active-parameter counts, so these predictions use total parameters and understate the real speed.
DeviceMemoryBandwidthBest qualityWeightstok/sFastest
Cerebras WSE-344.0 GB21000000.0 GB/sq530.7 GB~478801.2~598501.5 q4
NVIDIA GB200 Grace Blackwell Superchip372.0 GB16000.0 GB/sbf1698.25 GB~114.0~456.0 q4
AMD Instinct MI355X288.0 GB8000.0 GB/sbf1698.25 GB~57.0~228.0 q4
NVIDIA B200180.0 GB7700.0 GB/sbf1698.25 GB~54.9~219.5 q4
Google TPU7x (Ironwood)192.0 GB7380.0 GB/sbf1698.25 GB~52.6~210.3 q4
AMD Instinct MI325X256.0 GB6000.0 GB/sbf1698.25 GB~42.8~171.0 q4
AMD Instinct MI300X192.0 GB5300.0 GB/sbf1698.25 GB~37.8~151.1 q4
NVIDIA H200 SXM141.0 GB4800.0 GB/sbf1698.25 GB~34.2~136.8 q4
NVIDIA H100 NVL94.0 GB3900.0 GB/sfp849.12 GB~55.6~111.2 q4
NVIDIA H100 SXM80.0 GB3350.0 GB/sfp849.12 GB~47.7~95.5 q4
AMD Instinct MI250X128.0 GB3200.0 GB/sfp849.12 GB~45.6~91.2 q4
Google TPU v5p95.0 GB2765.0 GB/sfp849.12 GB~39.4~78.8 q4
NVIDIA A100 80GB SXM80.0 GB2039.0 GB/sfp849.12 GB~29.1~58.1 q4
NVIDIA H100 PCIe80.0 GB2000.0 GB/sfp849.12 GB~28.5~57.0 q4
AMD Instinct MI21064.0 GB1600.0 GB/sq636.84 GB~30.4~45.6 q4
NVIDIA A100 40GB SXM40.0 GB1555.0 GB/sq424.56 GB~44.3~44.3 q4
NVIDIA L40S48.0 GB864.0 GB/sq530.7 GB~19.7~24.6 q4
Apple M3 Ultra512.0 GB819.0 GB/sbf1698.25 GB~5.8~23.3 q4
Apple M2 Ultra192.0 GB800.0 GB/sbf1698.25 GB~5.7~22.8 q4
Apple M4 Max128.0 GB546.0 GB/sfp849.12 GB~7.8~15.6 q4
Apple M1 Max64.0 GB400.0 GB/sq636.84 GB~7.6~11.4 q4
Apple M2 Max96.0 GB400.0 GB/sfp849.12 GB~5.7~11.4 q4
Apple M3 Max128.0 GB400.0 GB/sfp849.12 GB~5.7~11.4 q4
Apple M4 Pro64.0 GB273.0 GB/sq636.84 GB~5.2~7.8 q4

Where it runs

Taken from the card's availability section.

PlatformId on that platform
Hugging Facemoonshotai/Kimi-Linear-48B-A3B-Instruct

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_lcrKimi Linear 48B A3B Instruct28.0%2026-09-10 evaluatedindependent evaluatorsource
gpqa_diamondKimi Linear 48B A3B Instruct41.21%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