Stockmark Models by Stockmark

StockmarkMITOpen source
1 model2025Up to 32K ctx

Last refreshed 2026-09-08. Next refresh: weekly.

Details

ResearcherStockmark
LicenseMITOSI-approved
Commercial usePermitted
Models1
Released2025
Max context32K

About

Stockmark is the family of Japanese-focused large language models from the Japanese company Stockmark. Stockmark-2-100B-Instruct is a 100-billion-parameter model built from scratch and pretrained on about 2.0 trillion tokens (60% English, 30% Japanese, 10% code), then post-trained with SFT and DPO on synthetic Japanese data to improve instruction following. It supports Japanese and English with a 32K context length, and its development was supported by Japan's GENIAC program. Stockmark's Hugging Face organization also hosts the earlier stockmark-13b and stockmark-100b models and a Stockmark-2-VL-100B-beta vision-language model.

Decision facts

Best fit
coding
Capability starting point
Stockmark 2 100B Instruct with 32K context
Lowest tracked input
Not tracked
Closest related family
360Zhinao

Current Variants

Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.

1 in view

Use when the workload needs 32K context and 100B parameters.

2025-0632K context100B parameters

Release Timeline

1 release group
2025-06
1 current
Stockmark 2 100B Instruct
32K context100B parameters
Current

Specifications(1 models)

Stockmark model specifications comparison
ModelReleasedContextParameters
Stockmark 2 100B Instruct2025-0632K100B

Available From(1 provider)

Popular comparisons in this family