News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
Stockmark-100b

Stockmark-100b

A 100 billion parameter Japanese and English language model.

Stockmark Inc. developed Stockmark-100b, a large language model with 100 billion parameters. It was trained on a mix of Japanese and English text to support business and general queries, releasing its weights under an open licence.

Share on WhatsApp

  • Contains 100 billion parameters.
  • Trained on 910 billion tokens.
  • Released under MIT licence.
  • Scored 5.97 on Japanese Vicuna QA.
  • Supported by GENIAC initiative.

What it is

Stockmark-100b is a large language model containing 100 billion parameters. Stockmark Inc. built the system from scratch, focusing on both Japanese and English languages. The company designed it to handle general text tasks and specific business inquiries. It serves as a foundational model for developers who need strong performance in Japanese contexts while retaining English capabilities. The project received support from the GENIAC initiative.

How it came about

Stockmark Inc. developed the model using internal resources and public datasets. The training process relied on the GENIAC programme for support. The company released the base model alongside an instruct-tuned variant called stockmark-100b-instruct-v0.1. These releases allow the community to use and adapt the technology for various applications.

Inside the model

The model uses a decoder-only transformer architecture. Training involved approximately 910 billion tokens of text data. This corpus combined English samples from RedPajama-Data with Japanese sources. The Japanese data included 203.7 billion tokens from CommonCrawl snapshots between 2020 and 2024. An additional 8.8 billion tokens came from a proprietary Stockmark Web Corpus focused on business content. Training ran for seven weeks on 48 nodes of NVIDIA H100 GPUs using Megatron-LM.

What it can do

In internal evaluations, the instruct version scored 0.90 on the Stockmark Business Questions dataset. It outperformed the smaller 13b variant, which scored 0.80. On the Japanese Vicuna QA Benchmark, it achieved a score of 5.97. This result exceeded Swallow-70b-instruct and GPT-3.5. The model generated 100 Japanese characters in 1.86 seconds on AWS Inferentia2 hardware, showing faster inference than some competitors.

How to use it

Developers can access the model weights via Hugging Face. Stockmark Inc. distributes the software under the MIT licence. This permissive terms allows for broad commercial and research use. Users can load the model using standard libraries like Transformers and run it on compatible GPU hardware.

Questions people ask about Stockmark-100b

What languages does Stockmark-100b support?

The model supports both Japanese and English. It was pretrained on a large corpus containing text from both languages to ensure proficiency in each.

Is Stockmark-100b free to use?

Yes, the model weights are available under the MIT licence. This open licence permits users to download, modify, and use the model for commercial or research purposes.

How was Stockmark-100b trained?

It was trained from scratch on 910 billion tokens. The process used 48 nodes of NVIDIA H100 GPUs and took approximately seven weeks to complete using the Megatron-LM library.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Stockmark-100b yet, so the facts on this page come from the sources listed above: huggingface.co/stockmark/stockmark-100b. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.