CyberAgentLM
CyberAgent's open-weight Japanese LLMs, led by the 22B CALM3
Overview
CyberAgentLM (CALM) is CyberAgent's family of open-weight Japanese language models, from the 2023 OpenCALM models (CC BY-SA 4.0) and CALM2-7B to the 22B CALM3-22B-Chat of July 2024 (Apache 2.0), trained from scratch on Japanese and English. CyberAgent has since added CAT reasoning and translation models built on other bases.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
Use cases
What you can produce with CyberAgentLM
- Open weights for CALM3-22B-Chat, a 22B Japanese-English chat model with a 16,384-token context
- Apache 2.0 licensing on CALM2 and CALM3 for commercial use and fine-tuning
- Smaller OpenCALM models from 160M to 6.8B parameters (CC BY-SA 4.0)
- CALM2-7B-Chat with a 32,768-token context
- LLaVA-CALM2-SigLIP, a Japanese vision-language model
- CAT-Thinking and CAT-Paws 8B models that reason in Japanese (Apache 2.0)
- CAT-Translate Japanese-English translation models from 0.8B to 7B (MIT)
ASEAN Perspective
CyberAgentLM in Southeast Asia
CyberAgentLM is a Japan-first model with strong Japanese and English bilingual capability, making it directly relevant to Japanese-language operations across ASEAN — particularly in Singapore, Malaysia, and Indonesia where Japanese multinationals maintain significant business presence. The Apache 2.0 licence allows regional enterprises to deploy entirely on-premises, which aligns well with data-sovereignty requirements in markets like Indonesia and Thailand. However, the model has no demonstrated training data or fine-tuning for Malay, Indonesian, Thai, Vietnamese, or other ASEAN languages, so organisations seeking broad regional language coverage will need to look elsewhere or supplement with fine-tuning.
CALM3-22B-Chat is still a respectable open Japanese model: trained from scratch, Apache 2.0, and by CyberAgent's own July 2024 measure level with Llama-3-70B-Instruct on a Japanese leaderboard. But the CALM line has not moved since 2024, and CyberAgent's newer releases, the CAT models, sit on Qwen3 and other bases.
There is no hosted API, so the 22B model needs about 45GB of GPU memory at 16-bit precision, and there is nothing for Southeast Asian languages. Worth a look for Japanese teams with their own GPUs; the newer CAT models are the better place to watch.
What we found
CyberAgentLM, shortened to CALM, is the family of Japanese language models released by CyberAgent, the Tokyo internet advertising, media and games group. The first CyberAgentLM went public in May 2023 as OpenCALM, six models from 160 million to 6.8 billion parameters under a CC BY-SA 4.0 licence. CALM2 followed in November 2023: a 7-billion-parameter model trained on 1.3 trillion tokens, with a chat version handling 32,768 tokens of context, under Apache 2.0. Earlier versions of this page put the whole family under Apache 2.0; OpenCALM's share-alike licence asks more of anyone redistributing a modified model.
The flagship is still CALM3-22B-Chat, released on 9 July 2024. CyberAgent trained it from scratch on 2 trillion Japanese and English tokens rather than adapting an English base model, gave it a 16,384-token context and a 65,024-entry vocabulary, and said it matched Meta's Llama-3-70B-Instruct on the Nejumi LLM Leaderboard 3 as of July 2024. That comparison is CyberAgent's own. More than two years later there is no CALM4; the only later CALM3 release is an experimental variant tuned in February 2025 on a Japanese safety dataset.
CyberAgent's newer open models use other foundations. CAT-Thinking-8B and CAT-Paws-8B, published in May and June 2026 under Apache 2.0, start from Qwen3-Swallow, a Japanese-adapted Qwen3-8B, and reason in Japanese; CAT-Paws targets coding and terminal agents, and its card warns that a 40k context limits it. CAT-Translate, four MIT-licensed models from 0.8B to 7B, translates between Japanese and English, and CyberAgent's card claims it beats TranslateGemma-12B.
The team behind them, CyberAgent AI Lab, dates to 2016, and CyberAgent cites a 2022 Thundermark Capital ranking that placed it fourth in Japan. Every model is free to download from Hugging Face. CyberAgent runs no hosted API, so the 22-billion-parameter flagship needs roughly 45GB of GPU memory at 16-bit precision before any quantisation. Coverage is Japanese and English only.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- CALM3-22B achieves performance equivalent to Meta's 70B Llama model — despite having roughly one-third the parameters.
- CyberAgent AI Lab was founded in 2016 making it one of the earliest AI research labs established by an advertising company anywhere in the world.
- The parent company CyberAgent also owns Abema, Japan's major internet TV service, and reports annual revenue above ¥800 billion (~US$5 billion).
- CALM3 was trained from scratch on 2 trillion tokens rather than being fine-tuned from an existing English base model — a deliberate design choice for Japanese tokenisation quality.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including CyberAgentLM's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with CyberAgentLM unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to CyberAgentLM directly →
Spotted something out of date? Suggest an update →
Alternatives to CyberAgentLM
More in LLMs & Chat
