CyberAgentLM
Open Japanese LLMs from Japan's biggest ad group — Apache 2.0
Overview
CyberAgentLM is CyberAgent's family of open-weight Japanese LLMs, from 160M-parameter OpenCALM (2023) to the 22B CALM3 flagship (2024), released under Apache 2.0 for free commercial self-hosting; CALM3-22B-Chat benchmarks near Llama-3-70B-Instruct at under a third the size.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 11 Jul 2026 and may be out of date or incomplete.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.
- Apache 2.0 — commercial use allowed
- CALM3-22B rivals much larger models
- No fees, sign-up, or rate limits
- Runs on vLLM, Transformers, GGUF ports
- No hosted API from CyberAgent
- Fine-tuning permitted under licence
Use cases
What you can produce with CyberAgentLM
- Open-weights CALM3-22B-Chat model files (BF16 ~45GB; 4-bit quant ~13GB) downloadable from Hugging Face
- Apache 2.0 licence permitting commercial deployment, modification, and redistribution without royalties
- ChatML-formatted instruction-tuned model ready for conversational use out of the box
- Compatibility with vLLM and SGLang for production-grade, OpenAI-compatible inference server setup
- Earlier model tiers (OpenCALM 160M–6.8B; CALM2 7B) for resource-constrained or edge deployments
- CyberAgent AI Lab research papers and benchmark results on GitHub (CyberAgentAILab org)
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 remains one of the stronger open-weight Japanese LLMs available, matching Meta's Llama-3-70B-Instruct on the Nejumi leaderboard at less than a third the parameter count — a genuinely good result for a 2024 release CyberAgent hasn't superseded since. Apache 2.0 licensing means no fees and no rate limits, and from-scratch training (rather than fine-tuning an English base model) gives it a Japanese-optimized ~60K-token vocabulary. CyberAgent AI Lab, the team behind it, ranks 4th in Japan for AI research output, behind only NTT and Fujitsu among non-pure-AI companies. The catch is infrastructure: there's no hosted API, so the 22B model's ~45GB BF16 footprint (or ~13GB at 4-bit) is entirely the user's problem. English and ASEAN-language coverage is essentially nonexistent, and community support is thinner than the Llama or Mistral ecosystems. Best suited to Japanese enterprises with in-house ML teams willing to run their own inference.
What people say
22.5 billion parameters, trained from scratch on 2 trillion Japanese and English tokens, and it still holds up against models three times its size — CALM3-22B-Chat, released July 2024, remains competitive with Meta's Llama-3-70B-Instruct on the Nejumi LLM Leaderboard despite being over a year and a half old with no CALM4 in sight. That's either a sign the architecture holds up or a sign CyberAgent hasn't prioritized a refresh; both readings are plausible.
The model comes from CyberAgent AI Lab, founded in 2016 and ranked 4th in Japan for AI research output — a notable result for a digital advertising company (CyberAgent posted ¥874 billion in FY2025 revenue) rather than a dedicated AI lab. The earlier OpenCALM (160M-6.8B, 2023) and CALM2 (7B, 1.3T tokens, 2023) releases show the progression, and all of it ships under Apache 2.0 with no licensing fees.
What you don't get is a hosted API. CyberAgent doesn't operate one, so running CALM3-22B means provisioning your own GPU — roughly 45GB of VRAM in BF16, less at 4-bit quantization. That's a real barrier for teams without in-house ML infrastructure, and a sharp contrast to SoftBank's SB Intuitions, which at least offers a paid API alongside its open weights.
Language coverage is Japanese and English only; there's nothing here for Southeast Asian markets. For a Japanese enterprise with GPU capacity and no interest in per-token API billing, CALM3 is still a legitimate free option in 2026 — just not a maintained-and-growing one.
Summary of public user & expert reviews, compiled by RECATOOLS.
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
