Skip to main content
Rase AI Intelligence Newsletter

Models & assistants Model or AI service

Swallow adapts Llama 3.3 70B for Japanese

The Swallow project released Japanese-adapted Llama 3.3 70B weights as base and instruction-tuned checkpoints. It documents the v0.4 release and its training approach. [1][2]

Model access
Downloadable Llama 3.3 Swallow 70B v0.4 base and instruction-tuned checkpoints. [1][2]
License
Llama 3.3 Community License. The project also identifies Gemma usage restrictions. [1][2]

What the model does

The project’s March 10, 2025 release page lists both v0.4 checkpoints. The team describes additional training and evaluations of Japanese-language understanding, generation and dialogue. [1][2]

Why it matters

A Japanese-adapted version gives researchers a way to compare a model with its original Llama lineage. Publishing both checkpoints lets them evaluate the base and the effects of instruction tuning. [1][2]

Who can use it

Developers building Japanese text applications can evaluate the downloadable instruction version. Researchers who need a starting point for further training can use the base version. [1][2]

Key dates

  1. Swallow documents the 70B v0.4 public release [1]

Ways to use it

Give the model Japanese tasks from the intended domain and have fluent reviewers check the output. Record the distinction between responses from the base and instruction versions. [1][2]

Model limitations

The release inherits the Meta Llama 3.3 license, and the project directs users to relevant Gemma restrictions as well. Results in the project’s benchmark do not cover every Japanese use case. [1][2]

Sources and verification

Catalog evidence checked How we verify records

  1. Primary source Swallow research team’s dated 70B v0.4 release and license notes
  2. Primary source tokyotech-llm/Llama-3.3-Swallow-70B-Instruct-v0.4
  3. Primary source Hugging Face registry metadata for tokyotech-llm/Llama-3.3-Swallow-70B-Instruct-v0.4

Related profiles

Models & assistants

GPT-6.2 Astra and Sol

OpenAI

Astra targets demanding projects. Sol targets everyday tasks. Both support text and image input in the API.

High confidenceEvidence checked

Models & assistants

MAI Code 2 and Code 2 Flash

Microsoft AI

Code 2 targets complex software work. Flash targets frequent edits, explanations and routine fixes.

High confidenceEvidence checked

Models & assistants

MAI Realtime 1

Microsoft AI

Users can try a MAI Labs demo. Developers can register for a limited Foundry preview.

High confidenceEvidence checked