Third Intelligence Releases Haas-ja-v1.0 Japanese Speech Recognition Model
Third Intelligence released Haas-ja-v1.0, a Japanese-specialized speech recognition model with 615.8 million parameters, as open source. The company says it has the top accuracy among the open-weight models it evaluated, scoring above the commercial model Gemini 3.5 transcribe.
1 source 2h ago
Reporting from 1 source: GIGAZINE.
Why it matters
A 615.8 million parameter Japanese ASR model that outscores a commercial model on the company's own accuracy chart and runs at low cost puts a lightweight open option in front of teams that would otherwise pay API fees.
Key Facts
- Third Intelligence released Haas-ja-v1.0, a Japanese-specialized speech recognition model with 615.8 million parameters.
- The company says Haas-ja-v1.0 scored highest among the open-weight models it evaluated, above the commercial model Gemini 3.5 transcribe.
- A free demo transcribed 4 minutes 19 seconds of audio in 0.32 seconds after upload, with some errors requiring manual correction.
- Haas-ja-v1.0 is available on Hugging Face under Apache License 2.0, with inference code on GitHub.
A free demo transcribed 4 minutes 19 seconds of audio in 0.32 seconds after upload, with some errors that needed manual correction. The model is on Hugging Face under Apache License 2.0, with inference code on GitHub.
Third Intelligence published the model weights on Hugging Face and the inference code on GitHub under Apache License 2.0. The company's cost chart compares commercial models by API fees and open models by the fee for inference on an H200 in the Japan region, and places Haas-ja-v1.0 as the lowest-cost model at equivalent accuracy.
In a hands-on test, a 4 minute 19 second clip finished transcribing 0.32 seconds after upload. The output still needed manual fixes: a name rendered as a similar-sounding wrong name, and a title likewise. Speed and accuracy claims come from Third Intelligence's own evaluation, not an independent benchmark.
Written with AI by Yomimono from the 1 cited source below, including Japanese-language reporting where cited. See our editorial policy.
Add Yomimono as a preferred source on Google