PFN releases a 53-language translation model built with NICT

Background

Translation models are rarely trained from an empty start. A team usually takes a base model that has already read a large amount of multilingual text and adapts it to the job of turning one language into another. How many languages that base model covers sets the ceiling for the finished system, and how many sizes are released sets which machines can run it.

What was released

On 26 October 2026, PFN released the weights of PLaMo 3 Translate ^[1]. The model covers 53 languages, and four sizes shipped at once, from 0.8B to 31B ^[1]. Releasing the whole ladder in one go rather than starting with a single checkpoint means the same model family spans the small end, where a model has to fit on modest hardware, and the large end, where more memory is available.

Where the base model comes from

PLaMo 3 Translate rests on PLaMo 3 NICT, a base model that came out of joint research between PFN and NICT ^[1]. The split of labour is a common one: a base model supplies broad language competence, and a translation model is built on top of it so that it inherits that coverage instead of learning each language pair from scratch. Because the base model was produced in a collaboration, the translation release carries that lineage with it.

Formats and packaging

Alongside the weights, the release includes quantised builds in FP8, NVFP4, GGUF and Apple MLX ^[1]. Quantisation means storing a model's weights in fewer bits than the original training run used. The files get smaller and the memory needed to run them usually drops, at some cost in numerical fidelity. FP8 and NVFP4 are low-precision number formats of that kind. GGUF and Apple MLX are packaging and runtime formats aimed at local inference: GGUF is the format used by desktop tools for running models on a personal machine, while MLX targets Apple-silicon hardware. Offering several of these at once lets a user pick the build that matches their own setup rather than converting weights themselves. PFN also published a technical blog together with the weights ^[1].

Licence

The models are distributed under the PLaMo Community License rather than an OSI licence ^[1]. An OSI licence is one approved by the Open Source Initiative as meeting its open-source definition; a community licence instead sets terms written by the publisher. What the release states explicitly is that free use covers personal, hobby and academic work without registration ^[1].

Why it matters

Putting four sizes, four quantisation and runtime formats, and a technical write-up into a single release gives users a menu rather than a single artefact: the 0.8B end is small enough for local and resource-limited use, while the 31B end suits machines with more memory. The terms of reuse are the other half of the story. Because distribution runs through the PLaMo Community License instead of an OSI licence, anyone who wants to build on the weights has to read the publisher's own terms, with personal, hobby and academic use available without registration ^[1].