---
type: "post"
id: "preferred-networks-releases-open-weight-plamo-3-translate-56cdb4"
slug: "preferred-networks-releases-open-weight-plamo-3-translate-56cdb4"
title: "Preferred Networks releases open-weight PLaMo 3 Translate in 0.8B–31B sizes"
description: "Preferred Networks has released downloadable weights for PLaMo 3 Translate in four sizes, from 0.8B to 31B parameters, with formats for GPUs, CPUs and Apple silicon."
retrieval_nugget: "Preferred Networks has released downloadable weights for PLaMo 3 Translate in four sizes, from 0.8B to 31B parameters, with formats for GPUs, CPUs and Apple silicon. The practical proposition is a translation model that can be run locally, including on smaller devices; licensing conditions and the company’s own evaluation claims matter when deciding whether it fits a production workflow."
published_at: "2026-11-05"
updated_at: "2026-11-05"
record_date: "2026-11-05"
date_kind: "published_at"
topics: ["models","open-models","local-models"]
entities: ["Preferred Networks","PLaMo 3 Translate"]
related_patterns: []
novelty_label: "restatement"
what_changed: null
prior_refs: []
source_url: "https://www.preferred.jp/en/blog/tech/plamo-3-translate-weight-release"
source_urls: ["https://www.preferred.jp/en/blog/tech/plamo-3-translate-weight-release"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:preferred-networks-releases-open-weight-plamo-3-translate-56cdb4"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/preferred-networks-releases-open-weight-plamo-3-translate-56cdb4/","markdown":"https://newruntime.com/posts/preferred-networks-releases-open-weight-plamo-3-translate-56cdb4.md","json":"https://newruntime.com/posts/preferred-networks-releases-open-weight-plamo-3-translate-56cdb4.json"}
---

# Preferred Networks releases open-weight PLaMo 3 Translate in 0.8B–31B sizes

## Retrieval answer

Preferred Networks has released downloadable weights for PLaMo 3 Translate in four sizes, from 0.8B to 31B parameters, with formats for GPUs, CPUs and Apple silicon. The practical proposition is a translation model that can be run locally, including on smaller devices; licensing conditions and the company’s own evaluation claims matter when deciding whether it fits a production workflow.

## Summary
Preferred Networks has released downloadable weights for PLaMo 3 Translate in four sizes, from 0.8B to 31B parameters, with formats for GPUs, CPUs and Apple silicon. The practical proposition is a translation model that can be run locally, including on smaller devices; licensing conditions and the company’s own evaluation claims matter when deciding whether it fits a production workflow.

## What happened
Preferred Networks (PFN) says it has released all four sizes of PLaMo 3 Translate on Hugging Face: 0.8B, 2B, 8B and 31B. The company describes the family as translation-specialized and says it supports translation across 53 languages. Its listed capabilities include taking surrounding context and glossaries as input, and allowing users to choose between plain and polite Japanese. PFN says the smallest, 0.8B version outperforms existing translation-specialized LLMs, including models more than 30 times larger, on Japanese–English translation and translation between Japanese and other languages. Those are the publisher’s reported results; the supplied material does not include independent verification.

Each size is offered in BF16, several quantized variants, GGUF and MLX formats. PFN points builders to different runtime and hardware combinations: BF16 for compatible GPUs with Transformers or vLLM; FP8 for NVIDIA Ada Lovelace, Hopper and Blackwell GPUs with vLLM; W4A16 for Ampere or later; W4AFP8 for Hopper; and NVFP4 for Blackwell. GGUF is intended for CPUs and GPUs through llama.cpp, while MLX-4bit targets Apple silicon using mlx-lm. The post says its quantized formats retained nearly the same accuracy as BF16 in PFN’s benchmarks and directs readers to the model cards for details.

The weights are released under the PLaMo Community License, not under a license described in the post as unrestricted commercial use. PFN says personal study and hobbies are free, as is noncommercial use including academic use, without an income or revenue limit and without prior registration. Commercial use is also free if the user or its affiliates had no more than JPY 1 billion in income or revenue in the most recent fiscal year, but requires advance registration through PFN’s form. Users above that threshold must arrange a separate license agreement with PFN.

PFN presents on-device inference as one possible use of the smaller model. Its post shows an internal demonstration of PLaMo 3 Translate 0.8B translating directly on a smartphone, including fast translation in airplane mode without an internet connection. The company explicitly says it has not released an app like the one shown. The post links detailed usage instructions through each Hugging Face model card; it does not announce a hosted inference service or a consumer application.

PFN attributes the models’ development to data refinement and a multi-stage training process. It says it starts from a third-generation PLaMo base model and applies two stages of supervised fine-tuning followed by direct preference optimization. The overview describes training on source-text and translation pairs, using large volumes of data in the first SFT phase and more varied source text with more thoroughly refined synthetic translations in the second. The overview also names tournament-based data synthesis, back-translation, iterative self-distillation and distillation into smaller models as techniques used. The post points readers to a separate Japanese-language release post for comparisons with frontier LLM and translation APIs.

## What the archive adds
This is a new translation-model release, with a useful parallel in our earlier coverage of [Mistral’s open weights for Large 4](https://newruntime.com/posts/mistral-releases-open-weights-for-large-4-4cc443/). That earlier note concerns a different model and use case; it does not establish a direct connection to PLaMo. The new fact here is a multilingual translation-specific family, including compact local formats and explicit commercial-use conditions.

## Trend signal
No supplied SITE_PATTERNS thesis directly fits this translation-specific weight release. It is a concrete model distribution choice, but the evidence does not establish a broader change in agent economics, protocols, runtime security or another listed pattern.

## Connections to other events
The release sits alongside other recent open-weight releases in the archive, including [Mistral Large 4](https://newruntime.com/posts/mistral-releases-open-weights-for-large-4-4cc443/) and [Perplexity’s contextual embedding models](https://newruntime.com/posts/perplexity-releases-pplx-embed-v2-context-open-weight-d1aa40/). These are distinct products rather than evidence of a shared technical effort. PLaMo’s distinguishing emphasis in the supplied source is translation, multiple local deployment formats, and model sizes that include a 0.8B option.

## Why it matters
Teams handling multilingual text often face a trade-off between translation quality and where data can be processed. Downloadable weights and formats for llama.cpp and Apple silicon give builders an option to evaluate local inference rather than sending every request to a hosted translation API. A smaller model may also expand the hardware choices available for prototyping or offline workflows, though the smartphone example is only an internal demo and does not prove production readiness on a particular device.

The license is part of the engineering decision, not administrative fine print. Noncommercial use is described as unrestricted by revenue, but commercial use has a revenue threshold and registration requirement, with a separate agreement needed above it. Teams should check the license and their organization’s revenue status before deployment. Likewise, PFN’s accuracy and quantization statements are a starting point for evaluation, not a substitute for testing on the languages, terminology and quality criteria that matter to a product.

## Limits and uncertainty
The supplied source reports PFN’s own benchmark conclusions but does not provide independent replication here, nor enough benchmark detail to compare quality across every language and domain. It gives no specific latency, memory requirement, deployment cost or complete hardware minimums. The smartphone demonstration is internal, and PFN says it has not released a corresponding app. The post also does not establish that all 53 supported languages perform equally well, or that the quantized variants preserve quality on every task. Commercial permissions depend on the license terms and the organization’s circumstances; the summary above is not a substitute for reviewing the license.

## Skill to build now
Run a small, controlled translation evaluation before considering deployment. Select representative source passages from the languages and domains your product handles, include terminology that should be preserved, and compare output across the available model sizes and quantization formats. Record quality, latency and the hardware/runtime used, then check the relevant model card and license conditions. This directly tests the source’s practical claims in your environment without assuming that PFN’s benchmarks or demo predict your workload.

## Reader takeaway
PLaMo 3 Translate offers a newly downloadable set of translation-focused weights, including compact and local-runtime options. The practical opportunity is to test whether local inference suits a real translation workload; the key cautions are to validate quality independently and verify the applicable license before commercial use.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- topic hub: [Open Models](https://newruntime.com/topics/open-models.md): Explore the open-models topic hub.
- topic hub: [Local Models](https://newruntime.com/topics/local-models.md): Explore the local-models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)