An MLX version of PLaMo 3 Translate 0.8B, the smallest model in PFN's translation-focused LLM family, now has weights of only about 423MB. It is a mixed-quantization model called "mixed_3_4", published on Hugging Face by an individual converter.
Unless noted otherwise, every size, speed, and memory figure in this post is the converter's own measurement (Mac mini M2 Pro, 32GB, mlx-lm 0.32.2). The main measurements for the mixed_3_4 version are:
- Weights: about 423MB
- Decode speed: about 352 tok/s
- Peak memory (MLX-managed): about 565MB
This follows my previous article, "Using PLaMo 3 Translate on a Mac". This time I cover how the mixed-quantization version is built, the measurement results, what to watch for when choosing it, and the license terms.
PLaMo 3 Translate and the available model versions
PLaMo 3 Translate is PFN's translation-focused LLM. It comes in four sizes (0.8B / 2B / 8B / 31B) and supports 53 languages. The weights were published on Hugging Face on 2026-11-05.
PFN's official announcement on X is below. It serves as a primary source for the release.
https://x.com/preferred_jp/status/2118134261912092801
Here are three sizes to use as reference points:
- Official MLX-4bit version (0.8B): about 0.98GB of weights (from the comparison in the 4bit-compact model card)
- The earlier uniform 4-bit version (4bit-compact): 501,002,761 bytes of weights, about 501MB
- This mixed_3_4 version: about 423MB of weights
The size reported for the uniform 4-bit version varies by source. The model card says about 501MB (about 505MB for the whole repository), but the converter's X post says "480MB". I use the model card's figure in this article.
The converter's post that says "480MB" is here:
https://x.com/sizumita/status/2119360329264271441
As a check on conversion fidelity, the model card for the uniform 4-bit version reports a per-tensor cosine similarity of about 0.996 between the dequantized weights and the BF16 original. The card says this is within the range expected for 4-bit affine quantization.
The model card for the uniform 4-bit version is at the page below. It describes the uniform 4-bit configuration and provides reference quality scores.
https://huggingface.co/sizumita/plamo-3-translate-0.8b-mlx-4bit-compact
The mixed_3_4 version was converted from the BF16 original, and its weights are smaller than those of the earlier uniform 4-bit version.
Using different bit widths for different parts of the model
The mixed_3_4 version is based on PFN's pfnet/plamo-3-translate-0.8b (BF16, unmodified). The conversion used mlx-lm 0.32.2's mixed_3_4 recipe (mlx_lm convert --quant-predicate mixed_3_4).
According to the model card, the bits are allocated as follows:
- 4-bit: lm_head, plus the first, the last, and every third down_proj across the layers
- 3-bit: everything else
- Quantization method: affine, group size 64
- Average: 3.804 bits/weight
The uniform all-4-bit version comes to 4.501 bits/weight. The mixed_3_4 version keeps part of the model at 4-bit and drops the rest to 3-bit, which brings the average down to 3.804 bits/weight. The model card also says this allocation follows the same strategy as llama.cpp's Q4_K_M.
The vocabulary is not trimmed; all 107,520 tokens remain.
You can read the full model card at the page below, including the allocation details and the original wording of the measurement conditions.
https://huggingface.co/sizumita/plamo-3-translate-0.8b-mlx-mixed34
The quality metric is slightly lower on average
With a smaller model, you naturally wonder what happened to translation quality. The baseline is the uniform 4-bit version. The converter compared the two versions using chrF++ under the same conditions (greedy decoding). These are the converter's own measurements.
- FLORES-200 devtest (en→ja / zh→ja / ko→ja, first 100 sentences in each direction): average 31.37 (4-bit version: 31.56)
- NTREX-128 (four directions including ja→en, first 200 sentences in each direction): average 29.57 (4-bit version: 29.90)
The model card states that the NTREX-128 news sentences were used neither for the conversion nor for choosing the method.
The mixed_3_4 version therefore scores slightly lower on average on both datasets. The converter's post describes the performance as "almost unchanged", but the measured chrF++ averages are slightly lower on both datasets.
The converter's post announcing the mixed_3_4 version (2026-11-09) is here:
https://x.com/sizumita/status/2119694681982541828
For reference, the model card says the uniform 4-bit version scores an average chrF++ of 31.56 on FLORES-200, roughly on par with the BF16 original's reference value of 31.20. The card itself notes, however, that the comparison includes implementation differences, so it should be read as "no degradation".
Even when the metrics are close, the translations themselves change
One point is easy to miss. Because the precision allocation is different, the mixed_3_4 version produces translations that differ from the 4-bit version's.
Even when the average scores are close, the two models do not translate the same sentence into the same text. On NTREX-128, the converter counted the sentences whose translation matches the 4-bit version byte for byte, and the match is far from complete (the per-direction counts are in the paid document below).
- If you need identical translations: the model card points you to the 4-bit version
- If you prioritize speed and memory: the mixed_3_4 version is an option
Minimal steps to try it on a Mac
Following the model card, install mlx-lm with pip install mlx-lm, load the model with mlx_lm.load, and translate with generate.
The tokenizer uses custom code bundled in the repository, so loading requires specifying trust_remote_code. The supported translation directions are the same as the original model's, including translation from multiple languages into Japanese and from Japanese into multiple languages.
License terms to check before you use it
The mixed_3_4 version is a derivative model under the PLaMo Community License. The model card states explicitly that it is an unofficial conversion and does not imply PFN's endorsement or affiliation.
The terms by type of use are:
- Non-commercial use: free
- Commercial use: prior registration with PFN is required. Limited to users with annual revenue (sales or income in the most recent fiscal year) of 1 billion yen or less; above that, a separate license is required
- Redistribution: you must include the license text, display "Built with PLaMo", and keep "PLaMo" in the model name (Article 4)
The original license text is at the page below. Consult it for the full terms governing commercial use and redistribution.
https://plamo.preferredai.jp/info/plamo-community-license-en
If you are considering commercial use or redistribution, read this original text before you start.
Compare them on the same sentences first
The measurements in this post come from a Mac mini M2 Pro (32GB) with mlx-lm 0.32.2.
If you try it yourself, load both the mixed_3_4 version and the 4-bit version, translate the same handful of input sentences with each, and compare. Seeing the speed and the translation differences side by side makes it easier to judge which one best suits your needs.
Get the complete measurement data and runnable code ($5)
This post covered the headline figures, the bit allocation, the average quality scores, and the license terms. The paid document adds the full side-by-side measurements, every per-direction chrF++ score, the byte-identical match counts, and the complete runnable Python example.
What's inside:
- Full size, speed, memory, and load-time comparison against the uniform 4-bit version, plus the whole-process RSS note
- All per-direction chrF++ scores for FLORES-200 and NTREX-128, with the 4-bit version's scores alongside
- The per-direction counts of translations that match the 4-bit version byte for byte
- The complete runnable Python code from the model card, with the
trust_remote_codenote
Originally published in Japanese on note. This English edition was translated and edited with AI assistance; all commands and measurements are unchanged from the original.
Top comments (0)