DEV.co
Open-Source LLM · llm-jp

llm-jp-3-150m

llm-jp-3-150m is a 150M-parameter transformer language model developed by Japan's National Institute of Informatics, pre-trained on 2.1T tokens of Japanese, English, code, and multilingual text. It supports both Japanese and English generation with a 4096-token context window. The model is open-source under Apache 2.0, ungated, and suitable for inference and fine-tuning on resource-constrained hardware.

Source: HuggingFace — huggingface.co/llm-jp/llm-jp-3-150m
152M
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
49.4k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerllm-jp
Parameters152M
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads49.4k
Likes8
Last updated2025-02-04
Sourcellm-jp/llm-jp-3-150m

What llm-jp-3-150m is

Transformer-based causal LM with 12 layers, 512 hidden size, 8 attention heads, and 4096 context length. Pre-trained on mixed-language corpus (Japanese: Wikipedia, Common Crawl, WARP, Kaken; English: Dolma, Wikipedia, Reddit; Code: The Stack; totaling 2.1T tokens). Uses Unigram byte-fallback tokenizer derived from llm-jp-tokenizer v3.0. Distributed in Hugging Face Transformers format with safetensors. Requires torch>=2.3.0, transformers>=4.40.1, and optional flash-attn>=2.5.8 for acceleration.

Quickstart

Run llm-jp-3-150m locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="llm-jp/llm-jp-3-150m")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Japanese NLP Applications on Edge/Local Infra

Ideal for organizations needing Japanese language generation without cloud dependencies. 150M parameters enable deployment on modest GPU or CPU setups, suitable for customer support chatbots, content generation, and document summarization in Japanese contexts.

Fine-tuning for Domain-Specific Tasks

Small parameter count (150M) makes LoRA and QLoRA fine-tuning feasible on single consumer GPUs. Suitable for adapting to domain-specific Japanese corpora (legal, medical, technical documentation) with minimal computational overhead.

Private/On-Premise Deployment

Apache 2.0 license and compact size enable self-hosted deployment without licensing friction. Applicable to enterprises requiring data sovereignty or air-gapped environments processing Japanese-English mixed workloads.

Running & fine-tuning it

ESTIMATE: bfloat16 precision ~300–600 MB model weights. Inference on single GPU (e.g., RTX 3090, V100, L4) with 8GB+ VRAM is feasible. CPU inference possible but slower. Fine-tuning with LoRA on 16GB consumer GPU practical; full fine-tuning requires 32GB+ or gradient accumulation. Flash-attn optional but recommended for throughput.

150M parameters make LoRA and QLoRA highly feasible on single consumer GPUs (16GB+). No official fine-tuning instructions provided in card. Recommend Hugging Face PEFT library or TRL for instruction tuning. Full fine-tuning (requires 32GB+ VRAM or multi-GPU setup). Tokenizer is fixed; no indication of instruction-tuning or chat-format variants in base model—community or custom adaptation likely required for dialogue use cases.

When to avoid it — and what to weigh

  • Requirement for Production-Grade Safety/Alignment — Model card explicitly states 'early stages of research' and 'not tuned to ensure outputs align with human intent and safety considerations.' Unsuitable for customer-facing applications without additional safety guardrails or RLHF fine-tuning.
  • Long-Context or Reasoning-Heavy Tasks — 4096-token context window and 150M parameters limit performance on complex reasoning, long document analysis, or multi-hop QA. Larger variants (3.7B, 7.2B) in the LLM-jp-3 family may be needed.
  • High-Volume Inference at Scale — No stated benchmarks for throughput, latency, or multi-request serving. Unknown performance at scale compared to distilled models from larger base LLMs. Requires empirical testing before committing to production inference clusters.
  • Non-Japanese/English Language Focus — Pre-training corpus is heavily weighted toward Japanese (1T+ tokens) and English. Chinese (0.8B tokens) and Korean (0.3B tokens) support is minimal; expect degraded performance for these languages.

License & commercial use

Apache License 2.0 (OSI-compliant). Permissive open-source license allowing modification, redistribution, and private use. No derivative attribution required in binary form, but license text must be included. Commercial use is permitted under Apache 2.0 terms.

Apache 2.0 is an OSI-approved permissive license that explicitly permits commercial use, modification, and redistribution. No license restrictions on commercial applications. However, model card disclaimers note 'early stages of research' and lack of safety tuning; independent risk assessment and additional safety measures (filtering, monitoring, oversight) are strongly recommended before deploying in production customer-facing systems.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityLow
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Model trained on web-scraped data (Common Crawl, Wikipedia, Reddit, Dolly) and code; inherent risk of memorized sensitive content (emails, credentials, PII). No stated deduplication or PII filtering applied to pre-training corpus. Tokenizer uses byte-fallback; ensure input sanitization if deployed in untrusted environments. No cryptographic integrity checks mentioned. Recommend scanning model card and repo for adversarial examples or prompt-injection vectors if used in chat/interactive contexts.

Alternatives to consider

Mistral-7B

Larger (7B params), English-optimized, more mature community support and serving infrastructure. Better for high-accuracy reasoning; overkill if Japanese is primary requirement and inference latency is critical.

Rinna Japanese GPT-2 (1.2B)

Japanese-specific, comparable size. Trade-off: smaller pre-training, older architecture (GPT-2). Simpler to deploy but lower quality than llm-jp-3-150m Transformer.

LLM-jp-3-440M or 980M

Variants within same family. Larger parameter counts (440M, 980M) offer better quality at cost of higher VRAM/latency. Choose if 150M quality insufficient and deployment budget allows.

Software development agency

Ship llm-jp-3-150m with senior software developers

llm-jp-3-150m offers compact, permissive open-source inference and fine-tuning. Evaluate on your hardware, test with your data, and deploy on-premise or edge infrastructure without licensing constraints. Check the model card for evaluation benchmarks and start with the provided Python snippet.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

llm-jp-3-150m FAQ

Can I use this model commercially?
Yes. Apache 2.0 permits commercial use without royalties or license fees. However, the model is in early research stages and lacks safety tuning. Independent risk assessment, content filtering, and monitoring are strongly recommended before deploying in production customer-facing systems.
What GPU VRAM do I need?
ESTIMATE: 300–600 MB for model weights in bfloat16. A consumer GPU with 8GB VRAM (RTX 3060, L4, V100) is sufficient for batch inference. Fine-tuning with LoRA requires 16GB+. Full fine-tuning or large batch serving requires 32GB+. CPU inference is possible but slow.
Is there a chat or instruction-tuned version?
Not in the base model. The llm-jp-3 family includes fine-tuned variants (see model card references). The 150M base is pre-trained on raw text; instruction tuning requires community adaptation or custom fine-tuning using PEFT (LoRA) or full fine-tuning.
How does it compare in quality to larger models like GPT-3.5?
Not directly comparable. 150M parameters are orders of magnitude smaller than GPT-3.5 (175B+). Quality is lower for complex reasoning and English tasks. Strength is Japanese language support at low latency/cost. Evaluation results available in the referenced blog (2025-02-05).

Work with a software development agency

From first prototype to production, DEV.co delivers software development services around tools like llm-jp-3-150m. Our software development agency staffs experienced software developers and web developers for custom software development, web development, integrations, and ongoing support across open-source llms and beyond.

Ready to Deploy Japanese LLM Locally?

llm-jp-3-150m offers compact, permissive open-source inference and fine-tuning. Evaluate on your hardware, test with your data, and deploy on-premise or edge infrastructure without licensing constraints. Check the model card for evaluation benchmarks and start with the provided Python snippet.