DEV.co
Open-Source LLM · ibm-granite

granite-4.0-h-small

Granite-4.0-H-Small is a 32-billion-parameter instruction-tuned language model from IBM, released October 2025. It supports 12 languages, handles general NLP tasks (summarization, classification, QA, RAG, code), and includes tool-calling for function integration. Apache 2.0 licensed, ungated, and designed for enterprise applications. No context window length disclosed.

Source: HuggingFace — huggingface.co/ibm-granite/granite-4.0-h-small
32.2B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
330.1k
Downloads (30d)

Key facts

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

FieldValue
Developeribm-granite
Parameters32.2B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads330.1k
Likes308
Last updated2025-11-03
Sourceibm-granite/granite-4.0-h-small

What granite-4.0-h-small is

32.2B parameter dense transformer, instruction-tuned via supervised finetuning, reinforcement learning alignment, and model merging on permissive open-source + synthetic datasets. Supports structured chat format with system prompts, tool-calling via OpenAI function schema, and Fill-In-the-Middle code completion. Published 2025-11-03 on Hugging Face. Context length unknown. Trained on English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, Chinese.

Quickstart

Run granite-4.0-h-small locally

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

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="ibm-granite/granite-4.0-h-small")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

Enterprise AI Assistants

Default system prompt guides professional, accurate, safe responses. Tool-calling enables integration with business APIs and functions. Suitable for customer service, internal knowledge retrieval, and domain-specific automation.

Multilingual Document Processing

Native support for 12 languages enables summarization, classification, and information extraction across diverse document corpora without per-language fine-tuning.

RAG and Knowledge-Grounded Applications

Model card explicitly lists RAG capabilities and improved instruction following, making it suitable for context-aware Q&A systems and retrieval pipelines.

Running & fine-tuning it

ESTIMATE: 32B dense model in bfloat16 requires ~64 GB GPU VRAM (e.g., A100 80GB, H100, or dual A100 40GB). FP8 quantization reduces to ~32 GB. CPU-only inference possible but slow. Exact requirements depend on batch size and serving framework; test on target hardware.

Model card confirms open-source instruction datasets and synthetic data were used. Fine-tuning via LoRA/QLoRA is feasible and documented in the transformers library. Model card does not detail LoRA rank recommendations or fine-tuning hyperparameters; refer to IBM's GitHub repository and huggingface training docs. Permissive license (Apache 2.0) allows commercial fine-tuning.

When to avoid it — and what to weigh

  • Extremely Long Context Requirements — Context window length is not disclosed. If you require very long sequences (>8K tokens), verify context length or test empirically before committing.
  • Real-Time Inference with Tight Latency Budget — 32B model requires significant compute. For sub-100ms latency, consider smaller models or quantized variants; evaluate on your hardware first.
  • Specialized Domains Without Fine-Tuning Validation — Model is general-purpose instruction-tuned. Domain-specific tasks (medical, legal, scientific) may require fine-tuning and validation on representative data.
  • License-Restricted Deployment Scenarios — While Apache 2.0 is permissive, verify compliance with internal IP policies, export controls, and any downstream commercial licensing requirements in your jurisdiction.

License & commercial use

Apache License 2.0. Permissive open-source license allowing use, modification, and distribution. No field-of-use or commercial restrictions.

Apache 2.0 is an OSI-approved permissive license. Commercial use, modification, and redistribution are allowed provided the license and copyright notice are retained. No patent protection guarantees. Verify internal IP and compliance policies, and confirm no downstream restrictions from your deployment platform (e.g., cloud provider terms).

DEV.co evaluation signals

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

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

Model card updated to include a default system prompt guiding 'professional, accurate, and safe' responses. No explicit red-teaming, adversarial robustness, or data leakage evaluation disclosed. Model is deployed via Hugging Face (gated=false) and Azure endpoints; verify endpoint security posture with your platform. Input sanitization and output validation are your responsibility. No formal security audit details published.

Alternatives to consider

Llama 2 / Llama 3 (Meta)

Similar parameter range, permissive license (though non-OSI for Llama 2 in some regions). Llama 3 offers comparable or better performance on many benchmarks and larger community. Trade-off: less recent IBM enterprise focus.

Mistral 7B/8x7B (Mistral AI)

Smaller, faster inference, Apache 2.0 licensed. Suitable if you can tolerate lower capability for cost/speed gains. MoE variant available for efficiency.

Comparable 32B variant, multilingual (Chinese-optimized), Apache 2.0 licensed. Good alternative if multilingual performance or cost in specific regions is a priority.

Software development agency

Ship granite-4.0-h-small with senior software developers

Start with a self-hosted or cloud deployment to evaluate tool-calling and multilingual capabilities. Our team can help you plan GPU requirements, fine-tuning strategy, and production architecture.

Talk to DEV.co

Related open-source tools

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

granite-4.0-h-small FAQ

Can I use Granite-4.0-H-Small commercially?
Yes. Apache 2.0 is a permissive OSI-approved license allowing commercial use. You must retain the license notice and copyright attribution. Check your organization's IP policies and any restrictions imposed by your deployment platform or customers.
What is the exact context window length?
Not disclosed in the model card. Test empirically on your target hardware, or check the GitHub repository (ibm-granite/granite-4.0-language-models) for technical specifications. Typical dense 32B models support 4K–8K tokens; confirm before deploying.
How much GPU memory do I need to run this model?
Estimate: ~64 GB for bfloat16 (A100 80GB), ~32 GB for FP8 quantization. Requirements vary by batch size, sequence length, and serving framework. Test on your hardware; consider quantization (GPTQ, AWQ) or model parallelism for cost reduction.
Can I fine-tune this model for my domain?
Yes. Use LoRA/QLoRA or full fine-tuning. Transformers library supports both. Model card confirms the base was trained on open-source data, so fine-tuning is permissible. Hyperparameters and best practices are not detailed in the card; refer to IBM's GitHub repo and community guides.

Software developers & web developers for hire

DEV.co helps companies turn open-source tools like granite-4.0-h-small into production software. Our software development services cover the full lifecycle — architecture, web development, integration, and maintenance — delivered by software developers and web developers who ship. Engage our software development agency to implement or customize it for your open-source llms stack.

Ready to Deploy Granite-4.0-H-Small?

Start with a self-hosted or cloud deployment to evaluate tool-calling and multilingual capabilities. Our team can help you plan GPU requirements, fine-tuning strategy, and production architecture.