DEV.co
Open-Source LLM · ibm-granite

granite-4.0-h-tiny

Granite-4.0-H-Tiny is a 7B-parameter instruction-tuned LLM from IBM released in October 2025. It supports 12 languages and is designed for enterprise applications including summarization, text classification, RAG, code tasks, and tool-calling. Licensed under Apache 2.0, it is freely available and ungated on HuggingFace.

Source: HuggingFace — huggingface.co/ibm-granite/granite-4.0-h-tiny
6.9B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
268.6k
Downloads (30d)

Key facts

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

FieldValue
Developeribm-granite
Parameters6.9B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads268.6k
Likes205
Last updated2025-11-03
Sourceibm-granite/granite-4.0-h-tiny

What granite-4.0-h-tiny is

A 6.9B-parameter dense transformer model finetuned from Granite-4.0-H-Tiny-Base using supervised finetuning, RLHF alignment, and model merging. Trained on permissively licensed open-source and synthetic datasets. Features structured chat format with default system prompt (Oct 2025 update), OpenAI-compatible tool-calling, and Fill-In-the-Middle code completion. Context length not specified. Safetensors weights compatible with transformers, vLLM, and Azure deployment.

Quickstart

Run granite-4.0-h-tiny 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-tiny")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

Designed for business applications with improved instruction following and alignment. Default system prompt (Oct 2025) guides professional, accurate responses suitable for customer-facing deployments.

Tool-calling workflows

Native support for OpenAI function-calling schema enables seamless integration with external APIs and microservices without custom parsing.

Code-related tasks

Demonstrated capabilities in code completion, analysis, and generation with Fill-In-the-Middle support for on-device inference.

Running & fine-tuning it

Estimated ~14–16 GB VRAM for full precision (bfloat16/float32) inference on A100/H100 or similar. 7–8 GB with int8 quantization. Exact memory footprint depends on batch size, sequence length, and serving framework. CPU-only inference possible but slow (transformers with device_map='cpu'). GPU acceleration recommended for production.

Model card does not specify LoRA/QLoRA compatibility or instruction-tuning datasets used. Apache 2.0 license permits modification and derivative works. QLoRA is compatible with transformers library. For domain-specific finetuning, expect moderate compute (single GPU feasible for small LoRA ranks). No proprietary training code documented; use standard transformers + PEFT tooling.

When to avoid it — and what to weigh

  • Very long-context requirements — Context length is not specified in the model card. If you need guaranteed long-context capabilities (32k+), verify against your use case or test empirically.
  • Cutting-edge reasoning benchmarks — BBH (3-shot CoT) scores 66.34 for H-Tiny variant. If your application requires state-of-the-art chain-of-thought reasoning, consider larger models or specialized reasoning architectures.
  • Non-English or unsupported languages — Supports 12 specific languages (English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, Chinese). Not suitable for other languages without finetuning.
  • Maximum performance in constrained environments — 7B parameters will require moderate GPU VRAM. If deploying on edge devices or CPUs only, quantization or smaller models (MoE variants) may be necessary.

License & commercial use

Apache 2.0 license. Permissive open-source license allowing commercial use, modification, and redistribution with attribution and no warranty. No restrictions on gating or enterprise deployment. No proprietary dependencies noted.

Apache 2.0 is a permissive OSI-approved license. Commercial use, modification, and redistribution are explicitly permitted. No license-based restrictions on building commercial products, SaaS, or closed-source derivatives. Attribution required. Verify compliance with any training data licenses (model card states 'open source instruction datasets with permissive license and internally collected synthetic datasets')—ensure your use of synthetic IBM data aligns with your commercial model if reusing internally.

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

No security audit, adversarial robustness testing, or known vulnerability disclosure documented. Model includes default system prompt (Oct 2025 update) intended to guide 'professional, accurate, and safe' responses—a soft control, not a guarantee. Typical LLM risks apply: prompt injection, hallucination, potential for misuse if deployed without input/output filtering. Recommend running evals on your specific use case and applying application-level safeguards (guardrails, moderation, rate limiting). No mention of red-teaming results or bias mitigation strategies.

Alternatives to consider

Mistral 7B Instruct

Similar 7B parameter density, strong instruction-following, permissive license (Apache 2.0), lower barrier to adoption. May offer comparable or better reasoning on some benchmarks (BBH). No tool-calling native but compatible with vLLM structured output plugins.

Granite 4.0 H Small MoE

Same family, sparse Mixture-of-Experts variant. Larger model size (78.44 MMLU vs 68.65 for Tiny) with lower active parameter count. Better for reasoning-heavy workloads if compute budget allows. Requires MoE-aware serving (vLLM, TGI).

LLaMA 3.2 (1B/3B/8B)

Meta's dense models, strong community ecosystem, permissive license (Llama 2/3.x license allows commercial use), available in multiple sizes. 1B/3B suitable for edge; 8B comparable to Granite-Tiny. Better documented; no tool-calling native. Consider if you prioritize inference speed over instruction-tuning quality.

Software development agency

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

Download from HuggingFace, benchmark against your workload, and evaluate tool-calling and RAG capabilities. Check context length and finalize integration with your serving framework (vLLM, TGI, Ollama). For custom deployment or domain-specific finetuning, reach out to explore /ai/custom-llm-apps or /ai/private-llm options.

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.

granite-4.0-h-tiny FAQ

Can I use Granite-4.0-H-Tiny commercially without restrictions?
Yes. Apache 2.0 is permissive and explicitly allows commercial use, modification, and redistribution. Attribution required. No proprietary restrictions documented. However, verify any third-party dependencies or training data licensing in your deployment chain.
What GPU do I need to run this model?
Estimate 14–16 GB VRAM for bfloat16 inference (A100, H100, RTX 4090, or A10G). Smaller GPUs (V100, T4, RTX 3090) support inference with int8 quantization (~8 GB). CPU-only inference is feasible but slow. Exact memory depends on batch size and sequence length—test in your environment.
Does it support long-context (>8K tokens)?
Context length is not stated in the model card. Assume default transformer context (~2K–4K) unless verified empirically or in upstream documentation. Contact IBM Granite team or test on your dataset before committing to long-context workloads.
How does it compare to Granite 4.0 H Small MoE in terms of performance?
H Small MoE is larger and outperforms Tiny on benchmarks (e.g., MMLU 78.44 vs 68.65, MMLU-Pro 55.47 vs 44.94). Tiny is faster and cheaper to run; choose Tiny for latency/cost-sensitive applications, Small MoE for accuracy. Both support tool-calling and multilingual tasks.

Work with a software development agency

DEV.co helps companies turn open-source tools like granite-4.0-h-tiny 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 Tiny?

Download from HuggingFace, benchmark against your workload, and evaluate tool-calling and RAG capabilities. Check context length and finalize integration with your serving framework (vLLM, TGI, Ollama). For custom deployment or domain-specific finetuning, reach out to explore /ai/custom-llm-apps or /ai/private-llm options.