Open LLMs/ibm-granite

Open-Weight LLM · Private & Custom AI

granite-4.0-h-tiny

A 7B instruction-tuned model built for enterprise ops automation, tool-calling, and private deployment in mid-market workflows.

Granite-4.0-H-Tiny is a 7B parameter instruct model from IBM, optimized for instruction following, function calling, and multilingual tasks across 12 languages. For ops teams, it's sized to run on modest hardware while handling RAG, classification, summarization, and agentic workflows—all controllable in your own environment.

6.9B
Parameters
apache-2.0
License (OSI/permissive)
Unknown
Context
268.6k
Downloads

Model facts

Developeribm-granite
Parameters6.9B
Context windowUnknown
Licenseapache-2.0 — OSI/permissive
Tasktext-generation
GatedNo
Downloads268.6k
Likes205
Updated2025-11-03
Sourceibm-granite/granite-4.0-h-tiny

Private deployment

Run granite-4.0-h-tiny in your own environment

Self-hosting is straightforward: load via `transformers` + `torch` on CUDA or CPU. At 7B params, estimate 14–28 GB VRAM depending on precision (fp16 vs int8 quantization). Companies deploy to on-prem Kubernetes, edge servers, or air-gapped data centers to keep operational data (customer interactions, internal docs, financial records) entirely within their network.

Operational AI use cases

01

Customer Support Automation & Routing

Route inbound tickets, summarize support threads, and extract intent without shipping customer data to external APIs. Use tool-calling to integrate with Zendesk, Jira, or internal ticketing systems. Model decides which queue, suggests responses, escalates—all privately.

02

Document Classification & Compliance Extraction

Process contracts, invoices, and regulatory filings to extract entities, classify risk level, and flag exceptions. Structured chat format + function calling enables direct API calls to archival or workflow systems, keeping sensitive docs in your infrastructure.

03

Internal Knowledge Chatbot & Question-Answering

Build a private chatbot over company wikis, procedure docs, and institutional knowledge. RAG capabilities mean the model retrieves and synthesizes your own data without exposing it to third parties. Ideal for onboarding, HR Q&A, and reducing support load.

Custom AI

As a base for custom AI

Strong foundation for fine-tuning on domain-specific tasks: train on your own labeled support tickets, contract language, or operational procedures. Apache 2.0 license permits commercial reuse. The model merging techniques and instruction-following focus make it suitable as a backbone for vertical-specific AI products (e.g., financial ops tools, legal tech).

In the operating system

Where it fits

Sits at the **execution layer** of an AI operating system: receives task requests from an agent/workflow orchestrator, executes them via built-in tool-calling, and returns structured outputs. Pair with a retrieval service (vector DB) for RAG, orchestration layer (LangChain, LlamaIndex) for multi-step workflows, and your internal APIs for data access.

Data control & security

Self-hosting ensures operational data (customer queries, internal records, financial data) never leaves your network. No external API logging or model training on your inputs. Sensitive workflows stay inside your firewall. Note: data governance, model monitoring, and compliance testing are your responsibility; the model itself carries no built-in audit trail or encryption.

Hardware footprint

**Estimate:** fp32 (dense) ~28 GB VRAM; fp16 ~14 GB; int8 quantization ~7–9 GB. Suitable for a mid-range GPU (RTX 4090, A100 40GB) or two consumer GPUs. CPU inference possible but slow (~2–5 tokens/sec). Context length unknown—requires testing; assume similar to base Granite (likely 4K–32K, verify on HF model card or GitHub repo).

Integration

Chat template + tool-calling schema are OpenAI-compatible, easing integration into existing LLM frameworks. Supports structured outputs via JSON schema in function calls. Can wire to REST/GraphQL APIs, SQL databases, internal webhooks. Tokenizer is HuggingFace standard; use `safetensors` format for safe loading. Monitor latency: 7B on single GPU typical inference time ~100–300ms depending on context length.

When it's not the right fit

  • Reasoning tasks requiring very long chains of thought—7B is constrained; H-Small or larger variant recommended.
  • Highly specialized domains (genomics, quantum physics) without heavy fine-tuning; general instruction data only.
  • Real-time, sub-100ms latency required at scale—7B + self-hosting may bottleneck; consider quantization or distillation.
  • Context length unstated; if you need >32K token windows, verify on GitHub or conduct load testing before committing.

Alternatives to consider

Llama 2 / 3 (7B / 8B)

Similar size & speed. Llama 3 has stronger instruction-following; no built-in tool-calling in base model (requires additional training). Broader ecosystem; more community fine-tunes.

Mistral 7B / Mistral Small

Lighter, faster inference. Weaker tool-calling support natively. Strong RAG baseline; better for latency-sensitive ops but less enterprise-focused.

Granite-4.0-H-Small MoE

IBM's own 27B MoE variant; better reasoning & MMLU (~78 vs 69), but larger footprint and sparse compute overhead. Use if your hardware supports it and reasoning is critical.

FAQ

Can we fine-tune Granite-4.0-H-Tiny on our own customer data?

Yes. Apache 2.0 license permits fine-tuning. Use standard HuggingFace trainer or PEFT (LoRA) to adapt to your domain. Synthesized data in the pre-training was permissively licensed, so derived models remain freely usable.

What's the license for commercial products built on this model?

Apache 2.0 is permissive: you can build commercial products, modify, and distribute—with attribution and a copy of the license. No restrictions on enterprise use. Verify liability clauses in your legal review.

How do we run this privately without touching external APIs?

Download the model from HuggingFace, host on your own GPU / CPU servers (on-prem, private cloud, or air-gapped). Load via `transformers` library. Wire function-calling outputs to internal APIs only. No telemetry or external dependencies required.

What's the model's context length?

Unknown from the HuggingFace card. Check the GitHub repo (ibm-granite/granite-4.0-language-models) or run a test. Likely 4K–32K based on Granite family conventions; confirm before production use.

Run Your Own AI—Securely & Affordably

Granite-4.0-H-Tiny is a lean, enterprise-ready model for building private AI systems on your terms. Let LLM.co help you architect a complete ops AI stack: integration, fine-tuning, deployment, and ongoing ops. Start building.