DEV.co
Open-Source LLM · ibm-granite

granite-4.0-tiny-preview

Granite-4.0-Tiny-Preview is a 6.7B parameter open-source LLM from IBM with mixture-of-experts (MoE) architecture, released May 2025. It is instruction-tuned and supports 12 languages, designed for general task execution (summarization, QA, code, RAG, long-context). Licensed under Apache 2.0 and ungated. Model card indicates it is preview/early release status.

Source: HuggingFace — huggingface.co/ibm-granite/granite-4.0-tiny-preview
6.7B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
154.6k
Downloads (30d)

Key facts

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

FieldValue
Developeribm-granite
Parameters6.7B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads154.6k
Likes184
Last updated2025-11-03
Sourceibm-granite/granite-4.0-tiny-preview

What granite-4.0-tiny-preview is

Fine-grained hybrid MoE instruct model (7B params) derived from Granite-4.0-Tiny-Base-Preview, trained with supervised fine-tuning and reinforcement learning alignment on permissively-licensed public datasets and internal synthetic data. Uses chat format with thinking capability. Implemented in transformers; native HF support pending (PR#37658). Requires transformers v4.45+ or source install. Evaluation shows modest performance vs. 8B predecessor on Arena-Hard (26.70) and AlpacaEval-2.0 (35.16), but stronger on code (HumanEval 82.41) and reasoning tasks. Context length and exact MoE configuration unknown.

Quickstart

Run granite-4.0-tiny-preview 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-tiny-preview")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

Long-Context Business Applications

Designed for long document/meeting summarization and QA. Suitable for contract analysis, policy review, or internal knowledge retrieval in enterprise workflows.

Code-Related Tasks & Function-Calling

HumanEval benchmark (82.41) and stated code capability indicate viability for code generation, refactoring, and API integration tasks within custom applications.

Multilingual Conversational Assistants

Native support for 12 languages (EN, DE, ES, FR, JA, PT, AR, CS, IT, KO, NL, ZH) enables deployment of AI assistants across geographies without additional training.

Running & fine-tuning it

Estimated 13–16 GB VRAM for inference (bfloat16, single GPU based on code example using device_map='cuda'). MoE architecture may reduce per-token compute vs. dense 7B, but exact sparse routing overhead unknown. Fine-tuning memory requirements unknown; full-parameter or LoRA feasibility requires verification.

Model card states fine-tuning is supported and mentions SFT/RLHF alignment techniques used in training. LoRA/QLoRA compatibility not explicitly stated. Users can fine-tune for languages beyond the 12 supported. Requires transformers from source or v4.45+; dependency chain impact on adapter frameworks unclear.

When to avoid it — and what to weigh

  • High Accuracy General Knowledge Required — Arena-Hard (26.70) and AlpacaEval-2.0 (35.16) scores are significantly lower than contemporary 8B+ models. Not recommended for high-stakes QA or benchmark-critical applications.
  • CPU-Only or Edge Deployment — Model size and MoE architecture likely require GPU (transformer source install, bfloat16 precision in examples). CPU inference speed and quantization compatibility unknown.
  • Production with Stability/Longevity Concerns — Labeled 'preview' status as of May 2025, released November 2024. Early lifecycle; breaking changes or discontinuation possible. Evaluate migration risk.
  • Retrieval Accuracy or Factuality-Critical Use — TruthfulQA (58.07) and PopQA (22.93) scores suggest limitations in factual recall. Not suitable for medical, legal, or high-confidence knowledge domains without external retrieval.

License & commercial use

Apache License 2.0 (permissive, OSI-approved). Allows commercial use, modification, and redistribution with attribution. Full license text available at apache.org.

Apache 2.0 is a permissive OSI license that explicitly permits commercial use, including proprietary applications and SaaS. No additional licensing required. However, verify that your deployment infrastructure (hosting, orchestration) complies with any cloud provider or container licensing terms. IBM imposes no stated restrictions on commercial deployment.

DEV.co evaluation signals

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

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityModerate
DEV.co fitGood
Assessment confidenceMedium
Security considerations

Model trained on permissively-licensed public data plus IBM internal synthetic data. No explicit mention of adversarial testing, jailbreak evaluation, or prompt-injection hardening. TruthfulQA score (58.07) suggests potential for hallucination/confabulation. Recommend input validation, output filtering, and user instruction review for sensitive applications. No CVE or vulnerability data provided. Code execution capability (function-calling) requires sandbox/validation. Data provenance for internal synthetic datasets not disclosed.

Alternatives to consider

Granite-3.3-8B-Instruct

IBM's prior 8B instruct model shows higher Arena-Hard (57.56) and AlpacaEval (62.68) scores, suggesting stronger general performance. More mature; not preview status. Consider if higher accuracy outweighs resource cost.

Mistral 7B Instruct v0.2

Comparable parameter count, permissive Apache 2.0 license, wider adoption, better downstream benchmark parity. Stronger community support and established serving ecosystem.

LLaMA 3.2 (8B/70B)

Meta's 8B variant offers permissive Llama License, broader evals coverage, native transformer support. 70B option available for higher accuracy. Larger ecosystem and production track record.

Software development agency

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

Download the model from Hugging Face, test on your hardware, and review the code example and Granite Docs. For production deployment, verify transformer dependencies, evaluate fine-tuning on your domain, and validate factuality/hallucination risk with your use case. Contact IBM Granite team for preview-status clarifications.

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-tiny-preview FAQ

Can I use Granite-4.0-Tiny-Preview in a commercial product?
Yes. Apache 2.0 is a permissive OSI license that explicitly permits commercial use without additional licensing. You must include a copy of the license and attribute IBM/Granite Team in distribution. Verify your hosting and infrastructure vendors do not impose conflicting restrictions.
What GPU is required to run this model?
Based on code examples using bfloat16, estimate 13–16 GB VRAM for inference on a single GPU (e.g., RTX 4080, A100 40GB, H100). MoE sparse routing may reduce per-token memory vs. dense models. Fine-tuning memory unknown; test with your framework. CPU-only inference not evaluated.
Is this model production-ready?
Model is labeled 'preview' as of May 2025 release. No guarantees on stability, breaking changes, or long-term support. Evaluate risk for critical applications. Monitor Granite Docs and IBM GitHub for updates before full production rollout.
Can I fine-tune Granite-4.0-Tiny-Preview for my domain?
Yes, fine-tuning is supported (card mentions SFT/RLHF techniques). LoRA/QLoRA compatibility not explicitly stated; test with your framework. Requires transformers v4.45+ or source build. Contact IBM or test in sandbox before large-scale training.

Software development & web development with DEV.co

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If granite-4.0-tiny-preview is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Ready to Deploy Granite-4.0-Tiny-Preview?

Download the model from Hugging Face, test on your hardware, and review the code example and Granite Docs. For production deployment, verify transformer dependencies, evaluate fine-tuning on your domain, and validate factuality/hallucination risk with your use case. Contact IBM Granite team for preview-status clarifications.