Granite 3.3 8B

Positioning

IBM Granite 3.3 8B is a dense 8-billion-parameter language model released under the permissive Apache 2.0 license. It is an iterative refresh of Granite 3.2, sharing the same architecture but with improved instruction following and tool-call reliability. IBM positions this model for enterprise tool-calling on IBM stacks, making it a natural fit for organizations already invested in IBM's ecosystem. With a 131K context window, it supports long-document processing and multi-turn agentic workflows.

Strengths

• Permissive Apache 2.0 license: No restrictions on commercial use, modification, or redistribution — ideal for enterprise deployment.
• Long 131K context window: Supports processing of lengthy documents, codebases, or conversation histories without truncation.
• Dense 8B parameter architecture: Predictable memory and compute requirements; no routing overhead or load-balancing issues common with MoE models.
• Enterprise-focused tuning: IBM's iterative improvements target instruction following and tool-call reliability, directly addressing common pain points in production agentic systems.

Limitations

• No community-reported benchmarks available: Operators should treat vendor-published metrics as best-case until independent third-party measurements emerge.
• 8B dense model may lag behind larger models on complex reasoning: For tasks requiring deep multi-step reasoning or broad world knowledge, larger models (e.g., 70B+ dense or MoE) may be more capable.
• Tool-calling optimization may not generalize: Improvements are tailored to IBM stacks; performance with non-IBM tools or custom APIs is unverified.
• Iterative refresh, not a new architecture: Gains over Granite 3.2 are incremental; users already on 3.2 may see only modest improvements.

What it takes to run this locally

At 8B parameters, quantized sizes range from 16 GB (FP16) down to ~2.6 GB (Q2_K). For practical use with a moderate context length, add ~30-50% for KV cache and framework overhead. A Q4_K_M quant (4.5 GB) plus overhead fits comfortably on a single consumer GPU with 8-12 GB VRAM (e.g., RTX 3060 12GB, RTX 4070). For full FP16 precision, a 24 GB GPU (e.g., RTX 3090/4090) is recommended. This model is firmly in the consumer deployment class.

Should you run this locally?

Yes if you need a permissively licensed, enterprise-tuned model for tool-calling and instruction following, especially within IBM infrastructure, and you have a consumer GPU with at least 8 GB VRAM for quantized inference.

No if you require state-of-the-art reasoning or broad world knowledge beyond what an 8B dense model can provide, or if you prefer to wait for independent community benchmarks before committing to deployment.

Catalog cross-links

• Granite 3.2 8B
• IBM
• Apache 2.0 License