← All models

Granite 4.0 H Micro

Language Available Full comparison ↗

ibm/granite-4-0-h-micro · by IBM · decoder-only-transformer

Pricing — 1 offering(s)

OpenRouter

Input tokens

$0.017 / 1M tokens (input) Current 2025-10-02 → present

Output tokens

$0.112 / 1M tokens (output) Current 2025-10-02 → present
Same sourcing caveat as the input tier.

Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.

See how Granite 4.0 H Micro fits into a cost-aware routing setup

Capability profile

How Granite 4.0 H Micro rates across core capability dimensions, with the task-level evidence behind each rating.

instruction following Strong

IBM reports the Granite 4.0 family as class-leading among open-weight models on instruction following and function calling for its size, surpassing "nearly all open-weight models" at comparable cost.

tool use Strong

Function calling is a stated focus of Granite 4.0; positioned for agentic and RAG pipelines.

reasoning Weak

3B model — not a reasoning model. The 'H Micro' tier is about fast execution, not deep analysis.

coding Moderate

context window Moderate

This OpenRouter-hosted endpoint serves a ~131K window; IBM trained on up to 512K and validated to 128K (see registry note on the training-vs-served-context distinction).

multilingual Moderate

IBM lists 12 languages officially supported for Granite 4.0.

speed Strong

The Mamba-2 hybrid plus 3B size makes it one of the fastest / lowest-memory models in this registry — designed for edge, local, and high-throughput serving.

cost efficiency Strong

$0.017 / 1M input, $0.112 / 1M output via OpenRouter — near the bottom of the price range here, and Apache-2.0 for free self-hosting.

Operator guidance

Route here for cost-, latency-, or memory-bound tasks that are mostly instruction-following, tool-calling, or retrieval-grounded rather than open-ended reasoning. For anything needing real analytical depth, step up to a larger Granite 4.0 tier (H Small) or a general mid-size model. Its differentiators versus other cheap small models are the Mamba-2 hybrid efficiency, the strong function-calling story for its size, and IBM's enterprise provenance guarantees.

Use cases

Edge / on-device / local-first deployments where memory and latency are hard constraints
High-volume function-calling and RAG pipelines where a small, cheap, instruction-tuned model is sufficient
Enterprise settings that need Apache-2.0 licensing, signed checkpoints, and ISO 42001 provenance
The cheap tier in a cascade, handling routing / extraction / tool-dispatch before a larger model

Limitations

3B parameters — weak on hard reasoning; not for analysis-heavy work
No IBM first-party public per-token API (watsonx.ai is enterprise-platform pricing); this entry is hosted via OpenRouter (see registry note)
Served context (~131K) is lower than the 512K training / 128K eval figures IBM cites
Capability ratings are qualitative from IBM's announcements; the benchmark claims are IBM-reported