Granite-4.2-3B is the most compact model in the family, aimed at scenarios with strict computational-resource constraints, while still retaining the key feature of the 4.2 generation — built-in reasoning. Architecture: a dense transformer with 3 billion parameters, 40 layers, 40 query heads and 8 KV heads, and an MLP hidden size of 8192. The model uses the same GQA scheme with full attention within a 128K-token context, RoPE, SwiGLU and RMSNorm. Full attention is especially important here because it allows even a compact model to build logical connections across the entire declared input length.
The 3B model supports three thinking modes — full, non-thinking and low-effort — as well as native tool calling, which makes it suitable for lightweight agentic scenarios and conversational interfaces. However, unlike the 8B and 30B models, it does not go through the agentic RL stage (SWE, terminal, search) and relies on basic RLVR for code and instruction following. This is not a drawback but a deliberate choice: the 3B model is intended for cases where autonomous solving of complex engineering tasks is not required, but a reliable reasoning model is needed that can be run on an edge device or in a memory-constrained environment.
Among the benchmarks, AIME25 (78.33) stands out — for a model of this size, this is a very high result. MMLU-Pro (67.84) reflects broad knowledge and instruction-following ability at a level comparable to larger models, which is important for a general-purpose assistant. RULER 128K (55.30) shows that the model handles long context, although this is not a high score, it is an acceptable result that allows it to be used in summarization and question-answering systems that explore medium-length documents.
Granite-4.2-3B is best suited for edge deployment, embedded assistants, document summarization, code/text autocomplete suggestions, and other lightweight tasks in memory-constrained environments. All models are distributed under the Apache 2.0 license, which removes concerns about commercial use.
| Model Name | Context | Type | GPU | Status | Link |
|---|
There are no public endpoints for this model yet.
An endpoint is a server with pre-installed software that automatically downloads neural network weights and launches an OpenAI-compatible API to handle incoming requests.
A private endpoint is created within the user's project, belongs exclusively to the user, and is billed per server configuration in accordance with the infrastructure pricing.
We recommend creating private endpoints when you need to:
| Name | GPU | TPS | Max Concurrency | |||
|---|---|---|---|---|---|---|
131,072.0 |
1 | $0.53 | 1.957 | Launch | ||
131,072.0 tensor |
2 | $0.57 | 2.322 | Launch | ||
131,072.0 |
1 | $0.83 | 2.087 | Launch | ||
131,072.0 tensor |
2 | $0.97 | 1.098 | Launch | ||
131,072.0 |
1 | $1.02 | 2.082 | Launch | ||
131,072.0 tensor |
2 | $1.56 | 4.521 | Launch | ||
131,072.0 |
1 | $1.59 | 3.042 | Launch | ||
131,072.0 |
1 | $2.37 | 8.918 | Launch | ||
131,072.0 |
1 | $3.83 | 8.909 | Launch | ||
131,072.0 |
1 | $4.11 | 10.616 | Launch | ||
131,072.0 tensor |
2 | $4.61 | 18.183 | Launch | ||
131,072.0 |
1 | $4.74 | 16.345 | Launch | ||
131,072.0 tensor |
2 | $9.40 | 33.038 | Launch | ||
| Name | GPU | TPS | Max Concurrency | |||
|---|---|---|---|---|---|---|
131,072.0 |
1 | $0.53 | 1.785 | Launch | ||
131,072.0 tensor |
2 | $0.57 | 2.151 | Launch | ||
131,072.0 |
1 | $0.83 | 1.915 | Launch | ||
131,072.0 |
1 | $1.02 | 1.911 | Launch | ||
131,072.0 pipeline |
3 | $1.43 | 1.547 | Launch | ||
131,072.0 tensor |
2 | $1.56 | 4.350 | Launch | ||
131,072.0 |
1 | $1.59 | 2.871 | Launch | ||
131,072.0 |
1 | $2.37 | 8.747 | Launch | ||
131,072.0 |
1 | $3.83 | 8.738 | Launch | ||
131,072.0 |
1 | $4.11 | 10.445 | Launch | ||
131,072.0 tensor |
2 | $4.61 | 18.012 | Launch | ||
131,072.0 |
1 | $4.74 | 16.174 | Launch | ||
131,072.0 tensor |
2 | $9.40 | 32.867 | Launch | ||
| Name | GPU | TPS | Max Concurrency | |||
|---|---|---|---|---|---|---|
131,072.0 |
1 | $0.53 | 1.395 | Launch | ||
131,072.0 tensor |
2 | $0.57 | 1.761 | Launch | ||
131,072.0 |
1 | $0.83 | 1.525 | Launch | ||
131,072.0 |
1 | $1.02 | 1.520 | Launch | ||
131,072.0 pipeline |
3 | $1.43 | 1.156 | Launch | ||
131,072.0 tensor |
2 | $1.56 | 3.959 | Launch | ||
131,072.0 |
1 | $1.59 | 2.481 | Launch | ||
131,072.0 |
1 | $2.37 | 8.356 | Launch | ||
131,072.0 |
1 | $3.83 | 8.348 | Launch | ||
131,072.0 |
1 | $4.11 | 10.054 | Launch | ||
131,072.0 tensor |
2 | $4.61 | 17.622 | Launch | ||
131,072.0 |
1 | $4.74 | 15.784 | Launch | ||
131,072.0 tensor |
2 | $9.40 | 32.477 | Launch | ||
Contact our dedicated neural networks support team at support@immers.cloud or send your request to the sales department at sale@immers.cloud.