Granite-4.2-8B occupies the middle position in the family and offers a trade-off between reasoning quality and computational efficiency. It is a dense decoder-only transformer with 8 billion parameters, 40 layers, 32 query heads, 8 KV heads, and an MLP hidden size of 12,800. Architectural principles: GQA with full attention within the context, RoPE with θ = 10,000,000, SwiGLU, RMSNorm, and bfloat16. Full attention here preserves the model’s ability to build precise long-range dependencies required for agentic loops and long documents, while GQA keeps inference cost reasonable for a mid-sized model.
What distinguishes it from Granite-4.1 is primarily the addition of native reasoning and agentic RL. Like the 30B, the model supports three thinking modes (full, non-thinking, low-effort) and native tool calling with reasoning during function selection. During training, the model went through an agentic RL stage—the same SWE, terminal, and search stages as the flagship, although with fewer RLVR rounds (2 versus 3 for the 30B). This means that 8B is capable not only of following instructions but also of acting in real environments: editing code, executing commands, and conducting multi-step search. At the same time, the model remains compact enough for deployment, making it a practical choice for mid-scale agentic systems.
Among the benchmarks, SWE Bench Verified (47.67) is indicative—the ability to solve real tasks in repositories; AIME25 (86.67) demonstrates strong mathematical reasoning—the result is close to the flagship even with 8 billion parameters. RULER 128K (71.41) confirms long-context capability: the model retains significant accuracy on a 128K context.
Granite-4.2-8B is a general-purpose production choice for mid-scale agentic systems: tool calling, software development process automation, multilingual dialogues, and tasks with reasonably long context where both quality and inference cost matter.
| 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 tensor |
2 | $0.93 | 1.443 | Launch | ||
131,072.0 pipeline |
3 | $1.06 | 1.145 | Launch | ||
131,072.0 tensor |
4 | $1.26 | 1.717 | Launch | ||
131,072.0 tensor |
2 | $1.56 | 1.540 | Launch | ||
131,072.0 tensor |
2 | $1.56 | 1.540 | Launch | ||
131,072.0 tensor |
2 | $1.92 | 1.536 | Launch | ||
131,072.0 |
1 | $2.37 | 3.189 | Launch | ||
131,072.0 tensor |
2 | $2.93 | 2.257 | Launch | ||
131,072.0 |
1 | $3.83 | 3.186 | Launch | ||
131,072.0 |
1 | $4.11 | 3.826 | Launch | ||
131,072.0 tensor |
2 | $4.61 | 6.664 | Launch | ||
131,072.0 |
1 | $4.74 | 5.974 | Launch | ||
131,072.0 tensor |
2 | $9.40 | 12.234 | Launch | ||
| Name | GPU | TPS | Max Concurrency | |||
|---|---|---|---|---|---|---|
131,072.0 tensor |
2 | $0.93 | 1.280 | Launch | ||
131,072.0 tensor |
4 | $1.26 | 1.555 | Launch | ||
131,072.0 tensor |
2 | $1.56 | 1.378 | Launch | ||
131,072.0 tensor |
2 | $1.56 | 1.378 | Launch | ||
131,072.0 tensor |
2 | $1.92 | 1.374 | Launch | ||
131,072.0 |
1 | $2.37 | 3.027 | Launch | ||
131,072.0 tensor |
2 | $2.93 | 2.094 | Launch | ||
131,072.0 |
1 | $3.83 | 3.023 | Launch | ||
131,072.0 |
1 | $4.11 | 3.663 | Launch | ||
131,072.0 tensor |
2 | $4.61 | 6.501 | Launch | ||
131,072.0 |
1 | $4.74 | 5.812 | Launch | ||
131,072.0 tensor |
2 | $9.40 | 12.072 | Launch | ||
| Name | GPU | TPS | Max Concurrency | |||
|---|---|---|---|---|---|---|
131,072.0 tensor |
4 | $1.26 | 1.184 | Launch | ||
131,072.0 pipeline |
3 | $1.34 | 1.650 | Launch | ||
131,072.0 tensor |
2 | $1.56 | 1.007 | Launch | ||
131,072.0 tensor |
2 | $1.56 | 1.007 | Launch | ||
131,072.0 tensor |
4 | $1.57 | 2.637 | Launch | ||
131,072.0 tensor |
2 | $1.92 | 1.003 | Launch | ||
131,072.0 |
1 | $2.37 | 2.656 | Launch | ||
131,072.0 tensor |
2 | $2.93 | 1.723 | Launch | ||
131,072.0 |
1 | $3.83 | 2.652 | Launch | ||
131,072.0 |
1 | $4.11 | 3.292 | Launch | ||
131,072.0 tensor |
2 | $4.61 | 6.130 | Launch | ||
131,072.0 |
1 | $4.74 | 5.441 | Launch | ||
131,072.0 tensor |
2 | $9.40 | 11.701 | Launch | ||
Contact our dedicated neural networks support team at support@immers.cloud or send your request to the sales department at sale@immers.cloud.