Source: ibm-granite/granite-4.2-30b BF16 safetensors
Tooling: nvidia-modelopt 0.46.x with mtq.NVFP4_DEFAULT_CFG (group size 16)
Export: real NVFP4 checkpoint (quant_algo: NVFP4 in hf_quant_config.json; packed U8 weight + F8_E4M3 scale pairs — not BF16 passthrough)
Calibration: mixed instruction / math / code / tool-call prompts on DGX Spark (GB10)

Measured artifact size

Artifact	Size (measured)
Full NVFP4 repo (safetensors + config/tokenizer)	16.45 GB (17660794999 bytes) (weights 16.44 GB (17650902032 bytes))

Serve (vLLM on Blackwell / GB10)

bash

hf download vcruz305/granite-4.2-30b-NVFP4 --local-dir granite-4.2-30b-NVFP4

vllm serve ./granite-4.2-30b-NVFP4 \
  --quantization modelopt \
  --dtype auto \
  --host 127.0.0.1 --port 8200 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.9

If your vLLM build expects the FP4 alias, try --quantization modelopt_fp4. Verify hf_quant_config.json shows quant_algo: NVFP4 before serving.

Verification

hf_quant_config.json → quant_algo: NVFP4, group_size 16, producer modelopt 0.46.x
Safetensor dtypes (local export scan): U8=448, F8_E4M3=448, BF16=131 (unquantized heads/norms/etc.)
Host export: DGX Spark GB10 (wesche-spark-9f73), mtq.NVFP4_DEFAULT_CFG + export_hf_checkpoint

Shard sizes

File	Size
`model-00001-of-00002.safetensors`	9.32 GB (10004186744 bytes)
`model-00002-of-00002.safetensors`	7.12 GB (7646715288 bytes)

Companion GGUF ladder: vcruz305/granite-4.2-30b-GGUF.

Original model card (IBM Granite-4.2-30B)

Granite-4.2-30B

GitHub

Model Summary


Developers	Granite Team, IBM
Model Type	Decoder-only Dense Transformer (Reasoning)
Architecture	GraniteForCausalLM
Base Model	Granite-4.1-30B-Base
Parameters	30B
Context Length	Natively Supports 128K (Long-context extension to 512K)
Precision	bfloat16
Tested Languages	English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, Chinese (other languages may work but have not been fully tested)
Reasoning Mode	Built-in `<think>...</think>` chain-of-thought
Best For	Reasoning, Code Generation, Tool Calling, Agentic Workflows, Multilingual Dialog
License	Apache 2.0
HF Collection	Granite 4.2 Language Models
Release Date	August 25, 2026

Model Overview

What is IBM Granite?

Granite is a family of open-source large language models developed by IBM, designed for enterprise and research use. Granite models are built to be versatile, safe, and efficient — covering a range of sizes and capabilities from compact edge-deployable models to large-scale reasoning systems. All Granite models are released under the Apache 2.0 license, enabling unrestricted commercial and academic use.

The Granite 4.2 generation introduces native reasoning (thinking) capabilities, allowing models to perform step-by-step chain-of-thought reasoning before producing final answers. This significantly improves performance on complex math, coding, multi-step logic, and agentic tool-calling tasks.

Description

Granite-4.2-30B is the flagship reasoning model in the Granite 4.2 family. It delivers the strongest performance across reasoning-intensive tasks by leveraging built-in <think>...</think> chain-of-thought. It supports flexible thinking modes — full thinking (default), non-thinking, and low-effort — allowing users to balance depth vs. latency on a per-query basis.

Key capabilities:

Built-in Reasoning: Native chain-of-thought that significantly improves performance on math, coding, and complex multi-step problems.
Flexible Thinking Modes: Seamlessly switch between full thinking, non-thinking, and low-effort modes within a single model.
Reasoning-Augmented Tool Calling: The model reasons about which tools to invoke and why, producing more accurate function calls.
512K Context Window: Supports long documents, multi-turn conversations, and complex agentic workflows.
Apache 2.0 Licensed: Fully open for commercial and research use.

Model Design

Granite-4.2-30B is built on a decoder-only dense transformer architecture with the following core components:

Attention: Grouped Query Attention (GQA) with 32 attention heads and 8 KV heads
Position Embedding: Rotary Position Embedding (RoPE) with θ = 10,000,000
Feed-Forward: MLP with SwiGLU activation (hidden size 32768)
Normalization: RMSNorm (ε = 1e-5)
Embeddings: Separate input/output embeddings (not tied)
Precision: bfloat16

Component	3B Dense	8B Dense	30B Dense
Embedding size	2560	4096	4096
Number of layers	40	40	64
Attention head size	64	128	128
Number of attention heads	40	32	32
Number of KV heads	8	8	8
MLP hidden size	8192	12800	32768
MLP activation	SwiGLU	SwiGLU	SwiGLU
Sequence length	131072	131072	131072
Position embedding	RoPE	RoPE	RoPE
# Parameters	3B	8B	30B

Training Methodology

Granite-4.2-30B is post-trained from Granite-4.1-30B-Base through a rigorous multi-stage pipeline that progressively unlocks reasoning, tool use, and instruction-following capabilities. A full listing of training datasets is available in the Granite 4.2 GitHub repository. The training pipeline consists of three stages:

Stage 1: Pre-Training

Granite-4.2-30B builds on Granite-4.1-30B-Base, which was pre-trained on a large-scale English as well as multilingual corpus. For full pre-training details (data composition, training recipe, and infrastructure), refer to our Granite 4.1 Technical Blog.

Stage 2: Supervised Fine-Tuning

The SFT stage draws on instruction-following, chain-of-thought, and reasoning data to cultivate the model's reasoning and thinking abilities. For all the three, 3B, 8B and 30B models, the training corpus comprises four sources: (1) publicly available datasets under permissive licenses, (2) internally generated synthetic data targeting reasoning, tool calling, and chain-of-thought capabilities, (3) agentic traces collected across a diverse range of tasks, and (4) a curated selection of human-authored data. Hyperparameters were tuned before training was scaled to all three model sizes. For the 30B model, we conducted a second SFT phase, in which the agentic data was up-sampled while a smaller share of general replay data was retained. This phase trained for a single epoch, starting from a lower learning rate than Phase 1.

Stage 3: Reinforcement Learning

The final stage of training applies multi-phase, multi-environment reinforcement learning using Group Relative Policy Optimization (GRPO). Training spans a broad mix of environments including math, code, science, instruction following, tool use, general chat and structured output. Most environments provide verifiable rewards, while open-ended prompts are scored by a generative reward model. Training runs asynchronously: generation and policy updates occupy separate GPU pools rather than proceeding in lockstep, and weights are refreshed in flight.

After the reward-driven phases, a preference-alignment (RLHF) phase tunes helpfulness, conversational quality, and safety. Reinforcement learning is carried out with NeMo RL, and the RL environments run on NeMo Gym.

Infrastructure: We trained the Granite 4.2 Language Models utilizing an NVIDIA GB200 NVL72 cluster hosted in CoreWeave. Intra-rack communication occurs via the 72-GPU NVLink domain, and a non-blocking, full Fat-Tree NDR 400 Gb/s InfiniBand network provides inter-rack communication.

For further details on the post-training methodology, please refer to the Granite-4.2 Technical Blog.

Evaluation Results

Task	3B Dense	8B Dense	30B Dense
Agentic (Coding)
SWE Bench Multilingual	NA	30.78	41.89
SWE Bench Pro	NA	19.11	33.29
SWE Bench Verified	NA	47.67	57
Terminal-Bench 2.1	NA	20.56	29.24
Agentic (General)
τ³-bench (AVG)	45.78	58.06	62.00
BFCL (v4)	52.41	52.39	61.39
ProfBench	32.10	41.20	42.90
BirdBench	NA	41.07	41.85
GDPval	NA	1189	1225
Reasoning
AIME25	78.33	86.67	89.17
HMMT Feb25	66.67	78.33	89.17
GPQA	54.80	64.14	66.41
LiveCodeBench v6	69.71	73.24	75.77
SciCode	24.11	36.09	38.76
Chat & Instruction Following
MMLU-Pro	67.84	74.04	77.60
MMLU-ProX lite (IBM)	27.78	61.06	66.64
Arena-Hard-V2	34.96	65.19	67.93
IFBench (prompt)	74.33	79.33	77.17
Long Context
RULER 64K	67.52	80.99	89.96
RULER 128K	55.30	71.41	81.38

Evaluations are run with an evaluation framework based on NeMo Evaluator SDK.

Inference

Generation Parameters

Important: Use temperature=1.0 and top_p=0.95 across all tasks and serving backends, including general chat, reasoning, and tool calling.

Parameter	Value	Notes
`temperature`	`1.0`	Required for all modes
`top_p`	`0.95`	Nucleus sampling threshold
`max_new_tokens`	`8192`	Thinking mode (increase for complex reasoning)
`max_new_tokens`	`2048`	Non-thinking mode
`do_sample`	`True`	Required when temperature > 0

Thinking Modes

Mode	Template Parameters	Behavior
Thinking (default)	`enable_thinking=True`	Full chain-of-thought reasoning inside `<think>...</think>`
Non-thinking	`enable_thinking=False`	Direct answer with no reasoning overhead
Low-effort	`enable_thinking=True, low_effort=True`	Brief reasoning for simpler queries

How It Works

Thinking enabled — The generation prompt ends with <|im_start|>assistant\n<think>\n, causing the model to reason until it emits </think>, then produce the final answer.
Thinking disabled — The prompt ends with <|im_start|>assistant\n<think></think>, bypassing reasoning entirely.
Low-effort — Appends {reasoning effort: low} to the user message, signaling shorter reasoning chains.

History Truncation

In multi-turn conversations, thinking content from previous assistant turns is automatically stripped (truncate_history_thinking=True by default) to conserve context window space. Only the current generation produces full reasoning. Set truncate_history_thinking=False to preserve full reasoning history.

Serving with vLLM

Granite-4.2-30B is optimized for deployment with vLLM.

Reasoning parser: Use the custom granite_thinking_parser included in this repository (requires vLLM v0.20+). The model also works with the built-in nemotron_v3 parser, but granite_thinking_parser provides better formatting of reasoning output. Native support for granite_thinking_parser will be added to vLLM and SGLang very soon. Tool calling parser: Use qwen3_coder.

Starting the Server

bash

vllm serve ibm-granite/granite-4.2-30b \
    --served-model-name granite-4.2-30b \
    --dtype bfloat16 \
    --max-model-len 131072 \
    --reasoning-parser granite_thinking_parser \
    --reasoning-parser-plugin ./granite_thinking_parser.py \
    --tool-call-parser qwen3_coder \
    --enable-auto-tool-choice

OpenAI-Compatible API Usage

python

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="unused")

response = client.chat.completions.create(
    model="granite-4.2-30b",
    messages=[{"role": "user", "content": "Explain the Riemann hypothesis in simple terms."}],
    temperature=1.0,
    top_p=0.95,
    max_tokens=8192,
)

print(response.choices[0].message.content)

Tool Calling via vLLM

python

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="unused")

tools = [
    {
        "type": "function",
        "function": {
            "name": "get_current_weather",
            "description": "Get the current weather for a specified city.",
            "parameters": {
                "type": "object",
                "properties": {
                    "city": {"type": "string", "description": "Name of the city"}
                },
                "required": ["city"]
            }
        }
    }
]

response = client.chat.completions.create(
    model="granite-4.2-30b",
    messages=[{"role": "user", "content": "What's the weather like in Boston right now?"}],
    tools=tools,
    temperature=1.0,
    top_p=0.95,
    max_tokens=4096,
)

print(response.choices[0].message.tool_calls)

Using with Agentic Coding Harnesses

Granite-4.2-30B can be used as the backbone model for agentic coding tools. Since it supports reasoning and tool calling via the OpenAI-compatible API, it integrates with popular agentic harnesses out of the box. Start the vLLM server as shown in the Serving with vLLM section above, then follow the harness-specific instructions below.

OpenCode

OpenCode is an AI coding agent that runs in your terminal.

Install:

curl -fsSL https://opencode.ai/install | bash

Configure ~/.config/opencode/opencode.json:

json

{
  "$schema": "https://opencode.ai/config.json",
  "model": "local/granite-4.2-30b",
  "provider": {
    "local": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "vLLM (local)",
      "options": {
        "baseURL": "http://localhost:8000/v1",
        "apiKey": "EMPTY"
      },
      "models": {
        "granite-4.2-30b": {
          "name": "Granite 4.2 30B",
          "limit": {
            "context": 131072,
            "output": 8192
          }
        }
      }
    }
  }
}

Run:

bash

opencode
opencode run "your task description"

For full documentation, see opencode.ai/docs.

Pi

Pi is a minimal agent harness for AI-powered coding that runs in your terminal. It supports custom providers via a models.json configuration file.

Install:

curl -fsSL https://pi.dev/install.sh | sh

Configure ~/.pi/agent/models.json:

json

{
  "providers": {
    "vllm": {
      "baseUrl": "http://localhost:8000/v1",
      "api": "openai-completions",
      "apiKey": "EMPTY",
      "compat": {
        "supportsDeveloperRole": false,
        "supportsReasoningEffort": false
      },
      "models": [
        {
          "id": "granite-4.2-30b",
          "name": "Granite 4.2 30B",
          "reasoning": true,
          "input": ["text"],
          "contextWindow": 131072,
          "maxTokens": 8192,
          "samplingParams": {
            "temperature": 1.0,
            "top_p": 0.95
          },
          "cost": { "input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0 }
        }
      ]
    }
  }
}

Run:

pi

Then select the granite-4.2-30b model with /model or Ctrl+L in the interactive session.

For full documentation, see pi.dev/docs.

OpenHands

OpenHands is an AI software engineer that can plan, write code, and execute commands.

Install and launch OpenHands following the official installation guide.
Configure the LLM in the OpenHands settings with:
- Model: granite-4.2-30b
- Base URL: http://localhost:8000/v1
- API Key: your vLLM --api-key value

Note: The openai/ prefix is required when connecting to OpenAI-compatible endpoints like vLLM. Refer to the OpenHands local LLM documentation for detailed setup instructions, troubleshooting, and alternative installation methods.

Quick Start (Transformers)

Installation

bash

pip install torch torchvision torchaudio
pip install accelerate transformers

Basic Inference (Thinking Mode)

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = "ibm-granite/granite-4.2-30b"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path, device_map="cuda", torch_dtype=torch.bfloat16)
model.eval()

messages = [
    {"role": "user", "content": "How many r's are in the word 'strawberry'?"},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=8192, temperature=1.0, top_p=0.95, do_sample=True)

print(tokenizer.decode(output[0][inputs.input_ids.shape[-1]:], skip_special_tokens=False))

Example Output

<think>
Okay, let's see. The problem is to find how many 'r's are in the word 'strawberry'.

First, I need to write out the word: s t r a w b e r r y.

Now, I need to count the number of 'r' letters. Let's list each letter and check for 'r'.

1. s – not r
2. t – not r
3. r – yes, that's one
4. a – no
5. w – no
6. b – no
7. e – no
8. r – yes, that's two
9. r – yes, that's three
10. y – no

Total r's = 3.
</think>
There are **3** r's in the word "strawberry".<|im_end|>

Non-Thinking Mode

python

messages = [
    {"role": "user", "content": "What is the capital of France?"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=2048, temperature=1.0, top_p=0.95, do_sample=True)
print(tokenizer.decode(output[0][inputs.input_ids.shape[-1]:], skip_special_tokens=False))

Example Output

<think></think>The capital of France is Paris.<|im_end|>

Low-Effort Thinking

python

messages = [
    {"role": "user", "content": "What is 2 + 2?"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True,
                                     enable_thinking=True, low_effort=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=4096, temperature=1.0, top_p=0.95, do_sample=True)
print(tokenizer.decode(output[0][inputs.input_ids.shape[-1]:], skip_special_tokens=False))

Example Output

<think>
Simple answer.
</think>
2 + 2 = 4.<|im_end|>

Tool Calling

Granite-4.2-30B supports tool calling with integrated reasoning — the model thinks about which tool to call and why before making the call. Tools are defined using the OpenAI function definition schema.

Basic Tool Calling

python

tools = [
    {
        "type": "function",
        "function": {
            "name": "get_current_weather",
            "description": "Get the current weather for a specified city.",
            "parameters": {
                "type": "object",
                "properties": {
                    "city": {"type": "string", "description": "Name of the city"}
                },
                "required": ["city"]
            }
        }
    }
]

messages = [
    {"role": "user", "content": "What's the weather like in Boston right now?"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, tools=tools,
                                     add_generation_prompt=True, enable_thinking=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=4096, temperature=1.0, top_p=0.95, do_sample=True)
print(tokenizer.decode(output[0][inputs.input_ids.shape[-1]:], skip_special_tokens=False))

Example Output

<think>
Okay, the user is asking for the weather in Boston right now. Let me check the tools
available. There's a function called get_current_weather that takes a city parameter.
I need to call that with the city set to Boston.
</think>
<tool_call>
<function=get_current_weather>
<parameter=city>
Boston
</parameter>
</function>
</tool_call>
<|im_end|>

Multi-Turn with Tool Response

python

messages = [
    {"role": "user", "content": "What's the weather like in Boston right now?"},
    {"role": "assistant", "content": "<think>\nThe user wants to know the current weather in Boston. I should call get_current_weather.\n</think>",
     "tool_calls": [{"function": {"name": "get_current_weather", "arguments": {"city": "Boston"}}}]},
    {"role": "tool", "content": '{"temperature": "72°F", "condition": "Partly cloudy", "humidity": "65%"}'},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, tools=tools,
                                     add_generation_prompt=True, enable_thinking=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=4096, temperature=1.0, top_p=0.95, do_sample=True)
print(tokenizer.decode(output[0][inputs.input_ids.shape[-1]:], skip_special_tokens=False))

Example Output

<think>
The tool returned the weather data for Boston: temperature 72°F, partly cloudy, humidity 65%.
I need to present this information clearly to the user.
</think>
The current weather in Boston is 72°F, partly cloudy, with 65% humidity.<|im_end|>

Multi-Turn Conversations

History Thinking Truncation

python

messages = [
    {"role": "user", "content": "What is 15 * 37?"},
    {"role": "assistant", "content": "<think>\nLet me calculate 15 * 37.\n15 * 37 = 15 * 30 + 15 * 7 = 450 + 105 = 555\n</think>\n15 * 37 = 555"},
    {"role": "user", "content": "Now divide that by 5"},
]

# Default: previous thinking is stripped to save context
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True,
                                     enable_thinking=True, truncate_history_thinking=True)

# To preserve full history:
text_full = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True,
                                          enable_thinking=True, truncate_history_thinking=False)

Parsing Thinking vs. Final Answer

python

import re

def parse_model_output(text):
    """Separate thinking content from final answer."""
    think_match = re.search(r'<think>(.*?)</think>', text, re.DOTALL)
    if think_match:
        thinking = think_match.group(1).strip()
        answer_start = text.find('</think>') + len('</think>')
        answer_end = text.find('<|im_end|>', answer_start)
        answer = text[answer_start:answer_end].strip() if answer_end != -1 else text[answer_start:].strip()
    else:
        thinking, answer = "", text.strip()
    return thinking, answer

thinking, answer = parse_model_output(output_text)

Ethical Considerations and Limitations

Granite 4.2 models are primarily finetuned using instruction-response pairs mostly in English, but also multilingual data covering the supported languages listed above. Although this model handles multilingual dialog, its performance may vary compared to English. Few-shot examples can help in such cases.

While aligned for safety, the model may occasionally produce inaccurate, biased, or unsafe responses. The content within <think>...</think> tags represents internal reasoning and may contain unpolished or intermediate thoughts that do not represent final conclusions.

To enhance safety in deployments, we recommend using Granite 4.2 alongside Granite Guardian to detect and flag risks across key dimensions outlined in the IBM AI Risk Atlas.

Resources

Product Page: https://www.ibm.com/granite
Documentation: https://www.ibm.com/granite/docs/
Learning Resources: https://ibm.biz/granite-learning-resources
Technical Blog: https://huggingface.co/blog/ibm-granite/granite-4-2
GitHub: https://github.com/ibm-granite/granite-4.2-language-models
HF Collection: https://huggingface.co/collections/ibm-granite/granite-42-language-models