Instructions to use aisquared/bolt-instruct-v2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use aisquared/bolt-instruct-v2 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="aisquared/bolt-instruct-v2")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("aisquared/bolt-instruct-v2")
model = AutoModelForMultimodalLM.from_pretrained("aisquared/bolt-instruct-v2", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use aisquared/bolt-instruct-v2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "aisquared/bolt-instruct-v2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aisquared/bolt-instruct-v2",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/aisquared/bolt-instruct-v2

SGLang

How to use aisquared/bolt-instruct-v2 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "aisquared/bolt-instruct-v2" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aisquared/bolt-instruct-v2",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "aisquared/bolt-instruct-v2" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aisquared/bolt-instruct-v2",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use aisquared/bolt-instruct-v2 with Docker Model Runner:
```
docker model run hf.co/aisquared/bolt-instruct-v2
```

bolt-instruct-v2 / README.md

PrajwalN-AIS

Update README.md

cee32a6 verified 5 days ago

preview code

Raw History Blame Contribute Delete

7.6 kB

---
tags:
- text-generation
- causal-lm
- instruction-tuning
- chat
- rag
- code-generation
- summarization
- extraction
- synthetic-data
- generated_from_trainer
- conversational
license: other
pipeline_tag: image-text-to-text
library_name: transformers
language:
- en
---
# Bolt Instruct V2

Bolt Instruct V2 is a fine-tune of [meta-models/Muse-Glimmer-30B](https://huggingface.co/meta-models/Muse-Glimmer-30B/blob/main/README.md) designed optimized for **concise reasoning, accurate tool selection, and efficient execution of tool-based workflows**.  

---

# Model Overview

Bolt Instruct V2 provide **strong instruction-following capabilities** across **multi-turn agentic workflows** while reducing cost and latency as compared to other leading ~30B models.

Key design goals:

- Strong instruction adherence
- Multi-turn agentic coding & automated task completion
- Concise reasoning

---

# Model Details

- **Type:** Dense Causal Transformer w/ merged LoRA adapter
- **Max context:** 131,072+
- **Supported modalities:** Text, Image 

### Capabilities

- Chat / multi-turn dialogue  
- Instruction following  
- Structured output (JSON)  
- Summarization & transformation  
- Extraction  
- RAG generation  
- Code generation  

---

# Training


### Training Approach

Bolt Instruct V2 is a LoRA supervised fine-tune, trained to combine concise reasoning with precise tool use. The goal was to reduce unnecessary deliberation while preserving the ability to make tool calls. We chose supervised fine-tuning because we could demonstrate the behaviour we wanted directly: brief reasoning followed by the appropriate tool calls, including grouped calls for parallel execution.

Training Configuration

| | |
|---|---|
| Method | LoRA supervised fine-tuning (Unsloth + TRL SFTTrainer) |
| Base model | unsloth/Muse-Glimmer-30B |
| LoRA rank / alpha | 32 / 32 |
| Trained modules | language-layer attention and MLP (vision tower frozen) |
| Loss | response-only (assistant turns masked in) |
| Optimizer | adamw_8bit, linear schedule |
| Quantization | none (full-precision base) |
### Training Data
The training corpus contains 7,646 multi-step agentic sessions, filtered to include tool schemas, at least one reasoning block, and a genuine user request. Reasoning before tool calls had a median length of 28 tokens, and 91% of tool calls were preceded by reasoning. The corpus also included 14,204 turns with multiple tool calls, representing 17.4% of tool-calling turns and occurring in 47% of sessions. Sessions contained a median of 13 tool calls, with 62.7% containing at least 10. Together, these examples demonstrated concise reasoning, parallel tool calling, and sustained multi-step tool use.
| | |
|---|---|
| Sessions | 7,646 (7,263 train / 383 held out) |
| Messages per session (median) | 27 |
| Assistant turns per session (median) | 11 |
| Distinct tool schemas | 26 |
| **Tool use** | |
| Total tool calls | 106,219 |
| Tool-calling turns | 81,570 |
| Tool calls per session (median) | 13 |
| Sessions with ≥10 tool calls | 62.7% |
| **Parallel tool calling** | |
| Turns issuing multiple tool calls | 14,204 (17.4% of tool-calling turns) |
| Sessions containing a parallel turn | 47.1% |
| Calls per parallel turn (median) | 2 |
| **Reasoning** | |
| Tool calls preceded by reasoning | 90.6% |
| Reasoning length (median) | 28 tokens |
| Reasoning length (mean) | 48 tokens |
### Training Hardware
| | |
|---|---|
| GPUs | 2 × NVIDIA H200 |
| GPU memory | 141 GB each, 282 GB total |
| Training time | 18 hours |
---
# Intended Use
- Chat assistants  
- Enterprise copilots  
- RAG pipelines  
- Document processing  
- Structured extraction  
- Code assistance  
---
# Usage
```python
from transformers import AutoProcessor, AutoModelForMultimodalLM
processor = AutoProcessor.from_pretrained("aisquared/bolt-instruct-v2")
model = AutoModelForMultimodalLM.from_pretrained("aisquared/bolt-instruct-v2", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))
```
---
# Evaluation
We evaluated Bolt Instruct V2 on raw performance on general-purpose and agentic tasks, as well as reasoning efficiency and time to complete a given task.
### Evaluation results for aisquared/bolt-instruct-v2:
Measured with the sampling configuration published for Muse Glimmer 
**temperature 1.0 / top_p 0.95 / top_k 64**, high reasoning strength 
served with vLLM (`--generation-config auto`).
![efficiency_donuts](https://cdn-uploads.huggingface.co/production/uploads/6271334067ff0a433db23501/-U2ok6mZQ8si_7Ncx7gFz.png)
### Efficiency — per task
| Metric | Benchmark | Muse Glimmer 30B(High) | **Bolt Instruct V2** | Gemma 4 31B (think on) |
|---|---|---|---|---|
| **Reasoning tokens** | BFCL | 1,266 | **635** | 2,609 |
| | τ²-bench | 1,689 | **976** | 3,450 |
| | AutomationBench | 6,351 | **2,022** | 10,835 |
| **Output tokens** | BFCL | 459 | **281** | 313 |
| | τ²-bench | 2,514 | **1,931** | 4,177 |
| | AutomationBench | 9,910 | **5,988** | 14,839 |
| **Latency** | BFCL (mean) | 19.76 s | **12.60 s** | 7.47 s |
| | τ²-bench | 64.9 s | **51.4 s** | 129.8 s |

### Agentic — multi-turn

| Benchmark | Muse Glimmer 30B(High) | **Bolt Instruct V2** | Gemma 4 31B (think on) |
|---|---|---|---|
| τ²-bench @ temperature 0 | 85.37 | 76.83 | 84.15 |
| τ²-bench @ published sampling ¹ | 34.45 | 30.64 | 85.21 |

### Agentic — single-turn tool calling

| Benchmark | Muse Glimmer 30B(High) | **Bolt Instruct V2** | Gemma 4 31B (think on) |
|---|---|---|---|
| BFCL Live | 55.84% | **59.67%** | 69.21% |
| BFCL Non-Live | 10.12% | **24.38%** | — |
| ACEBench parallel function calling | 48% | **79%** | 81% |

### General capability

| Benchmark | Muse Glimmer 30B(High) | **Bolt Instruct V2** | Bolt Instruct V1 |
|---|---|---|---|
| BBH | 63.00 | **89.00** | — |
| GPQA Diamond (CoT, 0-shot) ³ | 33.6 | **57.5** | — |
| IFEval (prompt-strict) | 65.06 | **73.75** | 85.95 |
| ARC Challenge | 58.28 | 56.91 | 56.66 |
| TruthfulQA MC2 | 61.64 | 59.32 | 69.34 |

Bolt Instruct V1 and V2 are not successive versions of the same model.
V1 is built on Olmo 3 (32B, 64K context, text-only) and optimized for
instruction following; V2 is built on Muse Glimmer (30B, 128K context,
multimodal) and optimized for agentic tool execution. Each should be read
against its own foundation model, But Bolt Instruct V2 improves on Muse Glimmer's IFEval
score.


**Reproduction.** τ²-bench and AutomationBench under their official
harnesses; lm-evaluation-harness for BBH, GPQA Diamond, IFEval, ARC
Challenge and TruthfulQA; BFCL and ACEBench under their own.

---

# Limitations

- May hallucinate without grounding
- Not suitable for high-risk domains without oversight
- Requires low-temperature decoding for precise tool-call arguments
- Performance varies by task domain; validate against your own task set
- Built for short and medium-horizon workflows and regression is seen on long multi-step workflows
---

# License

Bolt Instruct is released under the [AI Squared Community License](https://docs.squared.ai/terms-of-use).

	---
	tags:
	- text-generation
	- causal-lm
	- instruction-tuning
	- chat
	- rag
	- code-generation
	- summarization
	- extraction
	- synthetic-data
	- generated_from_trainer
	- conversational
	license: other
	pipeline_tag: image-text-to-text
	library_name: transformers
	language:
	- en
	---

	# Bolt Instruct V2

	Bolt Instruct V2 is a fine-tune of [meta-models/Muse-Glimmer-30B](https://huggingface.co/meta-models/Muse-Glimmer-30B/blob/main/README.md) designed optimized for concise reasoning, accurate tool selection, and efficient execution of tool-based workflows.

	---

	# Model Overview

	Bolt Instruct V2 provide strong instruction-following capabilities across multi-turn agentic workflows while reducing cost and latency as compared to other leading ~30B models.

	Key design goals:

	- Strong instruction adherence
	- Multi-turn agentic coding & automated task completion
	- Concise reasoning

	---

	# Model Details

	- Type: Dense Causal Transformer w/ merged LoRA adapter
	- Max context: 131,072+
	- Supported modalities: Text, Image

	### Capabilities

	- Chat / multi-turn dialogue
	- Instruction following
	- Structured output (JSON)
	- Summarization & transformation
	- Extraction
	- RAG generation
	- Code generation

	---

	# Training


	### Training Approach

	Bolt Instruct V2 is a LoRA supervised fine-tune, trained to combine concise reasoning with precise tool use. The goal was to reduce unnecessary deliberation while preserving the ability to make tool calls. We chose supervised fine-tuning because we could demonstrate the behaviour we wanted directly: brief reasoning followed by the appropriate tool calls, including grouped calls for parallel execution.

	Training Configuration

	\| \| \|
	\|---\|---\|
	\| Method \| LoRA supervised fine-tuning (Unsloth + TRL SFTTrainer) \|
	\| Base model \| unsloth/Muse-Glimmer-30B \|
	\| LoRA rank / alpha \| 32 / 32 \|
	\| Trained modules \| language-layer attention and MLP (vision tower frozen) \|
	\| Loss \| response-only (assistant turns masked in) \|
	\| Optimizer \| adamw_8bit, linear schedule \|
	\| Quantization \| none (full-precision base) \|

	### Training Data
	The training corpus contains 7,646 multi-step agentic sessions, filtered to include tool schemas, at least one reasoning block, and a genuine user request. Reasoning before tool calls had a median length of 28 tokens, and 91% of tool calls were preceded by reasoning. The corpus also included 14,204 turns with multiple tool calls, representing 17.4% of tool-calling turns and occurring in 47% of sessions. Sessions contained a median of 13 tool calls, with 62.7% containing at least 10. Together, these examples demonstrated concise reasoning, parallel tool calling, and sustained multi-step tool use.

	\| \| \|
	\|---\|---\|
	\| Sessions \| 7,646 (7,263 train / 383 held out) \|
	\| Messages per session (median) \| 27 \|
	\| Assistant turns per session (median) \| 11 \|
	\| Distinct tool schemas \| 26 \|
	\| Tool use \| \|
	\| Total tool calls \| 106,219 \|
	\| Tool-calling turns \| 81,570 \|
	\| Tool calls per session (median) \| 13 \|
	\| Sessions with ≥10 tool calls \| 62.7% \|
	\| Parallel tool calling \| \|
	\| Turns issuing multiple tool calls \| 14,204 (17.4% of tool-calling turns) \|
	\| Sessions containing a parallel turn \| 47.1% \|
	\| Calls per parallel turn (median) \| 2 \|
	\| Reasoning \| \|
	\| Tool calls preceded by reasoning \| 90.6% \|
	\| Reasoning length (median) \| 28 tokens \|
	\| Reasoning length (mean) \| 48 tokens \|
	### Training Hardware


	\| \| \|
	\|---\|---\|
	\| GPUs \| 2 × NVIDIA H200 \|
	\| GPU memory \| 141 GB each, 282 GB total \|
	\| Training time \| 18 hours \|

	---

	# Intended Use

	- Chat assistants
	- Enterprise copilots
	- RAG pipelines
	- Document processing
	- Structured extraction
	- Code assistance

	---

	# Usage

	```python
	from transformers import AutoProcessor, AutoModelForMultimodalLM

	processor = AutoProcessor.from_pretrained("aisquared/bolt-instruct-v2")
	model = AutoModelForMultimodalLM.from_pretrained("aisquared/bolt-instruct-v2", device_map="auto")
	messages = [
	{
	"role": "user",
	"content": [
	{"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
	{"type": "text", "text": "What animal is on the candy?"}
	]
	},
	]
	inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
	).to(model.device)

	outputs = model.generate(**inputs, max_new_tokens=256)
	print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))
	```

	---

	# Evaluation

	We evaluated Bolt Instruct V2 on raw performance on general-purpose and agentic tasks, as well as reasoning efficiency and time to complete a given task.

	### Evaluation results for aisquared/bolt-instruct-v2:

	Measured with the sampling configuration published for Muse Glimmer
	temperature 1.0 / top_p 0.95 / top_k 64, high reasoning strength
	served with vLLM (`--generation-config auto`).

	![efficiency_donuts](https://cdn-uploads.huggingface.co/production/uploads/6271334067ff0a433db23501/-U2ok6mZQ8si_7Ncx7gFz.png)


	### Efficiency — per task

	\| Metric \| Benchmark \| Muse Glimmer 30B(High) \| Bolt Instruct V2 \| Gemma 4 31B (think on) \|
	\|---\|---\|---\|---\|---\|
	\| Reasoning tokens \| BFCL \| 1,266 \| 635 \| 2,609 \|
	\| \| τ²-bench \| 1,689 \| 976 \| 3,450 \|
	\| \| AutomationBench \| 6,351 \| 2,022 \| 10,835 \|
	\| Output tokens \| BFCL \| 459 \| 281 \| 313 \|
	\| \| τ²-bench \| 2,514 \| 1,931 \| 4,177 \|
	\| \| AutomationBench \| 9,910 \| 5,988 \| 14,839 \|
	\| Latency \| BFCL (mean) \| 19.76 s \| 12.60 s \| 7.47 s \|
	\| \| τ²-bench \| 64.9 s \| 51.4 s \| 129.8 s \|

	### Agentic — multi-turn

	\| Benchmark \| Muse Glimmer 30B(High) \| Bolt Instruct V2 \| Gemma 4 31B (think on) \|
	\|---\|---\|---\|---\|
	\| τ²-bench @ temperature 0 \| 85.37 \| 76.83 \| 84.15 \|
	\| τ²-bench @ published sampling ¹ \| 34.45 \| 30.64 \| 85.21 \|

	### Agentic — single-turn tool calling

	\| Benchmark \| Muse Glimmer 30B(High) \| Bolt Instruct V2 \| Gemma 4 31B (think on) \|
	\|---\|---\|---\|---\|
	\| BFCL Live \| 55.84% \| 59.67% \| 69.21% \|
	\| BFCL Non-Live \| 10.12% \| 24.38% \| — \|
	\| ACEBench parallel function calling \| 48% \| 79% \| 81% \|

	### General capability

	\| Benchmark \| Muse Glimmer 30B(High) \| Bolt Instruct V2 \| Bolt Instruct V1 \|
	\|---\|---\|---\|---\|
	\| BBH \| 63.00 \| 89.00 \| — \|
	\| GPQA Diamond (CoT, 0-shot) ³ \| 33.6 \| 57.5 \| — \|
	\| IFEval (prompt-strict) \| 65.06 \| 73.75 \| 85.95 \|
	\| ARC Challenge \| 58.28 \| 56.91 \| 56.66 \|
	\| TruthfulQA MC2 \| 61.64 \| 59.32 \| 69.34 \|

	Bolt Instruct V1 and V2 are not successive versions of the same model.
	V1 is built on Olmo 3 (32B, 64K context, text-only) and optimized for
	instruction following; V2 is built on Muse Glimmer (30B, 128K context,
	multimodal) and optimized for agentic tool execution. Each should be read
	against its own foundation model, But Bolt Instruct V2 improves on Muse Glimmer's IFEval
	score.


	Reproduction. τ²-bench and AutomationBench under their official
	harnesses; lm-evaluation-harness for BBH, GPQA Diamond, IFEval, ARC
	Challenge and TruthfulQA; BFCL and ACEBench under their own.

	---

	# Limitations

	- May hallucinate without grounding
	- Not suitable for high-risk domains without oversight
	- Requires low-temperature decoding for precise tool-call arguments
	- Performance varies by task domain; validate against your own task set
	- Built for short and medium-horizon workflows and regression is seen on long multi-step workflows
	---

	# License

	Bolt Instruct is released under the [AI Squared Community License](https://docs.squared.ai/terms-of-use).