Instructions to use dusersad12/NimbusLM-ReleaseRepo with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use dusersad12/NimbusLM-ReleaseRepo with Transformers:

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

pipe = pipeline("text-generation", model="dusersad12/NimbusLM-ReleaseRepo")

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("dusersad12/NimbusLM-ReleaseRepo", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use dusersad12/NimbusLM-ReleaseRepo with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "dusersad12/NimbusLM-ReleaseRepo"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dusersad12/NimbusLM-ReleaseRepo",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/dusersad12/NimbusLM-ReleaseRepo

SGLang

How to use dusersad12/NimbusLM-ReleaseRepo 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 "dusersad12/NimbusLM-ReleaseRepo" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dusersad12/NimbusLM-ReleaseRepo",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "dusersad12/NimbusLM-ReleaseRepo" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dusersad12/NimbusLM-ReleaseRepo",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use dusersad12/NimbusLM-ReleaseRepo with Docker Model Runner:
```
docker model run hf.co/dusersad12/NimbusLM-ReleaseRepo
```

NimbusLM

1. Introduction

NimbusLM is the newest open-weight release from Toola Labs, trained from a 7B base with an extended curriculum of verified reasoning traces and tool-use dialogues. This release concentrates on long-horizon reasoning: the chat template carries a scratchpad section by default, and function-calling data was trained with balanced positive and negative examples. Across public and internal evaluations the model closes most of the gap to frontier systems, especially on quantitative and coding suites.

Two practical upgrades ship together with the weights. First, hallucination on factual lookups is measurably lower: on an internal 4,000-question probe, the unsupported-answer rate dropped from 6.1% to 2.4%. Second, the tokenizer inherits a compact 92k vocabulary with clean tool-call delimiters, which keeps prompts short when agents interleave system commands.

2. Evaluation Results

Comprehensive Benchmark Results

	Benchmark	BaselineA	BaselineB	BaselineA-v2	NimbusLM
Core Reasoning Tasks	Math Reasoning	0.502	0.519	0.527	0.550
	Logical Reasoning	0.841	0.859	0.869	0.884
	Common Sense	0.730	0.741	0.752	0.775
Language Understanding	Reading Comprehension	0.697	0.712	0.726	0.745
	Question Answering	0.574	0.589	0.601	0.616
	Text Classification	0.798	0.812	0.823	0.841
	Sentiment Analysis	0.764	0.779	0.790	0.807
Generation Tasks	Code Generation	0.689	0.702	0.715	0.734
	Creative Writing	0.622	0.640	0.653	0.671
	Dialogue Generation	0.578	0.596	0.607	0.625
	Summarization	0.728	0.740	0.758	0.775
Specialized Capabilities	Translation	0.758	0.772	0.784	0.803
	Knowledge Retrieval	0.601	0.617	0.630	0.648
	Instruction Following	0.718	0.733	0.744	0.765
	Safety Evaluation	0.712	0.698	0.726	0.755

Overall Performance Summary

NimbusLM posts consistent gains over every baseline in the suite, with the largest margins concentrated in the reasoning and generation groups.

3. Chat Website & API Platform

A hosted playground and an OpenAI-compatible endpoint are available on the Toola Labs website; bring your own key or sign up for the free tier there.

4. How to Run Locally

The weights load with the standard transformers API, exactly like the base model the series was built on. Two notes for this release:

The scratchpad section is toggled by a chat template flag, no code patch needed.
Tool-call traces should be passed as plain JSON strings inside the scratchpad.

Temperature

A sampling temperature of 0.7 works well for general chat; drop to 0.3 for extractive tasks.

5. License

Weights and code are released under the Apache-2.0 License, including commercial use and distillation.

6. Contact

Bug reports and reproduction questions go to the public issue tracker or to nimbus@toola.ai.

Downloads last month: 23