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

Libraries

How to use dusersad12/HelixMind-ReleaseRepo with Transformers:

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

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

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("dusersad12/HelixMind-ReleaseRepo", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use dusersad12/HelixMind-ReleaseRepo with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "dusersad12/HelixMind-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/HelixMind-ReleaseRepo",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

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

SGLang

How to use dusersad12/HelixMind-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/HelixMind-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/HelixMind-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/HelixMind-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/HelixMind-ReleaseRepo",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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

HelixMind

1. Introduction

HelixMind has just gone through a major version refresh. With this release, HelixMind strengthens its reasoning and inference depth by tapping into far more compute and by bringing in new algorithmic optimization mechanisms during post-training. The model turns in a strong showing across benchmark evaluations spanning mathematics, programming, and general logic, and its overall standing is now closing in on other leading models.

Compared with its predecessor, the refreshed model handles complex reasoning tasks far better. On the AIME 2026 test, for example, accuracy has risen from 68% in the previous release to 84% in this one. The gain traces back to deeper thinking during reasoning: on the AIME set, the earlier model averaged 11K tokens per question, while the new release averages 21K.

Alongside the stronger reasoning, this version also cuts the hallucination rate and widens support for function calling.

2. Evaluation Results

Comprehensive Benchmark Results

	Benchmark	Vesper-7B	Corvid-8B	Vesper-7B-v2	HelixMind
Core Reasoning Tasks	Math Reasoning	0.512	0.531	0.520	0.573
	Logical Reasoning	0.775	0.798	0.806	0.838
	Common Sense	0.721	0.710	0.716	0.750
Language Understanding	Reading Comprehension	0.663	0.681	0.686	0.718
	Question Answering	0.579	0.595	0.598	0.618
	Text Classification	0.795	0.805	0.812	0.838
	Sentiment Analysis	0.771	0.780	0.785	0.800
Generation Tasks	Code Generation	0.610	0.628	0.635	0.674
	Creative Writing	0.581	0.574	0.596	0.636
	Dialogue Generation	0.615	0.630	0.633	0.660
	Summarization	0.738	0.749	0.754	0.779
Specialized Capabilities	Translation	0.772	0.791	0.795	0.811
	Knowledge Retrieval	0.645	0.662	0.664	0.688
	Instruction Following	0.726	0.743	0.745	0.770
	Safety Evaluation	0.710	0.699	0.719	0.750

Overall Performance Summary

HelixMind performs strongly in every evaluated benchmark category, with its most visible headroom-to-leadership gains concentrated in reasoning and generation tasks.

3. Chat Website & API Platform

A chat interface and an API are both available for trying out HelixMind; details are on our official website.

4. How to Run Locally

For more about running HelixMind locally, please consult our code repository.

The usage guidance for HelixMind changes in the following ways compared with earlier releases:

A system prompt is now supported.
Adding special tokens at the start of the output to force the model into a particular thinking mode is no longer required.

HelixMind-Small shares the architecture of its base model and reuses the tokenizer configuration of the main HelixMind release. It runs exactly like its base model.

System Prompt

We suggest using the following system prompt with a concrete date.

You are HelixMind, an AI assistant built by Toola Labs.
Today is {current date}.

For example,

You are HelixMind, an AI assistant built by Toola Labs.
Today is October 17, 2026, Saturday.

Temperature

A temperature of $T_{model}$ = 0.55 is recommended.

Prompts for File Uploading and Web Search

To upload files, build prompts with the template below, where {file_name}, {file_content} and {question} are placeholders.

file_template = \
"""[file name]: {file_name}
[file content begin]
{file_content}
[file content end]
{question}"""

For generation backed by web search, we recommend the template that follows, with {search_results}, {cur_date}, and {question} as placeholders.

search_answer_en_template = \
'''# The search results below relate to the user's message:
{search_results}
Each result arrives as [webpage X begin]...[webpage X end], where X numbers the article. When a sentence leans on the results, cite it inline as [citation:X]; if several results back one sentence, list them all, e.g. [citation:2][citation:6]. Keep citations spread through the answer instead of piling them at the end.
While answering, keep the following in mind:
- Today is {cur_date}.
- Weigh each search result for relevance before you use it; not everything in them is tied to the question.
- For list-style questions, hold the answer to about 10 key points and point readers to the sources for the full picture. Prefer the most complete, relevant items, and avoid adding material the results do not cover.
- For creative work, cite sources inside the body of the text (e.g. [citation:3][citation:5]) rather than only at the end, interpret the user's request, pick a fitting format, and put the search results to full use to produce something insightful, creative, and professional. Stretch the answer as far as it needs to go, point by point and angle by angle.
- When the answer runs long, structure it cleanly and wrap up in paragraphs; if a point list fits better, cap it near 5 points and merge related ones.
- For short factual questions, one or two extra sentences of context are welcome.
- Pick a clean, readable layout that matches the question and the answer.
- Pull together several relevant pages instead of citing one page over and over.
- Unless the user asks otherwise, reply in the language of the question.
# The user's message is:
{question}'''

5. License

This code repository is governed by the Apache 2.0 License, and the HelixMind model weights carry the same Apache 2.0 License. The series permits commercial use and distillation.

6. Contact

For any questions, open an issue on our GitHub repository or write to contact@helixmind.ai.

Downloads last month: 25