dusersad12/MyStellarModel-ProdRepo

MyStellarModel

---

1. Introduction

MyStellarModel is the refreshed open release of our model family. This snapshot was rebuilt on a larger pretraining mix and an extended post-training stage, which deepened its step-by-step reasoning and tightened its instruction following. Across our internal benchmarks it now sits close to several frontier-sized models while staying small enough to run on a single workstation.

Compared with the previous snapshot, the biggest change shows up on hard multi-step problems: on the MATH-500 set, accuracy moved from 66.4% in the prior version to 88.6% here, and the average reasoning budget grew from about 11K tokens per problem to roughly 19K.

The snapshot also ships a lower hallucination rate and more dependable tool / function-calling behavior than its predecessor.

2. Evaluation Results

Comprehensive Benchmark Results

	Benchmark	ModelA	ModelB	ModelA-v2	MyStellarModel
Core Reasoning Tasks	Math Reasoning	0.498	0.527	0.512	0.545
	Logical Reasoning	0.782	0.799	0.791	0.813
	Common Sense	0.704	0.719	0.711	0.732
Language Understanding	Reading Comprehension	0.663	0.681	0.672	0.696
	Question Answering	0.571	0.593	0.582	0.604
	Text Classification	0.796	0.812	0.803	0.825
	Sentiment Analysis	0.761	0.777	0.769	0.790
Generation Tasks	Code Generation	0.612	0.631	0.622	0.645
	Creative Writing	0.573	0.594	0.585	0.604
	Dialogue Generation	0.609	0.628	0.618	0.640
	Summarization	0.733	0.751	0.742	0.764
Specialized Capabilities	Translation	0.772	0.791	0.782	0.803
	Knowledge Retrieval	0.643	0.662	0.653	0.674
	Instruction Following	0.724	0.743	0.734	0.755
	Safety Evaluation	0.705	0.723	0.714	0.736

Overall Performance Summary

MyStellarModel keeps a steady lead across every evaluated category, with its widest margins on the reasoning-heavy and generation-heavy rows.

3. Chat Website & API Platform

A chat playground and a public inference API for MyStellarModel are hosted on our official website; check there for rate limits and the latest endpoints.

4. How to Run Locally

Check the model's source repository for full run instructions. A few things changed versus the older family:

1. A system prompt is now expected at the start of a session.
2. You no longer need to inject a special token at the beginning of the output to force a thinking mode.

The MyStellarModel-Small companion shares the tokenizer with the main release and runs like its base model.

System Prompt

A dated system prompt is recommended:

You are MyStellarModel, a helpful assistant.
Today is {current date}.

For example,

You are MyStellarModel, a helpful assistant.
Today is September 21, 2026, Monday.

Temperature

We recommend setting the temperature $T_{model}$ to 0.55.

Prompts for File Uploading and Web Search

When the user supplies a file, wrap it with this template, filling in {file_name}, {file_content} and {question}:

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

For retrieval-augmented answers, use this template where {search_results}, {cur_date} and {question} are filled in:

search_answer_en_template = \
'''# The search results related to the user's message are below:
{search_results}
Each result above is wrapped as [webpage X begin]...[webpage X end]; X is the result's index. Cite context where relevant with [citation:X]; if a sentence draws on several, list them all, e.g. [citation:3][citation:5]. Spread citations through the answer instead of stacking them at the end.
Notes:
- Today is {cur_date}.
- Filter the results for relevance; not every page matters.
- For list-style questions, cap the answer at ~10 key points and point the user to the sources for the rest.
- For creative writing, cite inline as [citation:3][citation:5] rather than only in a closing block.
- Keep the response well-structured; group related points and merge where possible.
- Prefer the same language as the user's question unless asked otherwise.
# The user's message is:
{question}'''

5. License

The code is released under the MIT License, and the MyStellarModel weights are likewise covered by the MIT License. The family permits commercial use and distillation.

6. Contact

Open an issue on our GitHub repository, or write to contact@stellarmodel.ai.