Focus V1 Pro

Focus V1 Pro is the agentic version of the Focus V1 language model, developed by BRSX Labs.

Focus V1 Pro is designed around a simple idea: instead of forcing a small language model to perform every operation internally, give it access to specialized tools when they are useful.

The current Pro release focuses on two tools:

🔢 Math — performs mathematical calculations.
🔎 Search — retrieves information from a search-result dataset.

The agentic design is intentionally lightweight and optimized for a ~100M parameter model. Focus V1 Pro currently follows a single-tool → single-result interaction pattern rather than attempting complex multi-tool agent loops.

Model Architecture

Focus V1 Pro uses the same core Transformer architecture as the Focus family.

Configuration	Value
Parameters	~100M
Architecture	Decoder-only Transformer
Layers	12
Hidden size	768
Attention heads	12
KV heads	4
Attention	Grouped-Query Attention (GQA)
Feed-forward size	2048
Context length	2048 tokens
Vocabulary	32,000
Positional encoding	RoPE
RoPE θ	10,000
Normalization	RMSNorm
Activation	SwiGLU-style gated MLP
Embeddings	Tied

The attention mechanism uses 12 query heads and 4 key/value heads.

The model uses RMSNorm before both attention and MLP blocks, with residual connections around each component.

The feed-forward block uses a gated formulation based on:

SiLU(W₁x) × W₃x

followed by the output projection.

Input embeddings and the output language-model head are tied.

From Focus V1 to Focus V1 Pro

Focus V1 Base
      │
      ▼
   Instruct
      │
      ▼
Focus V1 Pro
      │
      ├── Math Tool
      │
      └── Search Tool

Focus V1 Pro is trained to preserve normal instruction-following behavior while learning when and how to use external tools.

The goal is not to turn the model into a large autonomous agent. Instead, Pro is designed around controlled tool usage suitable for a small language model.

Agentic Tool Use

Focus V1 Pro currently uses a single-tool, single-result architecture.

User
 │
 ▼
Focus V1 Pro
 │
 ├── Normal response
 │
 └── Tool required
       │
       ▼
     Tool
       │
       ▼
   Tool result
       │
       ▼
  Final response

This design deliberately avoids multi-step tool loops in the current version.

For a ~100M parameter model, keeping the interaction simple reduces the amount of state the model has to manage and allows the model to concentrate on selecting the appropriate tool and interpreting its result.

Math Tool

The Math tool is designed to handle mathematical operations that are better delegated to a deterministic calculator than generated directly by the language model.

For example:

User:
What is 847293 × 92847?

Focus V1 Pro:
→ Math tool

Math tool:
→ exact result

Focus V1 Pro:
→ explains the result

Focus V1 Pro performs well on basic mathematical operations when using the Math tool, including straightforward arithmetic calculations.

The tool is intended to reduce errors caused by attempting exact arithmetic directly inside the language model.

The Math tool does not turn Focus V1 Pro into a general mathematical reasoning system. More advanced mathematical capabilities are planned for future versions.

Search Tool

Focus V1 Pro can use a Search tool when external information is useful.

The current implementation uses search results provided through search.jsonl.

⚠️ Search implementation note

The search results currently available through chat.py are limited to the contents of search.jsonl.

If you want to use a different search backend, live web search, your own index, or another information source, you will need to implement or replace the search infrastructure yourself.

Focus V1 Pro
      │
      ▼
   Search
      │
      ▼
 search.jsonl
      │
      ▼
 search result

The included search system should therefore be considered a demonstration/reference implementation, rather than a general-purpose live web search engine.

What Focus V1 Pro Is Good At

Focus V1 Pro is primarily intended for:

💬 General instruction following
🔢 Basic mathematical calculations with the Math tool
🔎 Information retrieval using the provided Search tool
🧠 Interpreting and presenting tool results
🛠️ Lightweight agentic workflows
⚡ Tool-assisted tasks where a small model benefits from external computation or information

Its strongest intended use case is the combination of small-model inference + specialized external tools.

Limitations

Focus V1 Pro is a relatively small language model (~100M parameters) and should not be expected to match large language models across all tasks.

Known limitations include:

Limited raw reasoning capacity compared with much larger models
Advanced mathematics may require stronger tooling
The current agentic system does not support arbitrary multi-tool loops
Search is limited to the provided search.jsonl data unless the search backend is replaced
Tool availability directly affects tool-assisted performance
The model may still produce incorrect answers when a tool is not used or when external information is insufficient

Successful tool-assisted behavior should not be interpreted as evidence that the underlying model can independently perform the same operation with the same reliability.

Design Philosophy

A small model does not need to do everything itself. It needs to know when a specialized tool can do something better.

Instead of increasing model size solely to improve capabilities such as arithmetic or information retrieval, Focus V1 Pro delegates selected tasks to external tools.

This makes the model's effective capability larger than its parameter count might suggest, while keeping the underlying model relatively compact.

Future Direction

Future versions of Focus are expected to expand the agentic system with additional capabilities, including:

🐍 Code execution
🧪 Syntax checking
🔧 Automatic error detection and correction
🧮 More comprehensive mathematical tooling
🔎 More advanced keyword-based search
🔁 More sophisticated agentic workflows
🛠️ Additional specialized tools

These capabilities are planned for future versions and are not part of the current Focus V1 Pro architecture.

Intended Use

Focus V1 Pro is intended for:

Research
Educational experimentation
Local inference
Lightweight AI applications
Agentic/tool-use experiments
Building applications around a compact language model

It is particularly suitable for developers interested in exploring how small language models can be augmented with specialized tools.

Notable Difference From Focus V1

Focus V1

Language model + instruction following

Focus V1 Pro

Language model + instruction following + controlled tool use

The Pro version does not simply increase the model's size. It changes how the model interacts with external computation and information.

License

Apache License 2.0

BRSX Labs

Focus is developed as part of the BRSX Labs ecosystem.

Focus V1 Pro is an experiment in making compact language models more capable through specialized tools rather than parameter count alone.

Downloads last month: 43

Model tree for brsx-labs/Focus-v1-pro

Base model

brsx-labs/Focus-v1-base

Finetuned

(1)

this model