137 lines
5.1 KiB
Markdown
137 lines
5.1 KiB
Markdown
---
|
||
license: apache-2.0
|
||
language:
|
||
- en
|
||
pipeline_tag: text-generation
|
||
tags:
|
||
- deepbrainz
|
||
- reasoning
|
||
- mathematics
|
||
- code
|
||
- enterprise
|
||
- 4b
|
||
- long-context
|
||
- 32k
|
||
library_name: transformers
|
||
|
||
---
|
||
|
||
### 🚀 Introducing DeepBrainz-R1 — Reasoning-First Small Language Models for Agentic Systems
|
||
|
||
Today we’re releasing **DeepBrainz-R1**, a family of **reasoning-first Small Language Models (SLMs)** designed for **agentic AI systems in real-world production**.
|
||
|
||
Agentic systems don’t ask once — they reason repeatedly. Tool calls, verification loops, schema-constrained outputs, retries, and long-context planning fundamentally change the economics and reliability requirements of language models. LLM-only stacks struggle under this load.
|
||
|
||
DeepBrainz-R1 is built from the opposite premise:
|
||
|
||
> **Reasoning is a trained behavior, not an emergent side-effect of scale.**
|
||
|
||
#### What DeepBrainz-R1 is designed for
|
||
|
||
* **Repeatable multi-step reasoning**, not one-shot chat
|
||
* **Agent-compatible behavior**: tool use, structured outputs, low-variance reasoning
|
||
* **Production economics**: lower latency, predictable cost, deployability
|
||
* **Inference-time scalability**: compute where needed, not everywhere
|
||
|
||
#### The R1 lineup
|
||
|
||
* **[DeepBrainz-R1-4B](https://huggingface.co/DeepBrainz/DeepBrainz-R1-4B)** — *Flagship production model*
|
||
Best starting point for reliable agentic systems.
|
||
* **[DeepBrainz-R1-2B](https://huggingface.co/DeepBrainz/DeepBrainz-R1-2B)** — *Balanced production model*
|
||
Strong reasoning with lower cost and latency.
|
||
* **[DeepBrainz-R1-0.6B-v2](https://huggingface.co/DeepBrainz/DeepBrainz-R1-0.6B-v2)** — *Canonical small model*
|
||
Cost-efficient baseline for small-model agent workloads.
|
||
* **[Long-context variants (16K / 40K)](https://huggingface.co/collections/DeepBrainz/deepbrainz-r1-reasoning-first-slms-for-agentic-systems)** — early and experimental
|
||
* **[Research checkpoints](https://huggingface.co/collections/DeepBrainz/deepbrainz-r1-research-checkpoints)** — raw artifacts for ablation and evaluation
|
||
* **[Community quantizations (GGUF, low-bit)](https://huggingface.co/collections/DeepBrainz/deepbrainz-r1-community-quantizations-gguf-and-low-bit)** — community-maintained, not officially supported
|
||
|
||
We publish **supported releases, experimental variants, and research checkpoints separately** to keep expectations clear for builders, enterprises, and researchers.
|
||
|
||
#### Why now
|
||
|
||
2026 is the year agentic AI stops being a demo and starts becoming infrastructure. Infrastructure cannot rely on LLM-only economics or LLM-only reliability.
|
||
**Reasoning-first SLMs are the only viable path to scaling agents sustainably.**
|
||
|
||
— **DeepBrainz AI & Labs**
|
||
|
||
---
|
||
|
||
# DeepBrainz-R1-4B
|
||
|
||
**DeepBrainz-R1-4B** is a compact, high-performance reasoning model engineered by **DeepBrainz AI & Labs**. It is part of the **DeepBrainz-R1 Series**, designed to deliver frontier-class reasoning capabilities in cost-effective parameter sizes.
|
||
|
||
This variant offers an extended context window (up to 32,768 tokens), making it suitable for medium-length document and code analysis.
|
||
|
||
---
|
||
|
||
## 🚀 Model Highlights
|
||
|
||
- **Parameter Count:** ~4B
|
||
- **Context Window:** 32,768 tokens
|
||
- **Context Type:** Extended (RoPE)
|
||
- **Specialization:** STEM Reasoning, Logic, Code Analysis
|
||
- **Architecture:** Optimized Dense Transformer
|
||
- **Deployment:** Ready for vLLM, SGLang, and local inference
|
||
|
||
---
|
||
|
||
## 🎯 Intended Use Cases
|
||
|
||
- **Agentic Workflows:** Reliability in multi-step planning tasks.
|
||
- **Math & Science:** Solving complex word problems and equations.
|
||
- **Code Generation:** Writing and debugging algorithms.
|
||
- **Structured Data Extraction:** Parsing and reasoning over unstructured text.
|
||
|
||
> **Note:** This model has undergone post-training to enhance reasoning quality and agentic reliability.
|
||
> It is not optimized for open-ended conversational chat without additional instruction tuning.
|
||
|
||
---
|
||
|
||
## 💻 Usage
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
model_id = "DeepBrainz/DeepBrainz-R1-4B"
|
||
|
||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
model_id,
|
||
torch_dtype="bfloat16",
|
||
device_map="auto"
|
||
)
|
||
|
||
prompt = "Analyze the time complexity of the following algorithm:"
|
||
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
||
|
||
outputs = model.generate(**inputs, max_new_tokens=256)
|
||
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
||
```
|
||
|
||
---
|
||
|
||
## 🏗️ Technical Summary
|
||
|
||
The model has undergone **post-training** to improve reasoning quality, output stability, and robustness under agentic workloads.
|
||
|
||
*Detailed post-training recipes and dataset compositions are not fully disclosed.*
|
||
|
||
---
|
||
|
||
## 🛡️ Limitations & Safety
|
||
|
||
While this model demonstrates strong reasoning capabilities, it may still produce inaccurate information ("hallucinations"). Users should implement appropriate guardrails for production deployments.
|
||
|
||
---
|
||
|
||
## 📜 License
|
||
|
||
This model is released under the **Apache 2.0** license, allowing for academic and commercial use.
|
||
|
||
---
|
||
|
||
<div align="center">
|
||
<b>DeepBrainz AI & Labs</b><br>
|
||
<i>Advancing General Intelligence through Scalable Reasoning</i>
|
||
</div>
|