Deep Cogito launches open-source hybrid reasoning models

Cogito v1 is a new suite of open-source language models built to bring o-series and DeepSeek-style "thinking" into everyday weights you can run yourself.

The lineup spans five parameter sizes: 3B, 8B, 14B, 32B, and 70B. What makes them unique is their hybrid architecture. You can query them for near-instant answers on simple tasks, or flip them into reasoning mode to let them pause, self-reflect, and work through complex problems before responding.

How it works

Instead of just distilling thought traces from a massive teacher model, Deep Cogito leans on a training loop called Iterated Distillation and Amplification (IDA).

The basic idea is straightforward. The model is given extra compute at generation time to explore deeper solution paths. Once it figures out a stronger answer, that improved reasoning trace is distilled back into the model's own parameters. The team compares this self-reinforcing feedback loop to AlphaGo's self-play, but tailored for natural language.

Under the hood, the models are fine-tuned from Meta's Llama family and Alibaba's Qwen2.5. They ship under the standard Llama community license, meaning you can deploy them commercially for free as long as your service stays under 700 million monthly active users.

Benchmark gains and the math trade-off

Across standard benchmarks, the self-reflection mode delivers measurable bumps, especially on smaller footprints. The tiny Cogito 3B jumps from 65.4% to 72.6% on MMLU in reasoning mode, outpacing the base Llama 3.2 3B by noticeable margins.

The larger checkpoints tell a similar story. Cogito 70B hits 91.7% on MMLU in standard mode, beating Llama 3.3 70B at 85.3% and edging out the 109B Llama 4 Scout on aggregate scores. The reasoning variant also posts 92.7% on MGSM multilingual math.

Native tool calling is the other big addition. Base Llama 3.2 3B lacks tool calling out of the box, whereas Cogito 3B handles parallel and multiple tool calls natively, scoring over 91% on multi-call tasks. Cogito 8B clears 89% across all tool-use benchmarks, leaving Llama 3.1 8B far behind.

The catch is pure mathematics. While Cogito trades blows on general knowledge, it lags behind DeepSeek R1. On the competitive MATH benchmark, Cogito 8B scores 60.2% compared to 80.6% from DeepSeek R1 Distill 8B. Even the 70B reasoning model trails R1 by several points (83.3% versus 89.0%).

How to get started

All five base checkpoints are live right now on Hugging Face and Ollama. If you have an Apple Silicon Mac or an Nvidia GPU, pulling down the 3B or 8B weights via Ollama takes just a single command.

For production setups or larger models, hosted APIs are available through Fireworks and Together AI. The team says MoE models reaching up to 671B parameters will follow in the coming months.

Why it matters

If you are building lightweight local agents that need reliable tool calling alongside basic chain-of-thought verification, Cogito v1 is well worth a spin. The 3B and 8B versions punch far above their weight for edge workflows.

Just be mindful of what you use it for. It is not currently unseating DeepSeek R1 on heavyweight math benchmarks. But by proving that smaller open models can amplify their own reasoning without relying purely on external teacher models, Deep Cogito opens up a very practical lane for self-improving open AI.