Short answer: Deep Cogito v2 is a family of open-weight reasoning models from the US-based startup Deep Cogito. The first four models (70B, 109B, 405B and 671B) arrived as a preview on 31 July 2025. The flagship, Cogito v2.1 671B, followed on 19 November 2025. Each model can answer directly or reason first before responding. The idea is to train the models to reach good answers with fewer reasoning tokens, so they cost less to run. You can try them for free in a web chat, call them through several APIs, or download the weights from Hugging Face.
Deep Cogito v2 at a glance
Feature | v2 Preview | v2.1 |
Released | 31 July 2025 | 19 November 2025 |
Sizes | 70B dense, 109B MoE, 405B dense, 671B MoE | 671B MoE only |
Base models | Llama 3 70B, Llama 4 Scout, Llama 3 405B, and a DeepSeek model | Fork of a November 2024 DeepSeek base |
Modes | Answer directly or reason first | Same |
Context length | Not stated by Cogito | 128K, as listed by Ollama |
Licence | Llama-based models follow Meta’s Llama terms; check each model card | MIT, as listed by Ollama |
Weights | Hugging Face | Hugging Face |
Try it | chat.deepcogito.com (free), APIs, local | Same |
What is Deep Cogito v2?
Deep Cogito is a US-based AI company. Its v2 series is a set of “hybrid reasoning” models. Each one works as a normal chat model, or it can reflect for a while before it replies. Other reasoning models use the same split.
The training is what sets it apart. Most reasoning models improve by thinking longer at answer time. Cogito’s pitch is the opposite: it tries to teach the model which lines of thought tend to work, so less thinking is needed later. If you’re comparing options, our guide to the best AI reasoning models in 2026 shows where open models like this one fit, and our DeepSeek V4 guide covers DeepSeek’s newer model line, which Cogito’s 671B models build on in an earlier form.
How does Cogito train its models?
Cogito describes two ideas. These two ideas are:
Iterated Distillation and Amplification (IDA): the model reasons through a problem, the system keeps the reasoning paths that worked, and those get trained back into the model’s weights. Repeat the loop, and the model builds a kind of “intuition” about where to look first. We explain this in more depth in our post on Deep Cogito and IDA-powered open LLMs.
Process supervision: for v2.1, Cogito says its models are trained with feedback on the reasoning steps themselves, not only on the final answer (Cogito v2.1 post). In plain terms, it rewards good thinking, not just correct answers.
The company also reports that it spent under $3.5 million to train all eight Cogito models (v1 and v2 preview, 3B to 671B), including data generation and over a thousand experiments (Cogito v2 Preview post).
Cogito v2 models and sizes
Model | Size | Built on | Best for |
Cogito v2 70B | Dense | Llama 3 70B | Smallest option; easiest to host |
Cogito v2 109B | MoE | Llama 4 Scout | Mid-size reasoning at lower compute |
Cogito v2 405B | Dense | Llama 3 405B | Large dense model |
Cogito v2 671B | MoE | DeepSeek (Unsloth lists R1) | Preview flagship |
Cogito v2.1 671B | MoE | DeepSeek base fork | Flagship |
The base-model details come from Unsloth’s guide. On Hugging Face, the 109B model is the only one tagged for image-and-text input.
Cogito v2.1: what changed
Cogito v2.1 is a single 671B mixture-of-experts model. Deep Cogito calls it the best open-weight LLM from a US company, which is the company’s own claim. The company says the model improves instruction following, coding, long queries, multi-turn chats and creativity.
Token efficiency is the main selling point. The Decoder reported on a chart showing Cogito v2.1 using the fewest tokens per benchmark task, 4,894 on average, against 9,178 for Gemini 2.5 Pro. The same article presumes the base model is DeepSeek R1-Lite, but that is The Decoder’s speculation, not something Cogito confirms. The company’s own benchmark charts don’t print exact scores in text, so check the charts yourself.
Cogito v2 benchmarks and token efficiency
For the preview release, Deep Cogito reports that the 671B MoE:
matches or beats DeepSeek R1 and V3 on its tests, and matches the later R1 0528 in reasoning mode;
beats DeepSeek V3 in non-reasoning mode;
uses about 60% shorter reasoning chains than DeepSeek R1 0528.
The company also says the model trails the best closed models, such as OpenAI’s o3 and Claude 4 Opus, by a small margin. Cogito itself warns that its benchmarks may not reflect its internal evaluations or the newest closed models. Run your own prompts before you trust any of these numbers. For a closed-model comparison, see our Claude Opus 5 review.
Can Cogito reason about images?
The models are trained on text, but Deep Cogito showed them comparing photos of a duck and a lion when thinking mode was on. The company puts this down to transfer from the multimodal base model and says it hasn’t tested the models on visual benchmarks. If you need image work, don’t count on it yet.
How to run Cogito v2
Free web chat
The quickest way to try it is chat.deepcogito.com. Deep Cogito says the chat doesn’t store your conversations.
APIs
Cogito v2.1 is available through OpenRouter, Fireworks AI, Together AI, Ollama’s cloud, Baseten and RunPod (Cogito v2.1 post. On Together AI, the price is $1.25 per million input tokens and $1.25 per million output tokens at the time of writing. Prices differ by provider, so check the provider’s page before you commit.
Locally
You can run Cogito with Ollama (ollama run cogito-2.1, which needs Ollama 0.13.0 or later) or with Unsloth’s GGUF builds. Unsloth recommends having combined RAM and VRAM close to the size of the file you download. With less memory, the model still runs, but more slowly.
Model | Recommended quant | Download size |
70B dense | UD-Q4_K_XL | 44 GB |
109B MoE | UD-Q3_K_XL | 50 GB |
405B dense | UD-Q2_K_XL | 152 GB |
671B MoE | UD-Q2_K_XL | 251 GB |
For the 671B model, there’s also a 1.78-bit build of about 168 GB. At full precision, it needs data-centre hardware: about 8 B200 GPUs or 16 H200s for BF16, according to llmrun’s hardware notes. For most people, the 70B or 109B model is the realistic local choice.
Want to put a self-hosted model into a real product or workflow? Our AI consulting team can help you pick the model and plan the infrastructure.
How to turn on thinking mode?
In the chat template, set enable_thinking=True to make the model reason before it answers, as shown in the Cogito v2 Preview post. Leave it off for fast, direct replies. The exact syntax can differ by tool, so check the model card for your setup.
Cogito v2 vs DeepSeek R1 and closed models
Feature | Cogito v2.1 671B | DeepSeek R1 | Closed models (o3, Claude Opus) |
Weights | Open-weight | Open-weight | Not available |
Starting point | Fork of a DeepSeek base | Original | Original |
Reasoning style | Hybrid, trained for fewer tokens | Long reasoning chains | Varies; usually the strongest |
Where to run | Self-host or API | Self-host or API | Vendor API only |
Licence | MIT, per Ollama’s listing | MIT | Proprietary |
Who should use Cogito?
Use it if you want an open-weight reasoning model, care about lower reasoning-token costs, or want to self-host or fine-tune.
Skip it if you need the strongest possible reasoning on hard problems (closed models still lead), reliable vision, or a clearly stated licence for every model size.
Limitations
Most benchmark figures are company-reported and shown as charts.
The 671B models need serious hardware to run well.
Cogito’s own pages don’t state a licence. The MIT licence for v2.1 comes from Ollama’s listing, and the Llama-based preview models carry Meta’s licence terms.
Image reasoning is untested.
v2.1 exists only at 671B, so there’s no small v2.1.
Conclusion
Deep Cogito v2 is a serious open-weight option if you care about reasoning cost. The bet is simple: train the model to think smarter, not longer. v2.1 671B is the one to look at now. The claims are mostly the company’s own, though, so test it on your own work first.
Start with the free chat, then move to an API or a quantised local build once it earns a place in your stack. For more on open models and reasoning, best AI reasoning models in 2026, DeepSeek V4 articles, or you can also talk to our AI consulting team.
FAQs
What is Deep Cogito v2?
It’s a family of open-weight hybrid reasoning models from Deep Cogito. The preview came out in July 2025 in four sizes, and v2.1 671B followed in November 2025.
Is Deep Cogito v2 open-weight or open source?
It is open-weight. The weights are public on Hugging Face, but the training data and code are not open. Ollama lists v2.1 under the MIT licence. The Llama-based preview models carry Meta’s licence terms, so check each model’s page.
What is Cogito v2.1?
Cogito v2.1 is the November 2025 update: a single 671B mixture-of-experts model that Deep Cogito says uses fewer reasoning tokens than similar models.
How do I try Cogito for free?
Use the web chat at chat.deepcogito.com. You can also run smaller models locally with Ollama or Unsloth.
Can I run Cogito locally?
Yes, with quantised builds. The 70B model needs about 44 GB, while the 671B model needs about 168 GB at minimum.
What is IDA in Cogito?
Iterated Distillation and Amplification: the model’s best reasoning paths are trained back into its weights, so it needs less thinking at answer time.
Is Cogito better than DeepSeek R1?
Deep Cogito reports that it matches or beats R1 with shorter reasoning chains. Those are the company’s own tests, so try both on your own tasks.
Does Cogito support images?
Not by design. The models show some image reasoning, but it hasn’t been tested on visual benchmarks.