91探花

Deep鈥疌ogito is a startup lab in San鈥疐rancisco. This week it introduced Cogito鈥痸1, a family of open source language models built on Meta鈥檚 Llama base. The smallest packs 3鈥痓illion parameters and the largest holds 70鈥痓illion. All 5 files sit on Hugging鈥疐ace, Ollama, Fireworks鈥疉I and Together鈥疉I for anyone to download or call through an API. The timing aligns with a busy season of AI meet鈥憉ps, giving developers brand鈥憂ew code to test.

The licence lets companies put the models into paid products for up to 700鈥痬illion users each month without extra fees. Deep鈥疌ogito says even larger checkpoints of up to 671鈥痓illion parameters, will follow in the next few months and will stay open source. The cap is like Meta鈥檚 licence terms and guards against huge traffic spikes before extra fees kick in.

The firm鈥檚 chief, Drishan鈥疉rora, once led language model work at Google Search. He calls the Cogito line 鈥渢he strongest open models at their scale,鈥� beating Llama, DeepSeek and Qwen in head to head tests. The claim drew fast attention, with thousands of downloads logged within hours.

听

How Does The New Training Method Raise Brainpower?

听

Deep鈥疌ogito trains its models with a method called Iterated Distillation and Amplification, or IDA. During each loop the model spends extra compute time thinking through a task, finds a better answer, and then stores that thinking inside its own weights.

The team repeats the loop many times. Each cycle gives the model more skill without asking people to judge every answer, which often slows classic reinforcement learning from human feedback. As a result, the code learns quicker and at lower cost. Fewer human hours means small teams can train models that once needed giant budgets.

The firm drew inspiration from AlphaGo鈥檚 self鈥憄lay tactic. Instead of board games, the language model plays with text, builds a stronger reasoning chain, then condenses that chain into a shorter form that fits inside the network. The next time the model sees a similar prompt it reaches the strong answer in one pass.

Deep鈥疌ogito finished the whole line in roughly 75鈥痙ays with a team of fewer than 20 engineers. According to the blog post, the crew ran training on rented cloud GPUs rather than custom hardware, a setup they say slashed cost and sped up iteration. The lab argues that IDA lets a small crew punch above its weight, trimming training time while lifting quality.

听

听

What About The Benchmark Scores?

听

Public test numbers look bright for Cogito鈥痸1. On the MMLU knowledge exam, the three鈥痓illion standard model scores 65.5%. Llama鈥�3鈥痮f the same size scores 58.7%. When the Cogito model flips to reasoning mode, the score goes up to 72.6%.

The 8鈥痓illion standard model posts 80.5% on MMLU and eighty鈥憃ne鈥痯oint鈥痮ne percent on Hellaswag, leaving its Llama rival far behind. In reasoning mode, ARC reaches ninety鈥憈wo percent, while MMLU ticks up to 83.1%. Maths stays the lone weak spot, where DeepSeek鈥疪1 keeps a lead.

Mid鈥憆ange models of 14 and 32鈥痓illion parameters pass Qwen鈥�2.5 in combined tests. The 32鈥痓illion reasoning model crosses 90% on both MMLU and the MATH benchmark, a level few open models touch.

The headline act, Cogito鈥痸1鈥�70 billion, matches or beats the newly launched Llama鈥�4鈥疢oE 109 billion model on many checks. It scores 90.7% on MMLU and holds 92.7% on MGSM. Tool鈥慶alling accuracy tops 92%, while Llama models stay between 35%-54%.

Each Cogito file gives 2 modes in the same weight set, so builders can pick quick answers or deeper thought without swapping models. Early testers praise the 128 000 token context window and strong multilingual support across 30 languages. They also praise the built鈥慽n tool calling, which lets apps pull weather data, run maths or send emails with a single structured prompt.