PulseFlow Tecnologia
← Insights
·7 min read·executive-brief·open-models·deepseek·open-weight
PulseFlow Tecnologia

What DeepSeek, Kimi, and open models changed in AI strategy

In less than two years, the argument that "only whoever has the best closed model wins" became untenable. DeepSeek-R1, Kimi K2, Llama 3.1, and OpenAI's own gpt-oss show a consistent pattern: frontier capability stopped being the exclusive domain of a handful of closed labs.

DeepSeek-R1: parity with o1, MIT license, a fraction of the price

DeepSeek-R1 arrived with "performance on par with OpenAI-o1" on math, code, and reasoning tasks, using large-scale RL in post-training. But the most disruptive detail isn't technical, it's commercial: the model is fully open source under an MIT license, allowing free distillation and commercialization, with API pricing of $0.14 per million cached input tokens - a fraction of what leading closed competitors charged at the time. Smaller distilled versions (32B and 70B) reach parity with o1-mini.

Kimi K2: a trillion parameters, modified MIT

Kimi K2, from Moonshot AI, is a mixture-of-experts (MoE) model with 1 trillion total parameters, but only 32 billion activated at a time, spread across 384 experts with 8 selected per token. It was pretrained on 15.5 trillion tokens "with zero training instability," with a 128K-token context window. On benchmarks, it hits 53.7% on LiveCodeBench and 69.6% on AIME 2024 - competitive with reference closed models - licensed under a modified MIT license, with open weights and code.

Llama 3.1: Meta calls it "the first frontier-level open source model"

Meta described Llama 3.1 405B as "the first frontier-level open source AI model," trained on more than 15 trillion tokens using 16,000 H100 GPUs, evaluated on more than 150 benchmark datasets, competitive with GPT-4, GPT-4o, and Claude 3.5 Sonnet by Meta's own account. The license was modified to let developers use "outputs from Llama models... to improve other models" - a strategic change that encourages an entire ecosystem to build on top of Meta's model. More than 25 partners, including AWS, NVIDIA, Google Cloud, and Databricks, offered services on day one.

gpt-oss: even OpenAI opened its weights

The strongest signal that this pressure worked came from OpenAI itself: gpt-oss-120b and gpt-oss-20b, released under an Apache 2.0 license, with weights freely downloadable on Hugging Face. gpt-oss-120b achieves near-parity with o4-mini on reasoning benchmarks while running on a single 80GB GPU; gpt-oss-20b delivers results similar to o3-mini and runs on edge devices with just 16GB of memory. OpenAI partnered ahead of launch with Azure, Hugging Face, AWS, Databricks, and other platforms to ship with broad support from day one.

What this changes in enterprise strategy

When a Chinese model (DeepSeek), a model from a lab less known in the West (Kimi K2), Meta's largest open model, and even a release from OpenAI itself all converge on the same pattern - frontier capability available as open weights, at a much lower price - the message for anyone deciding architecture is clear: betting everything on a single closed model as a competitive advantage is an increasingly fragile bet. The real differentiator shifts to what a company builds around the model - agent architecture, governance, data integration, observability - because raw model capability is now available to any competitor willing to run an open weight.

Sources