New FREE & Open Reasoning LLM Matches Open AI o1! + RTX 5090 Unboxing! AI News

DeepSeek R1 is landing as a fully open-source reasoning model that performs essentially on par with OpenAI’s o1—while also undercutting it on accessibility and cost. Benchmarks cited for DeepSeek R1 show it matching or slightly exceeding o1 on AIM 2024 (about 79–80% accuracy, with DeepSeek R1 at 79.8% and o1 at roughly the same level). The smaller DeepSeek R1 32B variant also holds up strongly: it posts 72.6% on AIM 2024, far ahead of OpenAI’s o1-mini at 63.6%, and dramatically above DeepSeek’s older V3 model at 39.2%. The pattern repeats across other reasoning-heavy tests—DeepSeek R1 is described as “closing the gap” with o1 on code-focused evaluations (96.3% for R1 vs 96.6% for o1), while the 32B version still reaches about 90%.

Not every benchmark is a tie. On GPQA, OpenAI’s o1 retains a modest edge (75.7% vs DeepSeek R1 at 71.5%), though the margin is framed as relatively small. Math results look especially strong for DeepSeek R1, with 97.3% accuracy—just 1% behind o1—while the 32B model sits close at 94%. On SWE-bench Verified, DeepSeek R1 is presented as nearly matching o1 again, with the 32B variant lagging behind both o1 and DeepSeek V3 and o1-mini on that particular measure. Even so, the overall takeaway is that DeepSeek’s reasoning stack has improved sharply compared with its own prior generation, and that the 32B model delivers near state-of-the-art performance for a fraction of the size.

Beyond the headline model, DeepSeek is also releasing distilled smaller models “stemming from DeepSeek R1,” including versions as small as 1.5B parameters. Those tiny models are claimed to outperform GPT-4o on some benchmarks, with larger distilled variants (7B, 14B, 32B, and even a 70B) pushing further. Pricing is another major point: the transcript cites extremely low token costs—14 cents per million input tokens and 55 cents per million input tokens, with output around $29 per million—positioning the system as practical for builders who want to run reasoning at scale.

The rest of the AI news shifts from model performance to infrastructure and generation workflows. A new open-source training framework, Yandex FSDP (fully shared data parallel), is pitched as a way to speed large language model training by reducing GPU communication bottlenecks, cutting memory usage, and saving resources across up to 150 GPUs. For local deployment, Nvidia’s RTX 5090 is mentioned via an upcoming unboxing, with 32GB VRAM highlighted as enough to run local models and potentially test DeepSeek R1 32B and even 70B-class models.

Finally, the transcript emphasizes a clear trend in AI video: consistency. Early-access “elements” features from Kling AI are shown combining multiple input images (including a person, a room, and a Red Bull can) into coherent scenes with consistent characters and backgrounds. Other tools and workflows—like LoRA fine-tuning for OneAI video generation—are framed as the path toward stable characters across clips, with examples such as a John Wick-style character that’s difficult to distinguish from real footage. The segment closes with additional research and product updates in relighting, lightweight speech generation, and interactive image editing, all aimed at making AI outputs more controllable and usable in real workflows.