On January 20, 2025, Chinese AI lab DeepSeek released DeepSeek-R1, an open-weight reasoning model under the MIT license that the company says performs on par with OpenAI's o1 on math, coding and reasoning tasks. The launch also included six smaller distilled models and an open training recipe, and within a week it had triggered a record sell-off in AI stocks. For developers, DeepSeek-R1 turned reasoning models from a premium, closed API feature into something you can download, inspect and run on your own hardware.

Key Facts#

  • Release: January 20, 2025, by Hangzhou-based DeepSeek, as DeepSeek-R1 and the experimental DeepSeek-R1-Zero.
  • Architecture: a mixture-of-experts (MoE) model with 671B total parameters, 37B activated per token and a 128K context window, built on the DeepSeek-V3 base model.
  • License: MIT for both code and weights, explicitly allowing commercial use, modification and distillation into other models.
  • Distilled models: six dense checkpoints (1.5B, 7B, 8B, 14B, 32B and 70B) fine-tuned on about 800,000 samples generated by R1 and based on Qwen2.5 and Llama 3.x. They keep the licenses of their base models.
  • API: offered as deepseek-reasoner through an OpenAI-compatible API at $0.55 per million input tokens (cache miss) and $2.19 per million output tokens, with reasoning tokens billed as output.
  • Market impact: on January 27, 2025, Nvidia shares fell about 17 percent and the company lost nearly $600 billion in market value, a record one-day loss for any company on Wall Street.
  • Cloud availability: Microsoft added R1 to the Azure AI Foundry model catalog and to GitHub Models on January 29, 2025.

What Happened#

DeepSeek published the weights on Hugging Face and a technical report on GitHub. The report described two models. DeepSeek-R1-Zero was trained with large-scale reinforcement learning (RL) directly on the base model, with no supervised fine-tuning first. It learned useful behaviors such as breaking problems into steps and checking its own work, but its output was often hard to read, mixed languages and repeated itself. DeepSeek-R1 fixed this with a small "cold-start" fine-tuning set before RL, followed by further RL and filtering stages that produced cleaner answers. Hugging Face's analysis notes that the RL stage used Group Relative Policy Optimization (GRPO) to make training more efficient.

The benchmark claims came from DeepSeek itself. The company reported results comparable to o1 across math, code and reasoning suites, and said its distilled Qwen-based 32B model outperformed o1-mini on several benchmarks. Unlike o1, R1 shows its full chain of thought, which made the model easy to study and easy to use as a teacher for smaller models.

DeepSeek's free assistant app, which exposed R1 through a "DeepThink" toggle, overtook ChatGPT in downloads on Apple's App Store. On Monday, January 27, investors reassessed how much high-end compute frontier AI really needs, and Nvidia posted a record one-day loss in market value. Two days later, Microsoft made R1 available as a serverless endpoint in Azure AI Foundry and in GitHub Models, and said distilled versions would come to Copilot+ PCs.

Background#

OpenAI introduced o1 in September 2024 and showed that letting a model "think" longer at inference time improves results on hard problems. OpenAI kept the training recipe secret and hid the raw reasoning traces. DeepSeek had already drawn attention in December 2024 with DeepSeek-V3, a 671B MoE model that used multi-head latent attention and multi-token prediction. Its technical report put the GPU cost of the final training run at under $6 million, a figure that excluded earlier research and experiments.

R1 combined open weights, visible reasoning, a permissive license and a detailed description of how the reasoning ability was trained. DeepSeek did not release its training data or code, so Hugging Face launched Open-R1 on January 28, a community project to rebuild the missing pieces in the open.

In September 2025, the R1 work appeared as a peer-reviewed paper in Nature. According to CNN's report on the paper, it put the cost of the reasoning-specific training at about $294,000 on 512 Nvidia H800 GPUs, on top of roughly $6 million spent on the base model.

Why It Matters for Developers#

The biggest practical change is that reasoning now runs locally. The distilled 7B to 32B models run on a single workstation GPU through runtimes such as Ollama, vLLM and SGLang. In .NET, OllamaSharp implements the IChatClient interface from Microsoft.Extensions.AI, so the same application code can target a local R1 distill during development and a hosted model in production:

C#
using Microsoft.Extensions.AI;
using OllamaSharp;

// Run `ollama pull deepseek-r1:14b` first; other R1 distill tags work the same way.
IChatClient client = new OllamaApiClient(new Uri("http://localhost:11434"), "deepseek-r1:14b");

var response = await client.GetResponseAsync(
    "A nightly job starts at 22:45 and takes 3 h 50 min. When does it finish?");

Console.WriteLine(response.Text);

Reasoning models also need different handling from chat models. R1 writes its reasoning between <think> tags before the final answer. Depending on the runtime, that text arrives inline or in a separate field, so strip it before you show it to users or write it to logs. Budget for long outputs, because reasoning tokens count as output tokens. DeepSeek recommends a temperature between 0.5 and 0.7 and no system prompt, with all instructions in the user message. That affects how you template prompts in Semantic Kernel or a hand-rolled client.

Licensing needs attention even though R1 itself is MIT. The Qwen-based distills inherit Apache 2.0, while the 8B and 70B distills inherit Meta's Llama 3.1 and 3.3 licenses. Check which checkpoint you actually ship.

Finally, think about data governance. DeepSeek's hosted API is operated by a company based in China, which may not fit your compliance requirements. For enterprise workloads, the more defensible options are self-hosting the open weights or using a managed deployment such as Azure AI Foundry. Whatever you choose, treat the vendor benchmarks as claims and run your own task-specific evaluations, for example with the tooling described in our AI evaluation guide.

What's Next#

At launch, the open questions were reproducibility, training data and whether US labs would respond with open reasoning models of their own. Open-R1 and other community projects set out to reproduce the distillation and RL steps with public datasets. DeepSeek kept iterating in the open: its V3.1-Terminus model was followed in September 2025 by DeepSeek-V3.2-Exp, which introduced DeepSeek Sparse Attention for cheaper long-context inference, still under the MIT license.

Our reading, which is interpretation rather than fact, is that R1 raised the bar for what "open" means in AI: weights plus a usable recipe under a license that allows distillation. For .NET teams, the lasting lesson is architectural. Keep model access behind an abstraction such as Microsoft.Extensions.AI, so that switching between a local distill, a cloud endpoint and next quarter's model is a configuration change rather than a rewrite. For more on running models on your own hardware, see our guide to local AI with ONNX Runtime, Ollama and Foundry Local.

Sources#