How DeepSeek R1 went to market with open weights and let X tell the story

The X angleResearchers and investors on X spread R1's results within days of release, and Marc Andreessen's January 2025 post calling it a Sputnik moment for AI drew wide attention.

DeepSeek R1 is the strangest launch in this series, because the company barely marketed it. DeepSeek, a Chinese AI lab, published R1 on January 20, 2025, as a reasoning model with open weights, a technical report and a permissive license, alongside its free chat app. There was no launch event aimed at Western audiences and no founder touring podcasts. Within a week, its app had overtaken ChatGPT at the top of the US App Store's free chart and its release had become a talking point in financial markets. The engine was X, where researchers, developers and investors argued about what R1 meant. For founders studying go to market strategy examples, it shows what happens when a product's proof is public enough for others to carry it.

The product and the bet

R1 was a reasoning model: it worked through problems step by step before answering, in the style that OpenAI's o1 had introduced a few months earlier. DeepSeek claimed that R1 matched o1 on several math, coding and reasoning benchmarks. It was built on DeepSeek V3, a large mixture of experts model the company had released in December 2024.

The distinctive choices were about openness rather than capability:

• Open weights. Anyone could download R1 and run it, inspect it or build on it.
• A permissive license. DeepSeek's release note said code and models were released under the MIT License, allowing commercial use and distillation.
• A technical report. The company published a detailed account of how R1 was trained, including its heavy use of reinforcement learning.
• Distilled versions. DeepSeek also released smaller models distilled from R1, which could run on far more modest hardware.

The bet, in analysis, was that openness itself would be the distribution strategy. A closed model has to be marketed. An open model with strong results gets tested, benchmarked, fine-tuned and discussed by thousands of people who have their own reasons to talk about it.

Who it was for

R1 had three audiences at once. AI researchers and engineers could study and build on it. Developers could use it cheaply through DeepSeek's API or run it themselves. Consumers could use it free in the DeepSeek app, which displayed the model's reasoning before its answer.

That last detail mattered. Showing the reasoning made the product feel different from rival chat apps and gave users something novel to screenshot. It turned an abstract technical feature into something visible.

The launch

DeepSeek's release note, dated January 20, 2025, announced R1 as a fully open source model with a technical report, under the MIT License. The model became available through the company's website, app and API, and the weights were posted for download.

The first wave of attention was technical. Researchers read the paper, developers ran the model and benchmark results circulated. Within days, the conversation widened. A key claim in public discussion, drawn from DeepSeek's earlier V3 report, was that its models had been trained far more efficiently than Western competitors' models. That claim was later disputed by others in the industry, but it became central to how R1 was perceived.

By January 27, 2025, TechNode and other outlets reported that DeepSeek's AI assistant had overtaken ChatGPT to reach the top of the US App Store's free apps list. The same day, DeepSeek said it was limiting new registrations because of what it described as large-scale malicious attacks on its services, while existing users could still log in. Technology stocks fell sharply that Monday, with Nvidia suffering a record one-day loss in market value, a move widely linked in coverage to concerns raised by DeepSeek's models.

How X carried it

DeepSeek's own posts were not the main story on X. The story was what others said.

Researchers as the first amplifiers

Open weights and a detailed paper gave researchers material to work with immediately. They posted replication attempts, benchmark comparisons, observations about the reasoning traces and commentary on the training method. Because these posts came from people with technical credibility, they carried weight with the broader AI audience on X. In analysis, this is the open source dynamic at its most powerful: the community does the product evaluation in public, and its conclusions become the marketing.

Screenshots of the reasoning

Consumers who downloaded the app posted screenshots of R1 thinking through problems. The visible reasoning was novel and often surprising, sometimes because it was impressive and sometimes because it was oddly human in its hesitations. It was exactly the kind of content that spreads on X.

Investors turned it into a macro story

On January 26, 2025, venture capitalist Marc Andreessen posted on X describing R1 as AI's Sputnik moment, comparing it to the satellite launch that kicked off the space race. He also praised it separately as one of the most impressive breakthroughs he had seen. The framing spread quickly, and it shifted R1 from an AI research story to a geopolitical and financial one. The next day's stock market reaction amplified that shift, and coverage moved from technology sections to front pages.

This is the clearest example in this series of an X post by someone outside a company reframing that company's launch. The Sputnik comparison gave journalists, investors and policymakers a ready metaphor, and it made R1 relevant to people who would never download a model.

What came next

Capacity and access

The surge in attention came with operational strain. Limiting new registrations during the attacks reported in late January showed the risk of sudden global demand for a company not set up for a consumer launch at that scale. Founders should note that the same viral dynamics that create a launch can overwhelm the infrastructure behind it.

Pricing as part of the story

DeepSeek's API pricing was widely discussed as being far lower than that of comparable Western reasoning models. Combined with free consumer access and open weights, that made R1 the default low-cost option for many developers. Qualitatively, the model was a free consumer app, a low-cost usage based API and free weights for anyone who wanted to host the model themselves.

Iterating in the open

DeepSeek continued to release updates. In May 2025, it shipped an updated checkpoint of R1 with improved benchmark results. In August 2025, DeepSeek V3.1 introduced a hybrid design with both thinking and non-thinking modes, letting users switch between fast answers and deeper reasoning in the app and on the web. Each release followed the same pattern: post the model, publish the details and let the community test it.

The go to market lessons

DeepSeek's situation is unusual. Few startups will release a model that moves stock markets. But several of its go to market mechanics apply to smaller companies, particularly in developer tools and infrastructure. The lessons below are analysis.

1. Openness can replace a marketing budget

Open weights, a permissive license and a detailed technical report gave thousands of people a reason to talk about R1. For technical products, giving away enough to be tested in public can generate more credible attention than any campaign. This is a key consideration in any go to market strategy for AI startups with a developer audience.

2. Make the differentiator visible

Showing the reasoning in the consumer app turned a technical feature into a shareable experience. Founders should ask which part of their product is invisible today and whether exposing it would make the value obvious.

3. Benchmarks against the leader create a frame

DeepSeek positioned R1 directly against OpenAI's o1. That comparison gave every reader a reference point. Positioning against a known leader works when the results are credible enough for outside experts to confirm.

4. Prepare for attention you cannot control

Andreessen's post turned R1 into a geopolitical story. Companies cannot script that kind of reframing, and it can bring scrutiny on security, policy and data handling. A go to market team launching anything ambitious should have a plan for how to respond if the story escapes the original audience.

5. Price as a positioning statement

Low API prices did more than win usage. They reinforced the efficiency narrative and invited comparison. For a b2b go to market strategy, price can be part of the message, not only the business model, though it has to be sustainable.

6. Capacity is part of the launch plan

Registration limits during the peak of attention were a reminder that infrastructure is a go to market concern. Any go to market plan for a product that might spread fast should include a capacity plan and a communication plan for when demand outruns supply.

Open release as a go to market framework

DeepSeek's approach can be laid out as a simple framework. The table below is analysis based on the public record.

The common thread is that every lever invites outside verification. Open releases work as distribution only when the results hold up under that verification. A weaker model released the same way would have been tested just as publicly, and the conversation on X would have moved on quickly.

Where it fits among go to market strategy examples

DeepSeek R1 completes a pattern visible across this series. ChatGPT showed what a free research preview could do on X. Manus showed how scarcity and video could create a global story. DeepSeek showed that openness, strong results and visible reasoning could let the community, and a handful of influential voices on X, carry a launch further than any campaign. The full set is in the examples library.