18 Dec 2024 AI Engineering 4 min read

DeepSeek-R1: Open Reasoning Is Coming

OpenAI's o1 made test-time reasoning the frontier capability of late 2024 — closed, expensive, and API-only. DeepSeek's R1-Lite-Preview and Alibaba's QwQ-32B-Preview signal that open alternatives are weeks away, not years. We examine what reinforcement-learned reasoning actually does, what open weights would do to its cost structure, and why that matters for European data sovereignty.

Reasoning Became a Product Behind an API

On September 12, 2024, OpenAI released o1-preview, the first commercial model trained to reason before answering. On December 5, o1 followed in ChatGPT. Yesterday, December 17, it reached the API as o1-2024-12-17. Reasoning is now the frontier capability of this industry — and it is closed. The chain of thought is hidden, the weights are inaccessible, and the API price stands at 15 US dollars per million input tokens and 60 dollars per million output tokens.

For European engineering teams this is a double dependency: the capability and the infrastructure that runs it belong to the same vendor. Every architecture decision downstream — caching, routing, fallbacks — inherits that dependency. Two announcements from China in the past four weeks indicate this state will not last.

A request arrives — the router classifies its difficulty. 1/4

Test-Time Compute Is the New Scaling Axis

o1-class models change where compute is spent. Instead of scaling parameters further, they are post-trained — reinforcement learning plays the central role — to produce long chains of thought before the final answer. OpenAI describes them as models that spend more time thinking before they respond. Accuracy then scales with thinking time, not model size. DeepSeek has published the curve: on AIME, its R1-Lite-Preview scores 21 percent with fewer than 1,000 reasoning tokens and 66.7 percent with more than 100,000.

The technique has clear limits. It does not make a model smarter per token; it buys accuracy with latency and output cost. For latency-sensitive or high-volume paths it is the wrong tool.

DeepSeek-R1 Shows Its Work

On November 20, 2024, DeepSeek released DeepSeek-R1-Lite-Preview in its web chat, limited to 50 messages per day. Unlike o1, which shows only a summary, R1 streams its full reasoning process in real time, self-corrections included. DeepSeek announces verbatim: "Open-source models & API coming soon." The self-reported benchmarks place the preview at or above o1-preview on math-heavy tasks.

The numbers are DeepSeek's own and not independently verified. On GPQA Diamond, LiveCodeBench, and ZebraLogic the preview trails o1-preview. Weights, license, and release date are still unpublished. Until they ship, R1 is a demo, not a dependency you can build on.

Benchmark	DeepSeek-R1-Lite-Preview	OpenAI o1-preview
AIME 2024 (pass@1)	52.5	44.6
MATH (accuracy %)	91.6	85.5
Codeforces (rating)	1450	1428

QwQ Puts Reasoning Weights on Hugging Face

One week later, on November 27, Alibaba's Qwen team released QwQ-32B-Preview — the first downloadable reasoning model under Apache 2.0. It has 32.5 billion parameters, a 32,768-token context window, and reports 65.2 percent on GPQA, 50.0 percent on AIME, and 90.6 percent on MATH-500.

The Qwen team names its defects itself: the model mixes languages mid-answer and falls into recursive reasoning loops. That candor matters. It shows that reasoning behavior fits into a 32B model an engineering team can run, quantized, on its own hardware — imperfectly, but inspectably.

What Open Weights Do to Reasoning Cost

Reasoning tokens are billed as output tokens. At o1's 60 dollars per million, a single answer that thinks for 20,000 tokens costs 1.20 dollars before the actual answer is written. This multiplier is invisible in benchmarks and dominant in production, and caching mitigates little because each request generates its own trace.

DeepSeek's pricing history explains the anticipation around R1. In May 2024 the company launched DeepSeek-V2, a 236-billion-parameter mixture-of-experts model, at 0.14 dollars per million input tokens — roughly one hundredth of GPT-4-Turbo — and set off a price war among Chinese providers. Open weights push this further: the marginal price of a reasoning token becomes your GPU amortization and utilization, not a vendor's margin. Self-hosting is not free; it trades API spend for operations effort and capacity planning.

Sovereignty Is a Deployment Property

Open weights turn model choice into a deployment decision. Weights on your own hardware mean data residency by construction, pinned versions instead of silent model updates, and no deprecation risk: API models are retired on the vendor's schedule, weights on your own disk are not. Under the GDPR and the EU AI Act, in force since August 1, 2024, these properties are worth real money to German enterprises.

Honest scoping: open weights are not open training data. Provenance, alignment, and failure modes of a model trained in China must be evaluated like any other supply-chain dependency. We treat model evaluation as an engineering discipline, not as leaderboard reading. And sovereignty over weights does not extend to the export-controlled GPUs they run on.

What Open Reasoning Models Do Not Solve

Both models are previews. R1's size, weights, license, and RL recipe are unpublished; what serving it requires is unknown. QwQ's own model card warns against unsupervised deployment. All benchmark numbers are self-reported.

The paradigm itself has costs no license fixes: answers take seconds to minutes, and a visible chain of thought is not a faithful explanation of the model's computation — it is output, and it can rationalize. Reasoning models complement fast models; they do not replace them.

Outlook From December 2024

We expect DeepSeek to publish R1 weights within weeks, plausibly with a technical report on the reinforcement-learning recipe. We also expect distillation: visible reasoning traces are training data, so small open models that imitate long chains of thought are the obvious next step.

If an open model reaches o1-level reasoning at DeepSeek-V2-level prices, the market must reprice what reasoning capability costs to build and to run — including the assumption that it requires frontier-scale capital. Our working plan for 2025: build model-agnostic evaluation harnesses now, and treat reasoning as a commodity you can host, not a feature you rent.