DeepSeek launches R1-Lite-Preview to rival reasoning models

On November 20, 2024, DeepSeek opened R1-Lite-Preview in its chat as a preview of a reasoning-oriented model. DeepSeek’s official note compares mathematics results and links more inference compute to higher scores, while postponing weights and API access. Proper reading begins by preserving the protocol: one answer and a vote across many attempts are not the same cost or test.

The announcement matters for two reasons. First, it confirms that Chinese labs are not limiting themselves to competing in general-purpose chatbots: they are also developing models that devote more computation to solving complex problems. Second, DeepSeek has made R1-Lite-Preview available for testing through its chat service, exposing part of its reasoning process before it delivers the final answer.

A model that tries before it answers

Conventional language models typically generate an answer almost immediately, one word at a time. R1-Lite-Preview takes a different approach: it produces intermediate steps, reviews its reasoning and reaches a conclusion afterward. That does not mean it thinks like a person, nor that its explanations are a reliable window into how it works internally, but it is optimized to devote more effort to tasks where a first intuition is often wrong.

This approach is particularly useful for math exercises, programming, logic puzzles and questions that require linking several conditions. It is also the path OpenAI popularized with o1, presented as a model that can think for longer before answering.

The practical difference for users is clear: responses may take longer, but they are more likely to be correct when a problem requires several calculations or some planning. The trade-off is that displaying a long chain of reasoning does not eliminate errors. A model can produce a convincing explanation and still start from a mistaken premise or make an arithmetic error.

Results close to o1-preview on math

DeepSeek compared R1-Lite-Preview with o1-preview on AIME 2024, a test of U.S. competition math problems commonly used to evaluate reasoning. The company puts its model at 41.5% accuracy with a single answer per question, versus 44.6% attributed to o1-preview. Source

The gap is small on a demanding test, although it should be interpreted cautiously. Benchmarks measure a specific ability under controlled conditions; they do not guarantee that a model will perform the same way in a spreadsheet, a legal query or when writing code for production. It also matters how the question is phrased, how much computation the system is given and whether it can generate multiple answers from which to choose the best one.

DeepSeek says that, through majority voting across multiple runs, R1-Lite-Preview reaches 86.7% on AIME 2024. This technique improves the result because the system tries to solve the same problem several times and selects the most frequently repeated solution. It is a useful tool for evaluating capability, but it consumes considerably more computing power than a single query and does not represent a chatbot’s typical performance. Source

Available to try, but not with open weights

R1-Lite-Preview is available through DeepSeek’s chatbot. The company allows users to observe the reasoning steps generated by the model, a notable decision at a time when OpenAI has chosen not to show users o1’s complete chain of thought.

That said, the launch does not make R1-Lite-Preview an open model in the strict sense. DeepSeek offers access to the service, but it has not released the model’s weights—the files that would allow users to download it, run it on their own servers or study it independently. The distinction matters: public access is not the same as open code or an open model.

The company has also not detailed R1-Lite-Preview’s architecture, training volume or computational cost. Without that information, it is difficult to assess how easily its approach could be reproduced outside DeepSeek or compare its efficiency with U.S. offerings.

The race enters a new category

Until a few months ago, discussion of language models focused largely on size, fluency and the ability to maintain a conversation. Reasoning systems shift the focus to a different question: how much additional work can a model do before providing an answer?

For companies and developers, that shift opens up possibilities in technical analysis, programming and the automation of processes with complex rules. It also makes it necessary to distinguish between tasks that need a fast, inexpensive answer and those where it is worth waiting and paying more for a second check.

R1-Lite-Preview is still a preliminary version, and its results depend heavily on the tests reported by its creator. But its arrival shows that o1 has not opened a category reserved for OpenAI. Reasoning is becoming the new battleground among major models, and DeepSeek has just entered the race.

“Reasoning” needs an observable criterion

A written sequence can help locate an error but does not show that those steps caused the answer. Preserve question, final answer, time, tokens and external verification. Check mathematics results, execute code tests and require sources for facts. Intermediate prose is supporting evidence.

Include problems with false premises or insufficient information. A system that always produces a long derivation may look diligent and be less reliable because it never stops. Recognising that no solution is available belongs to reasoning.

One run and a vote do not share an axis

Generating candidates and selecting a majority often raises accuracy when errors differ, but multiplies cost and latency. Report one attempt, multiple attempts and verifier-based selection separately. Comparing the last mode with another model’s first attributes a budget advantage to the model.

The useful curve plots quality against compute. When more time helps and then plateaus, an operating point appears. Measure dispersion too: a high average with unpredictable failures may be worse for an application than lower, stable capability.

The benchmark must touch the real task

AIME and MATH test particular forms of mathematics; they do not establish general planning, factual reliability or safe tool use. For transfer, prepare your own unseen and verifiable problems, preserve criteria and change one condition at a time.

The preview adds another limitation: without weights or API on that date, only the offered interface could be tested. Service configuration, answer selection and local reproducibility were unknown. Stating access conditions prevents a controlled experience from becoming a universal property.

The transferable skill is to compare reasoning models through the same unit: task, attempt, budget, verifier and cost. The technique’s name matters less than a reproducible curve of successes and failures.

Selection needs its own evaluator

A majority may select the common error. For verifiable problems, compare voting, formal checking and selection by another model. Measure what each layer adds and which new failures it introduces. A learned judge may prefer persuasive explanations even when the solution is false.

Repeated answers are not necessarily independent samples when they share prompt, data and bias. Changing temperature creates surface diversity without guaranteeing different strategies. Inspecting solution trees and error types shows whether the budget explores or merely paraphrases.

In production, set a time limit and an abstention output. When the system exhausts its budget, it should report that fact rather than present the last attempt as certain. Recording how many tasks hit the limit supplies the denominator missing from a selected demonstration.