https://huggingface.co/Menlo/Lucy

Lucy is a compact but capable 1.7B model focused on agentic web search and lightweight browsing. Built on Qwen3-1.7B, Lucy inherits deep research capabilities from larger models while being optimized to run efficiently on mobile devices, even with CPU-only configurations.

We achieved this through machine-generated task vectors that optimize thinking processes, smooth reward functions across multiple categories, and pure reinforcement learning without any supervised fine-tuning.

https://arxiv.org/html/2508.00360v1

The main theoretical idea: reasoning as a “dynamic task vector machine”

The paper’s most important conceptual contribution is its treatment of the model’s <think> ... </think> reasoning process.

Rather than viewing the text inside <think> tags as merely an explanation or chain-of-thought transcript, the authors argue that the generation process itself can act as a dynamic task vector machine. In their framing, the model’s internal task representation is progressively:

• constructed,
• updated,
• refined,
• and aligned with the next tool action.

A useful paraphrase of the paper’s argument is:

The model should not merely generate a reasoning trace after deciding what to do. Its reasoning process should help construct and maintain the representation of the task that guides subsequent searches and answers.

This is intended to address a common problem in agentic systems: a model may know how to call a search tool but still produce unreliable, repetitive, or incoherent search trajectories.

The authors therefore focus less on teaching a fixed search procedure and more on stabilizing the reasoning loop that determines when and how the model searches.

Training method

Search-efficiency reward

The model is rewarded for visiting and examining retrieved documents rather than repeatedly issuing new searches. If it performs more searches than document visits, the paper applies a penalty.

This encourages Lucy to make better use of information it has already retrieved instead of continually broadening the search. Efficient-thinking reward

The authors penalize excessively long internal reasoning. Their reward distribution is centered around approximately 35 reasoning tokens, with a strong penalty for unusually long reasoning spans.

The motivation is to combat “overthinking”—reasoning that is verbose, repetitive, or distracts the model from using tools effectively.

The paper explicitly says that these parameters were selected heuristically, based on observations from preliminary experiments. Data and evaluation

The training dataset contains 10,325 question-answer pairs. It is described as an adaptation of the MuSiQue-Ans dataset, which contains multi-step questions requiring information gathering and reasoning.

For evaluation, the authors use SimpleQA and connect the evaluation system to an MCP server providing web-search and scraping tools.

The headline result is:

Lucy: 78.3% on SimpleQA with MCP
1.7B parameters

The paper compares this with several systems:

DeepSeek-671B with MCP: 78.2%
Jan-nano, 4B: 80.7%
Jan-nano-128k, 4B: 83.2%

The authors emphasize that Lucy performs approximately on par with the much larger DeepSeek model in this setup, while being far smaller.

However, the comparison must be interpreted carefully: the result is for tool-augmented evaluation with MCP, not necessarily for a pure, closed-book language-model test. Lucy’s web access is an essential part of its performance. Emergent behavior: skipping unnecessary reasoning

One of the more interesting observations is that Lucy learned to suppress reasoning during predictable, low-decision operations.

The paper gives a simplified example:

When deciding what to search, Lucy generates a <think> segment.
When merely reading or processing a retrieved document, it may omit explicit thinking.
When deciding on a more complex follow-up search, it resumes thinking.

In paraphrase, Lucy appears to allocate reasoning selectively:

It spends computation on uncertain decisions, such as formulating a query, while skipping redundant deliberation during deterministic actions, such as reading a result.

The authors report a 17.8× latency improvement in the observed example. They also stress that this is only a single illustrative case and that its generality remains unproven.

This is one of the paper’s strongest practical ideas: reasoning need not be uniformly applied at every step. A useful agent may reason deeply only where the next action is ambiguous. An important limitation: more thinking cannot replace missing knowledge

The paper is unusually candid about Lucy’s limitations.

The authors report cases where Lucy failed not because the needed information was absent from the web, but because the model lacked basic knowledge about the entities involved. As a result, it formed incorrect assumptions and generated poor search queries.

Their conclusion is that test-time reasoning has a ceiling:

Thinking longer cannot reliably compensate for missing conceptual grounding.

The paper contrasts Lucy with the larger Jan-nano model, which was better able to notice and correct its own mistaken assumptions during research. Lucy, by comparison, sometimes continued searching within an incorrect conceptual frame.

caveats are that:

the evaluation depends on external web tools;
the training dataset is relatively small and specialized;
several reward components are heuristic;
the correctness reward is described as a substring match, which may not perfectly measure semantic correctness;
the latency improvement comes from an illustrative scenario;
and the authors acknowledge that Lucy can fail when it lacks foundational knowledge needed to formulate sensible searches.