PokeeResearch 7B

What is PokeeResearch 7B?

PokeeResearch 7B is a research agent rather than a chat model. Pokee AI fine-tuned Qwen2.5-7B-Instruct with reinforcement learning from AI feedback, so it plans searches, reads pages through outside tools, checks its own answer and can merge several independent research threads before replying. It reads and writes text only and works within a 32K-token context.

It is still current: Pokee offers it as its open-source agent beside a hosted research API, but the code on GitHub has not changed since launch. Running it takes your own GPU plus Serper, Jina and Gemini keys, so it is a project for developers, not a ready app.

Is PokeeResearch 7B still maintained?

It is still available and still Pokee's open research agent, but it has had no public updates since launch.

Weights

Still on Pokee's Hugging Face page under Apache 2.0, with no sign-up gate.

Code

The GitHub repository has not changed since release, and pull requests from outside contributors remain open.

Later checkpoint

Pokee has posted a second 7B research checkpoint on Hugging Face, but with no model card or notes, so it is not a documented upgrade.

Hosted option

Pokee calls its closed, hosted research agent its flagship and sells it through its Deep Research API.

Who is PokeeResearch 7B best for?

PokeeResearch 7B is best for developers and researchers who want a small, open research agent they can run, study and fine-tune on their own hardware.

• A research agent you can inspect

  Weights and code are Apache 2.0, so you can run it, read every research step it saves and change how it works.

• Studying agent training

  The repository includes the evaluation scripts and benchmark set behind Pokee's original numbers, and the paper explains each training choice.

• Fine-tuning for one domain

  At 7B parameters it is a cheap base to adapt for scientific or document research, a use Pokee's model card names.

• Multi-step fact questions

  It does best on questions that need several searches, scoring 92.0 on HotpotQA and 97.8 on NQ in Pokee's test.

What's included with PokeeResearch 7B?

Web search and page reading, self-verification, merged research threads, recovery from broken tool calls, and command-line and web apps.

Web search and page reading

It calls Serper to search and Jina to read pages, over up to 10 turns per question by default.

Self-verification

It writes a check of its draft answer against what it found and can correct itself when the check fails.

Research Threads Synthesis

It can run several independent research threads on one question and merge them, which lifted its HLE, GAIA and BrowseComp scores in Pokee's original tests.

Recovery from broken tool calls

A malformed tool call does not end the run; the agent recovers and keeps researching.

Command-line and web apps

A CLI for single or interactive questions and a Gradio web app, serving the model locally or through vLLM.

Evaluation scripts

Scripts rerun Pokee's benchmark test and save each question's full research trail and the judge's decision.

How does PokeeResearch 7B work?

1. Start the Docker image

   Pokee supplies a Docker image with the dependencies; clone the repository inside it and log in to Hugging Face.

2. Add your API keys

   Put your Serper, Jina and Gemini keys in a .env file in the project folder.

3. Start the tool server

   It handles search and page reading, with an optional cache that saves API credits; the web app starts it for you.

4. Ask a question

   Use the command line for one question or an interactive session, or open the Gradio app in your browser.

How much does PokeeResearch 7B cost?

PokeeResearch 7B is free: the weights and code are Apache 2.0, and you pay only for your own GPU and the search, page-reading and Gemini APIs it calls.

Pokee does not sell PokeeResearch 7B. The web searches, page reads and Gemini calls it makes are billed by those services at their own rates. Pokee's hosted Deep Research API runs a different, closed agent and shows no public price. Full pricing on github.com. We checked these prices on 28 Sep 2026.

How do readers rate PokeeResearch 7B?

No one has rated PokeeResearch 7B yet. Be the first.

PokeeResearch 7B pros and cons

It is free, open and the strongest open 7B research agent in Pokee's tests, but it leans on three outside services and its code has not moved since launch.

Is PokeeResearch 7B open source?

Yes. PokeeResearch 7B's code is public under the Apache 2.0 licence, so you can run it on your own server.

What are the best PokeeResearch 7B alternatives?

PokeeResearch 7B is an open source alternative to Gemini Deep Research Agent and Perplexity. Scira, Atria Dawn Preview, Khoj and Co-STORM do the same job and are open source too.

Gemini Deep Research AgentPaid

Choose it when you want deep research over an API without running a GPU; Google estimates about $1 to $3 per task.

PerplexityFree

Choose it when you want deep research in a ready app with no setup; it comes with the $20-a-month Pro plan.

SciraOpen Source

Research assistant that finds answers and acts across 100+ apps.

Atria Dawn PreviewOpen Source

Choose it when you want a much larger open research agent that also writes code and runs experiments; it is a 744B MoE model under the MIT licence, with a hosted API if you lack the hardware.

KhojOpen Source

Personal AI that answers from your notes, PDFs and the web, with any local or online model.

Co-STORMOpen Source

Stanford's multi-agent research tool that writes cited reports.

Questions about PokeeResearch 7B