IQuest-Q1 Debuts for RL Debugging as Anonymous Space Bunny Alpha Tops OpenRouter Calls

IQuest-Q1 uses a decoder-only Transformer backbone with sparse MoE. According to QbitAI, its total parameter count is about 320B, while only about 15B parameters are active. QbitAI reported that the model performed well on complex evaluations including NL2Repo for repository-level code generation, CyberGym for cybersecurity, Terminal-Bench 2.1, DeepSWE v1.1 for long-horizon code tasks, and JobBench for office scenarios. The official release also says IQuest-Q1 took part in its own research and development iteration: when its proposals were approved by human researchers, validated model updates, training assets and research workflows entered the next round, with data pipelines, training configurations, evaluation methods and tools accumulating as reusable assets.

In a QbitAI test, IQuest-Q1 generated an HTML game called National Day Bedroom Defense for a vertical mobile phone from a prompt of about 200 words. The game included NPCs with individual settings and health bars, a pillow weapon and a bed that restored health. It also randomly offered a daily goal from options such as milk tea, fried chicken, watching a movie, visiting a supermarket or taking a nearby walk, and it allowed the result to be saved and shared. QbitAI also said the model numerically solved the Three-Body problem's equations of motion within 100,000 steps and could render the result in a style compared to flowing golden lines.

A second QbitAI test involved a real reinforcement-learning training failure. The reward curve failed to climb as expected. IQuest-Q1 was asked to analyze training logs and saved trajectories and to trace the cause from the codebase using the Claude Code CLI framework. It located the problem in the text return and trajectory stitching step when the RL framework connected to Scaffold: the inference service inserted an extra space during text decoding. That space made generated text and returned history misalign, broke prefix matching and interrupted multi-turn generation, so earlier RL rounds that read code, ran commands and edited code were wasted, and only the final answer was trained. IQuest-Q1 fixed the bug and training resumed, QbitAI reported.

IQuest Research had built a series of related results before IQuest-Q1. From late July to August, the team reported work including MuonH optimization, UBio-MolFM and sparse weight decomposition. On Sept. 16, the ModularRSI self-evolution framework co-proposed by IQuest Research ranked second on the Hugging Face daily list. According to QbitAI, it raised an agent's accuracy on Terminal-Bench 2.0 from 47.57% to 52.43% and on SWE-Bench Verified from 73.40% to 76.45%, while allowing learned execution abilities to be reused across tasks and models.

Space Bunny Alpha appeared with less explanation. According to ifanr, it has no technical report and no large model company has claimed it. It was quietly launched on OpenRouter and OpenCode and then rose to first place in daily calls on both platforms through user tests. OpenRouter described it as fast and able to remember a lot of content. Users reported that it could view images, write code, open tools, modify files and carry a task forward step by step. One user compared it with GPT-6 Astra on a detailed voxel-style Japanese garden and found Space Bunny Alpha stronger in detail. Another uploaded a game video, and the model reconstructed the spaceship, storm, boss, combos and sound effects.

ifanr tested Space Bunny Alpha by giving it a rough sketch. A circle stood for a record, wavy lines for an introduction, a box marked 'buy' for a button, and an arrow beside a shopping bag for the animation direction. The model produced a website that could be opened in less than five minutes and filled in background, layout and other details that the sketch did not show. A more casual sketch of a record store, with an entrance below, a record wall on the left, a large window on the right, a listening station in the middle and a bent arrow marked 'choose one and put it here,' became a 3D webpage where the camera could rotate, a clicked record pulled its sleeve from the shelf, and a clicked listening station placed the record on a turntable.

Using an OpenRouter API key and DeepSeek Harness, ifanr also had Space Bunny Alpha write files, start projects and call a browser to check results. It produced projects including a space exploration scene, an ink-covered ground game and a liquid oxygen methane engine exhibit. The ink game placed players on a sunny rooftop park for 60 seconds, with one player using orange ink and the other cyan ink; whoever covered more ground won. The engine exhibit included silver-gray metal, pipes, turbopumps, a preburner, a combustion chamber and a bell nozzle. It offered cutaway and exploded views that separated components, cut open the nozzle and kept connecting lines between parts. Space Bunny Alpha also built a playable browser-based 3D Peach Blossom Spring scene, though ifanr said its visuals had less detail and texture than high-completion scenes made with Opus 5.5, and GPT-6 Astra handled complex 3D motion more smoothly. At the Max reasoning setting, ifanr observed the model could overthink and take longer than flagship models.

The identity of Space Bunny Alpha remained unconfirmed. Ifanr reported that the name invited guesses about moon-related companies such as Moonshot and Kimi, but tokenization evidence pointed more strongly toward MiniMax. A researcher used 50 strings in two rounds of testing and found that Space Bunny Alpha matched all eight MiniMax versions tested, and the open-source project stealthprint also reported tokenizer features consistent with MiniMax models. Ifanr added that test code for a 'MiniMax-M3.1' configuration, with one million context, forced thinking and low, high and max reasoning levels, was written into the MiniMax Code repository on Sept. 18, five days before the bunny appeared. On Sept. 27, MiniMax quietly launched M3.1-Flash-Preview with similar specifications. That evidence was not enough to determine the developer or exact version, ifanr said. The model is currently free on OpenRouter under the ID stealth/space-bunny-alpha and on OpenCode as space-bunny-free.

Editor's Summary

IQuest Research's IQuest-Q1 and the anonymous Space Bunny Alpha both put fast, long-context coding and agent delivery at the center of model competition. IQuest-Q1 showed its value in game generation and RL debugging, while Space Bunny Alpha's sketch-to-web and 3D tests made it a popular free option on OpenRouter and OpenCode. Space Bunny Alpha's developer has not been confirmed, though tokenization evidence reported by ifanr points toward MiniMax.