A state-of-the-art general-purpose model in its class, built through automated AI research and designed to advance AI and scientific discovery.
Today, Apex Intelligence is releasing its first model, Apex-35B-V1.0.
This 35B general-purpose model is based on Qwen3.6-35B-A3B and has been further developed through mid-training and post-training. Its development incorporated extensive proprietary datasets and research trajectories.
Apex-35B-V1.0 leads the similarly sized models evaluated in R&D capabilities while maintaining strong general-purpose performance, and is competitive with much larger models on selected benchmarks.
During its development, AI autonomously participated in data construction, experiment design, code execution, training-result analysis, and training optimization. In use, Apex-35B-V1.0 can handle standard tasks such as coding, agent workflows, question answering, and deep search. It can also carry out end-to-end computational research workflows in areas such as AI research, computational science, and robotics simulation.
Apex-35B-V1.0 is both a product of automated AI research and a tool for advancing further research in AI and science.
Benchmark Performance
Exceptional R&D Capabilities. General Excellence Preserved.
Apex-35B-V1.0 delivers leading research performance among the similarly sized models evaluated, outperforming both Qwen3.6-35B-A3B and Recursive Flint across all six research benchmarks. Its strengths extend to agentic tasks, coding, and reasoning, approaching or surpassing substantially larger models on selected benchmarks. It outperforms GLM-5.3 on FrontierScience and PaperBench, exceeds DeepSeek-V4.1-Flash-1017 on BioMysteryBench Human-difficult, and achieves the highest BFCL V4 score among all models compared. These results demonstrate strong R&D capabilities alongside broad general-purpose performance at the 35B scale.
Demo Cases
Robotics: From Task Execution To Real-To-Sim Reconstruction
Apex-35B-V1.0 demonstrates strong capabilities in robotic task execution, trajectory collection, and real-to-sim scene reconstruction. On RoboRSI and Real2Sim, our 35B model achieves scores of 44.5 and 21.55, respectively, significantly outperforming other models under 50B parameters and delivering performance competitive with the much larger, 300B-class GLM-5.3-Flash (52.4 and 21.4, respectively).
The results below compare Apex-35B-V1.0 against GLM-5.3, GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, Qwen3.6-35B-A3B, and Flint.
The following demonstrations highlight two complementary capabilities: using code and visual feedback to execute robotic tasks, and transforming real-world observations into executable physics simulations.
RoboRSI: Robotic Task Execution And Trajectory Collection
RoboRSI evaluates how effectively AI agents can control robots through code generation and iterative feedback in two simulation environments: RobotTwin and RoboCasa.
Given a robotic task, the model generates control code, executes actions in the simulator, observes the results through visual feedback, and revises its approach when an attempt fails. The benchmark measures task completion and trajectory collection under fixed time and computational resource budgets.
The video below demonstrates five representative tasks:
Block stacking: Stack multiple blocks into a stable structure.
Size-based arrangement: Arrange blocks in descending order of size.
Target placement: Move a T-shaped block to a specified target position.
Target placement with distractors: Move a T-shaped block to its target while avoiding irrelevant objects.
Object manipulation: Pour a red ball from a cup into a trash bin.
These examples demonstrate the model's ability to translate task instructions into executable robot actions, interpret visual feedback, and iteratively improve its execution.
Real2Sim: From Real-World Images To Executable Simulations
Real2Sim evaluates a model's ability to reconstruct a physics-enabled simulation environment from a single real-world image.
Given a 2D photograph, Apex-35B-V1.0 identifies the objects in the scene, selects simulation assets with similar visual and physical properties, and reconstructs their spatial arrangements. It then builds a physics-based virtual environment in which robotic actions and object interactions can be simulated and tested.
Unlike purely visual 3D reconstruction, the goal is to create a functional simulation that approximates both the scene's appearance and its physical behavior, enabling downstream robotic experimentation.
The six examples below compare the original real-world photographs with environments reconstructed by the Base and Apex models. The accompanying animations show the resulting physical interactions and object dynamics in each simulated scene.
Together, these examples demonstrate how Apex-35B-V1.0 bridges real-world perception and simulation-based robotic experimentation.
ReferenceBase (Qwen3.6-35B-A3B)Apex-35B-V1.0
Computational Fluid Dynamics
Apex-35B-V1.0 uses simulation feedback to optimize fluid dynamics systems, improving performance while satisfying physical and operational constraints. On CFD-Bench, it achieves a 79% win rate against the base model, with 79 wins, 0 draws, and 21 losses.
The benchmark compares Apex-35B-V1.0 with GLM-5.3, GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, and Flint, using Qwen3.6-35B-A3B as the shared baseline.
The following demonstrations highlight two complementary optimization capabilities in fluid dynamics: designing continuous motion profiles to improve propulsion efficiency and optimizing parameter assignments across interconnected fluid networks.
Flapping-Foil Motion Optimization
Flapping-foil simulation is widely used in the design of propulsion systems, including underwater vehicles and bio-inspired robotic systems. It allows engineers to explore motion strategies, evaluate performance under different flow conditions, and optimize designs before committing to physical prototypes.
In this task, Apex-35B-V1.0 optimizes a foil's vertical motion (heave) and rotation (pitch) to maximize net thrust across varying inflow conditions while limiting actuation costs and motion variability.
The motion is parameterized by 32 Fourier coefficients, which collectively define the shape and timing of each flapping cycle. By iteratively adjusting these coefficients based on simulation feedback, the model discovers more effective propulsion strategies.
Apex-35B-V1.0 achieves the highest net thrust of 0.164, outperforming GLM-5.3 (0.150) and Qwen3.6-35B-A3B (0.081).
Fluid Network Parameter Assignment
Complex fluid distribution networks are fundamental to many engineering systems, including municipal water supply, oil and gas pipelines, industrial cooling systems, and vascular circulation. These networks consist of interconnected branches, where changes in resistance, pressure, or flow at one outlet can influence the behavior of the entire system.
Optimizing such networks requires coordinating local parameters to achieve desired global performance, such as stable inlet pressure, balanced flow distribution, and efficient operation under varying conditions.
In this task, Apex-35B-V1.0 optimizes a vascular network with 64 terminal outlets by assigning predefined resistance and compliance parameters to match target inlet impedance and blood flow distribution. Each parameter set must be used exactly once, and changes at one outlet affect flow throughout the network, making this a challenging global optimization problem.
Apex-35B-V1.0 achieves the lowest mismatch of 1.24×10⁻⁵, outperforming GLM-5.3 (1.41×10⁻⁵) and Qwen3.6-35B-A3B (1.68×10⁻³).
AI For AI
Apex-35B-V1.0 applies its research and engineering capabilities to improve other AI systems. Through automated experimentation, training optimization, and iterative feedback, it develops effective training strategies and translates research findings into measurable improvements in model performance.
Training A Model To Solve Sokoban
In the AI4AI-Bench task ragen_sokoban_grpo, Apex-35B-V1.0 trains a Qwen2.5-3B-Instruct model to solve Sokoban puzzles, which require long-horizon spatial planning and reasoning. The task consists of two phases: 4 hours to explore training strategies and 12 hours to retrain the model, with performance evaluated on 512 unseen puzzle levels.
During exploration, Apex-35B-V1.0 followed a systematic research process: understanding the problem, designing experiments, analyzing results, and refining its training recipe. Across 207 commands and code edits, we identified 36 high-quality research and engineering actions, including 15 critical decisions that shaped the final training strategy. In contrast, the base model skipped reference materials and baseline evaluations, conducted poorly structured experiments, and failed to develop an effective training recipe.
Apex-35B-V1.0 achieved a 24.4% success rate on 512 unseen levels, more than doubling the base model's 11.7%.
How Apex Explores Training Strategies and Prepares the Final Retraining Recipe
● Apex-35B-V1.0▲ Base · Qwen3.6-35B-A3B
Swipe horizontally to explore the full timeline.
Select a circle or triangle to read the action and its timestamp.
Action Records And Chart Method
Circles show 36 substantive Apex actions plus 3 retrospective errors. Triangles show 9 Base milestones, rather than an exhaustive action count. Base exited after 101 minutes without a submission or a completed training step; the orange point is the untrained Qwen2.5-3B-Instruct reference, not a successful Base training run.
The retained curve combines separate training branches without connecting them. Display step 490 reuses checkpoint 500; step 455 is estimated. The final endpoint shows the reported 24.4% evaluation result (125 of 512 levels solved).
Reinforcement Learning For Image Generation
In the AI4AI-Bench task ddpo_sd15_aesthetic, Apex-35B-V1.0 uses DDPO to fine-tune Stable Diffusion 1.5 with LoRA, aiming to improve image quality measured by a fixed aesthetic reward model. The task follows the same two-phase workflow: 4 hours of exploration and 12 hours of retraining.
During exploration, Apex identified two key improvements to the training process:
Reducing computational cost. By inspecting the evaluation pipeline, Apex discovered that the benchmark used 20 denoising steps rather than the default 50. It tested the 20-step configuration and observed comparable aesthetic scores (5.499 vs. 5.487), while trial runtime decreased from 1,040 to 551 seconds.
Improving training utilization. Apex identified that a low fixed training-round limit could prematurely terminate retraining. It increased the limit to 1,000 rounds, introduced time-based stopping, and implemented complete checkpoint saving, enabling 409 training rounds within the available budget.
With these optimizations, the average aesthetic score improved from 5.397 to 6.638, a 23% increase.
Building Apex-35B-V1.0 Through Recursive Self-Improvement
We built Apex Research, an automated research system that enables recursive self-improvement (RSI) by continuously identifying weaknesses, generating hypotheses, running experiments, and turning discoveries into measurable model improvements.
Apex Research contributed throughout the development of Apex-35B-V1.0, from self-evolving training data and automated experimentation to training recipes, GPU kernels, and inference optimization.
Self-Evolving Data. Apex Research identifies model weaknesses and automatically generates targeted training data to address them. For example, when evaluations revealed failures in computational fluid dynamics (CFD), it analyzed the underlying reasoning gaps, synthesized and validated new scientific training examples, and incorporated them into subsequent training rounds. This self-evolving data pipeline improved performance on our internal CFD-Bench by 32.1 points.
Automated Experimentation. Apex Research analyzes evaluation results and agent trajectories to uncover systematic failure patterns and optimize training strategies. For example, it identified excessively long reasoning contexts and imbalanced data mixtures as sources of degraded agent performance. It then explored alternative sequence-length distributions, short-to-long trajectory sampling ratios, and token-aware batch compositions, discovering more effective training configurations that improved agent reliability and training efficiency.
Training and Inference Infrastructure. Apex Research extends automated optimization to the underlying compute stack, exploring more efficient GPU kernels, workload balancing, communication strategies, and execution scheduling. On the training side, it optimizes GPU load balancing, communication-computation overlap, dynamic batching, and token-aware workload distribution to reduce synchronization overhead, minimize GPU idle time, and improve hardware utilization, achieving 45–60% higher training throughput. On the inference side, it optimizes attention kernels, KV-cache management, and prefill/decode execution to improve serving efficiency, delivering 1.6–1.8× faster decoding and 30–35% lower time-to-first-token under comparable workloads.
Together, these discoveries compound across data, training, and infrastructure, creating a continuous feedback loop in which each research iteration makes the next more effective.
How To Use Apex-35B-V1.0
The weights for Apex-35B-V1.0 are available on Hugging Face.
The accompanying code, deployment guide, and usage examples are available on GitHub. Developers can download and run the model for AI development, scientific computing, and other research tasks. For details on model versions, hardware requirements, supported inference frameworks, and licensing, please refer to the project pages.
Next Steps
Apex Intelligence was founded by Yongchao Chen, an assistant professor at Tsinghua University’s College of AI. Its goal is to build the next generation of self-evolving foundation models for research.
By “self-evolving,” we mean recursive self-improvement (RSI): enabling AI to participate in research, experimentation, and validation, then using the resulting discoveries to improve the model itself.
We aim to help AI evolve from an assistant into a research collaborator, moving beyond accelerating known work toward discovering the unknown. Our long-term vision is to develop models whose collective research capability matches and progressively surpasses that of 100,000 top AI scientists across different disciplines, continuously contributing to scientific research and discovery.
Following the 35B model, Apex Intelligence plans to release a larger self-evolving foundation model. The new model will further strengthen its AI research capabilities and expand into scientific fields such as chip design and quantum computing.
More details on the model’s capabilities, benchmark results, and availability will be announced soon.
Footnotes
All Apex-35B-V1.0 results reported below are obtained with the reasoning effort set to 'max', temperature = 1.0, top-p = 0.95, and top-k = 20. Unless noted otherwise, all models are evaluated with the Claude Code harness; Qwen3.6-35B-A3B, Recursive Flint, GLM-5.3, GLM-5.3-Flash, and DeepSeek-V4.1-Flash-1017 are evaluated by us with each model's default sampling parameters, and scores cited from official reports are identified in the notes below. Every task is run once (pass@1).
Research
SciCode. All models are evaluated on the 64 main problems of SciCode with tool use disabled. We report main-problem accuracy, where a main problem counts as solved only if all of its subproblems pass, followed by the subproblem pass rate.
FrontierScience. All models are evaluated on the 60 problems of the FrontierScience Research set with web tools enabled. Responses are graded against the official rubrics.
BioMysteryBench. We report the Human-solvable (73 tasks) and Human-difficult (17 tasks) subsets. All models are evaluated with web tools disabled and a 200-turn limit. Answers are graded against the task rubrics.
PaperBench. All models are evaluated on the PaperBench Code-Dev split: 20 papers, 4 of which require Hugging Face access, with web tools enabled. Submissions are graded against the official rubric trees using the official grading prompts. We report the mean replication score over all 20 papers.
Agentic Tasks
BrowseComp. All models are evaluated on the full 1,266-task set with the same web-search setup across models, relying on Claude Code's built-in context compaction. We report task-level accuracy.
WideSearch. All models are evaluated on all 200 tasks (100 English and 100 Chinese) with the same web-search setup across models. We report average item-level F1.
BFCL V4. All models are evaluated with the official BFCL v4 evaluation pipeline (gorilla), using native function calling instead of the Claude Code harness. Scoring uses the official checker over all 5,106 test entries, covering the non-live, live, multi-turn, and agentic (web search and memory) categories. We report the unweighted pass rate across all entries rather than the official weighted overall score.
APEX-Agents. All models are evaluated on all 240 tasks of APEX-Agents v1.1, using the task-declared time limit. Web tools are disabled; agents interact only with the provided task environments. We report the fraction of tasks with a reward greater than 0.
SWE-bench Pro v2. All models are evaluated on the full public set of 642 tasks (v2.0.0) with web tools disabled. A task is resolved when the official test suite passes.
SWE-bench Verified. All models are evaluated on all 500 tasks. The agent sandbox has no internet access. A task is resolved when the official FAIL_TO_PASS and PASS_TO_PASS tests pass.
LiveCodeBench v6. All models are evaluated on the 1,055 problems of the v6 release without internet access. A problem counts as solved only if the submitted solution passes all hidden test cases within the task's time limit.
SWE-Atlas QnA. All models are evaluated on all 124 tasks with web tools disabled.
Reasoning
MMLU-Pro. All models are evaluated on the full 12,032-question test set (10 options per question) with a 5-shot prompt and web tools disabled. The answer is extracted from the model's final response, and we report accuracy.
AA-LCR. All models are evaluated on all 100 AA-LCR v1.1 tasks with web tools and other external tools disabled. We report accuracy.
GPQA Diamond. All models are evaluated on the 198 questions with web tools disabled.
Internal Benchmarks
RoboRSI. We evaluate robotic coding agents in two simulation environments covering task execution and trajectory collection. RoboDojo employs the RoboAgent framework and includes tasks such as sorting, stacking, pouring, and pushing to assess agents' ability to achieve goals via robot operation interfaces and control code. RoboTwin encompasses tasks including pick-and-place, bimanual handovers, stacking, sorting, and container manipulation, requiring agents to autonomously develop and validate control policies, complete operations, and export valid trajectories. All models use the Claude Code harness with a shared vision MCP tool, performing planning, coding, and iterative debugging within prescribed time and resource budgets.
Real2Sim. We evaluate models on a suite of tasks requiring autonomous reconstruction of loadable, physically interactive MuJoCo scenes from single indoor photographs. All models use the Claude Code harness and obtain visual information through a shared vision MCP tool, with web tools disabled and no access to hidden ground truth. Evaluation executes the model-submitted build script in a clean environment and assesses the reconstructed scene's joints, objects, layout, and appearance, while accounting for physical validity, completeness, and scale error.
CFD-Bench. We evaluate models on computational fluid dynamics and related numerical optimization tasks. All models are evaluated with the Claude Code harness in isolated environments, developing and refining solutions within predefined resource and time budgets. We report wins / draws / losses against Qwen3.6-35B-A3B.
