step-3.5-flash

The flagship language reasoning model from Jueyue Xingchen boasts top-tier reasoning capabilities and fast, reliable execution. It can decompose and plan complex tasks, quickly and reliably invoke tools to carry out tasks, and excel at a wide range of challenging assignments—including logical reasoning, mathematics, software engineering, and deep research. With a context length of 256K, its core design is a “light-activation, high-density, efficient agent engine” specifically tailored for real-time interaction, complex reasoning, and coding tasks.

• Sparse and Efficient Architecture: Based on an MoE model with 196B total parameters, this model activates only 11B parameters per token, striking the ultimate balance between large-model capabilities and small-model speed.
• Ultra-High-Speed Generation Engine: Featuring 3-path multi-token prediction (MTP-3), it achieves speeds of 100–300 tokens/s in typical scenarios and peaks at 350 tokens/s in single-stream encoding, delivering near-real-time responses.
• Leading Agent Performance: It excels in authoritative benchmarks such as SWE-bench Verified (74.4%) and Terminal-Bench 2.0 (51.0%), demonstrating exceptional stability and reliability.
• Long-Context Optimization: Supporting a 256K context length, it employs a 3:1 sliding-window attention (SWA) mechanism, significantly reducing computational overhead without sacrificing performance.
• Locally Friendly Deployment: It can run on consumer-grade high-end hardware such as Mac Studio M4 Max and NVIDIA DGX Spark, ensuring data privacy and low latency.

───────────────────────────────────────────────────────────────────

Core Capabilities

⚡ Ultra-Fast Reasoning Engine:

• The MTP-3 technology performs forward prediction on four tokens at once, dramatically accelerating generation and making encoding tasks lightning-fast.
• With just 11B activated parameters, its inference cost is about 1/6 to 1/18 that of comparable MoE models, offering outstanding cost-effectiveness.

🧰 Professional Agent Foundation:

• It integrates a scalable RL framework that supports continuous self-improvement and excels at long-term, multi-step tasks.
• It leads comprehensively in Chinese and English reasoning benchmarks including BrowseComp-ZH (73.7%), GAIA (84.5%), and AIME 2025 (97.3%).

🧠 High-Density Intelligence:

• A fine-grained routing system featuring 288 expert specialists plus 1 shared expert retains the “memory” of the 196B model while achieving execution efficiency comparable to an 11B model.
• It supports the Parallel Thinking mechanism, further enhancing performance on tasks like xbench-DeepSearch.

───────────────────────────────────────────────────────────────────

Test Data