Step 3.5 Flash
Step 3.5 Flash, developed by StepFunAI, is a large-scale foundation model with approximately 1.96 trillion parameters. Equipped with a wide 256K context window, it provides advanced chat functionality.
Parameters
1960.0B
Context Window
256K
License
Apache 2.0
Release Date
2026-02-02
API Pricing
API pricing for this model is not yet available
Strengths
- ・1.96 trillion massive parameter count
- ・256K long context support
- ・Apache 2.0 open license
Weaknesses
- ・Massive 399GB file size
- ・High hardware requirements
- ・Requires enormous computational resources
Use Cases
- ・Long document analysis and summarization
- ・Advanced conversational AI chatbot
- ・Large-scale data context processing
Deep Analysis
Architecture
Sparse MoE (196B total, 11B active)
3-way Multi-Token Prediction
Context Window
256K tokens
3:1 SWA ratio for efficiency
Output Speed
100-300 tok/s
Peaks at 350 tok/s for coding
SWE-Bench Verified
74.4%
AIME 2025
97.3
99.9 with PaCoRe
Average Score
81.0
Across 8 key benchmarks
Release Date
February 2026
License
Open-source
Strengths
- ・Exceptional intelligence density (11B active from 196B)
- ・74.4% SWE-Bench Verified
- ・97.3 AIME 2025 (99.9 with PaCoRe)
- ・100-300 tok/s generation speed
- ・256K efficient context with SWA
- ・Runs on consumer hardware (Mac Studio M4 Max)
- ・Open-source
Weaknesses
- ・Sparse MoE may have occasional routing issues
- ・Newer model with less community tooling
- ・196B total still requires significant VRAM for self-hosting
Competitor Comparison
| Model | Arena | SWE | GPQA | Price |
|---|---|---|---|---|
| GLM-4.7 355B | - | - | - | Higher |
| DeepSeek V3.2 671B | - | - | - | Comparable |
| Kimi K2.5 1T | - | - | - | Higher |
| Claude Opus 4.5 | - | - | - | Higher (closed) |
Step 3.5 Flash is StepFun's most capable open-source model with 196B total / 11B active parameters. It achieves 81.0 average score across 8 benchmarks, 74.4% SWE-Bench Verified, and 97.3 AIME 2025, rivaling models many times its size while running at 100-300 tok/s.
Analysis generated: 2026-05-24