Meta Llama 4 Maverick: First Open Model to Reach Human-Parity on SWE-Bench Verified
[AI-NEWS-ENTRY]
Date: 2026-02-20
Title: Meta Llama 4 Maverick: First Open Model to Reach Human-Parity on SWE-Bench Verified
Content: Llama 4 Maverick (open-weight 405B variant) achieves 68.7% resolution rate on SWE-Bench Verified (matching or exceeding closed models like Claude 4 Opus and GPT-5.3), demonstrating production-ready autonomous software engineering capabilities in an open-source setting for the first time.
Keywords: SWE-Bench, autonomous coding, software engineering agents, Llama 4 Maverick, open-source frontier model, code generation
new
Mistral Devstral: Specialized 24B Model Optimized for Developer Tool-Chain Integration
Meta Llama 4 Maverick: First Open Model to Reach Human-Parity on SWE-Bench Verified
xAI Grok-4.1 Adds Native Video Understanding and Temporal Action Localization
DeepMind AlphaGeometry 2 Solves 92% of IMO Geometry Problems Autonomously
OpenAI o4-proto-2 Demonstrates Self-Evolving Prompt Engineering Loop
Mistral Codestral-Mamba: State-Space Model Achieves SOTA Efficiency on Long-Code Completion
OpenAI o4-proto: First Model to Demonstrate Self-Improving Code Synthesis Loop Over Multiple Iterations
Google DeepMind AlphaEvolve: Evolutionary Algorithm Guided by Gemini for Protein Engineering
xAI Grok-4 Early Preview: Real-Time Multi-Modal World Model with Physics-Aware Forecasting
Anthropic Claude 4.3 Sonnet Introduces Native Hierarchical Task Decomposition for Long-Horizon Agents