Today for AI
HOT RADAR
toolsHEAT 0.0°

AI CURATED HOT STORY · 10/21/2026

Recursive Launches Flint Model with Day-0 vLLM Support

6 reports archived0 independent sourcesupdated 10/22/2026, 22:40:16
Synthesis & Latest Updates

AI startup Recursive has launched Flint, its first open-source coding model built by the automated research system Eureka Machine in under four months. This 40B Mixture-of-Experts (MoE) model, licensed under Apache 2.0, validates the potential of AI-assisted development while remaining under human supervision.

Cross-referenced data highlights Flint's exceptional efficiency: despite 40B total parameters, it activates only ~2.5B per token. It achieves a 74.2% score on SWE-bench Verified and maintains high quality while delivering speeds of 125+ tokens/s on a 48GB MacBook Pro—7-10x faster than comparable models like Qwen3.8. On a single GB200 GPU, decoding reaches 500 tokens/s with NVFP4 quantization, supporting a 262K context window.

The release significantly impacts the developer ecosystem. vLLM provided Day-0 support via dedicated Docker images, enabling immediate deployment. Flint sets a new Pareto frontier for local coding models, offering high-speed, privacy-preserving code generation and accelerating the shift toward efficient, lightweight AI infrastructure.

LATEST/vLLM officially announced Day-0 support for Flint, enabling rapid deployment via dedicated Docker images. Additionally, OpenRouter has listed the model for free access, allowing users to test its native tool calling and long-context capabilities.

Hourly heat curve

168 fully observed hours
Now
0.0
Peak
65.423:00
24h change
–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 6 reports · Latest first
  1. The DecoderT2·78 pts

    Recursive Releases Flint: First MoE Coding Model Built by AI

    • Flint is the first result from Eureka Machine's autonomous proposal, build, and evaluation loop, developed in four months
    • Uses a Mixture-of-Experts architecture with 40B total parameters but only 2.5B active per token, optimized for local agentic coding
    • Claims superior performance on SWE-Bench and other coding benchmarks among sub-50B MoE models, though results are unverified
  2. 机器之心 (微信公众号)T2·88 pts

    Yuan-Tian's New Firm Releases Flint-40B-A2B, Built Entirely by AI

    • Flint-40B-A2B is a 40B total parameter (2B active) MoE model designed specifically for agentic coding tasks.
    • The model was developed in under 4 months with full participation from Recursive's proprietary automated research system, Eureka Machine.
    • It outperforms similar-sized MoE models on benchmarks like SWE-Bench Pro v2 and achieves over 125 tokens/s generation speed on an M4 Pro MacBook.
  3. OpenRouter (@OpenRouter)T1·85 pts

    Recursive Launches Flint: 40B MoE Coding Model on OpenRouter

    • Flint is Recursive's first model, featuring a 40B total parameter MoE architecture (~2.5B active) optimized for coding agents.
    • Built from scratch in under 4 months by the automated research system Eureka Machine, it supports a 262K context window and native tool calling.
    • The model demonstrates strong performance on benchmarks like Terminal-Bench 2.1 and SWE-bench Verified, while offering fast local inference options.
  4. TechCrunch AIT2·85 pts

    Recursive Launches Flint: AI-Assisted On-Device Coding Model

    • Flint is a free, locally runnable coding model developed by a small team in under four months with AI assistance.
    • The key innovation is the use of the 'Eureka Machine' automated research system to accelerate model development, validating AI-assisted R&D.
    • Current iterations remain under human supervision, not yet achieving fully autonomous recursive self-improvement.
  5. Rohan Paul (@rohanpaul_ai)T2·88 pts

    Recursive Flint Launches: 40B MoE Coding Model for Fast Local Inference

    • Flint is a 40B MoE model built by the Eureka Machine's automated research loop in under 4 months.
    • It runs at 125+ tokens/s on a 48GB MacBook, offering ~7-10x speed improvement over comparable quality models like Qwen3.8 while maintaining ~81% average benchmark scores.
    • Weights are open-sourced on Hugging Face under Apache 2.0, enabling fully local, unmetered, and private inference.