@cline: We’re about to see a breakout in local model capabilities. Flint scores 54.9% on Terminal-Bench 2.1 with ~2B active par…
Summary
Flint scores 54.9% on Terminal-Bench 2.1 with only ~2B active parameters, outperforming other small MoE models and some dense ~30B models while running at 125+ tok/s on a MacBook Pro with ~24GB memory.
View Cached Full Text
Cached at: 10/21/26, 06:35 PM
We’re about to see a breakout in local model capabilities.
Flint scores 54.9% on Terminal-Bench 2.1 with ~2B active parameters. That’s ahead of every other small MoE and two of the three dense ~30B models.
It only needs about 24GB and runs at 125+ tok/s on a MacBook Pro. https://t.co/9QQFQjgtN4
Similar Articles
Recursive releases Flint
Recursive 发布了 Flint，一款 40B 总参数、仅激活 2.5B 参数的 MoE 本地 agentic 编码模型，在 50B 以下 MoE 模型中于 SWE-Bench Pro v2、Terminal-Bench 2.1 等基准上领先，并借助自研 Eureka Machine 自动化研发系统在四个月内完成构建。
Benchmark: Recursive's Flint 40B A2.5B Q4 achieves 128 tok/s on MacBook Pro M4 Pro 48GB using Kapitsa, an open-source Apple Silicon optimized engine, Claude/Codex/Pi/OpenCode compatible.
Kapitsa, an open-source inference engine optimized for Apple Silicon, achieves 128 tok/s decode on a MacBook Pro M4 Pro 48GB running Recursive's new Flint 40B Q4 model, outperforming llama.cpp and Recursive's own engine in long-context decode and prefill benchmarks.
Recursive launches Flint, a laptop coding model it says was largely built by AI
Recursive released Flint, a free 40B-parameter mixture-of-experts coding model that runs on laptops, claiming it was built largely by its automated AI research system, the Eureka Machine. Flint achieves high token throughput and strong SWE-Bench Verified results for its size, and is available under the Apache 2.0 license on Hugging Face.
@cline: GLM-5.3 (max) outperforms GPT-5.6 Sol (max) on the new Terminal-Bench 4.0. Incredible seeing open weights compete with …
GLM-5.3 (max) outperforms GPT-5.6 Sol (max) on Terminal-Bench 4.0, highlighting the competitiveness of open-weight AI models, with Cline promoted for discounted access.
@VictorKaiWang1: 95.3% on Terminal Bench 2.1 Deepseek V4 Flash + StateM, 88.8% on TB2.1
Researchers achieve 95.3% accuracy on Terminal-Bench 2.1 using DeepSeek V4 Flash and StateM, matching GPT-5.6 Sol Max performance and exploring agent improvement beyond model scaling.