Models Published:
Recursive releases Flint, an Apache 2.0 coding model with 40B parameters and about 2.5B active, already in Ollama for Mac
What happened
Recursive released Flint on October 21, a mixture-of-experts model for coding agents with 39.8B total parameters and about 2.5B active per token, trained from scratch and published on Hugging Face under the Apache 2.0 licence. It supports reasoning and multi-turn tool use, with a documented 262,144-token context, and Recursive reports 54.9% on Terminal-Bench 2.1 and 74.2% on SWE-bench Verified. Recursive says its NVFP4 version scores within 0.7 points of the 16-bit model and runs at more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64GB RTX Spark at short contexts. Ollama v0.40.8 added it the same day as `ollama run flint`, on macOS only for now.
Why it matters locally: A coding agent model with only about 2.5B active parameters and a permissive licence can run quickly on a single laptop, so source code never has to leave the company.
Three ways to read it
For the market
Flint enters the space held by open coding models such as Qwen3-Coder-30B-A3B and Devstral Small 2, and Recursive says it rivals dense 25–40B models while running more than five times faster. A newcomer shipping day-one support in Ollama and tuned kernels for both Apple Silicon and RTX Spark suggests local coding agents are becoming a hardware selling point, though its benchmark lead still needs independent confirmation.
In plain words
Flint is an AI helper for programmers that reads a codebase, finds bugs and writes fixes and tests. It is large in total, but only a small slice of it works on each word, a bit like a big team where only the few relevant specialists answer each question, so it stays fast on a laptop.
For practitioners
The BF16 checkpoint is about 80 GB and Recursive's SGLang kit targets one B300 or H200 per replica; on a Mac or RTX Spark, use the NVFP4 or other quantized builds, and note Ollama's GPU support outside macOS comes in a later release. Only 14 of the 49 layers use full attention (the other 35 use a 513-token sliding window), which keeps the cache small at long context. Use the recommended sampling (temperature 1.0, top-p 0.95, top-k 20), keep `reasoning_content` in the history, and budget at least 32,768 output tokens for hard tasks.
Sprout's insight
Sprout compares this with other brands and with where real applications are heading. Free for members.
Sign in free to read it →Subscribe to Sprout's weekly
One email a week: the top stories and a taste of Sprout’s insights. Unsubscribe any time.
Source
Read the official source: Introducing Recursive Flint (Recursive) ↗Summaries are written by OnPrem World from the official source. We link to the original and never republish it.
On this site
More news
- AMD and Intel agree on oneDNN as PyTorch's shared x86 CPU acceleration library
- vLLM v0.32 brings 4-bit KV cache to more GPUs and stops allowing every CORS origin by default
- DeepSeek-V4.1-Flash-1017: a same-size update with better agent scores and shorter answers
- Open WebUI 0.12.3 adds end-to-end encrypted chats and one search box for the whole workspace
- llama.cpp v0.7.0 adds a GPU cache for offloaded MoE experts and moves the server to port 9931