flint:40b-a2b-qat-int4

879 17 hours ago

Flint is a 39.8 billion-parameter mixture-of-experts model from Recursive, built for coding agents.

tools thinking 40b 
ollama run flint:40b-a2b-qat-int4

Details

17 hours ago

a503b9032955 · 23GB

json
{ "bos_token_id": 150000, "eos_token_id": [ 150001, 150003, 150005 ], "pad_token_id": 150001, "do_sa
json
{ "additional_special_tokens": [ { "content": "<|system|>", "lstrip": false, "normalized": false, "r
json
{ "version": "1.0", "truncation": null, "padding": null, "added_tokens": [ { "id": 150000, "content"
json
{ "add_bos_token": false, "add_eos_token": false, "added_tokens_decoder": { "150000": { "content": "
license
Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR US
params
{ "temperature": 1, "top_k": 20, "top_p": 0.95 }
json
{ "architectures": [ "FlintForCausalLM" ], "attention_bias": false, "attention_dropout": 0.0, "atten
template
{{ .Prompt }}
779 tensors

Readme

Flint logo

Flint is Recursive’s open model for local coding agents. Give it a repository task—investigating a bug, implementing a feature, or writing tests—and it can reason through the work and use tools across multiple turns. It is built for development workflows that move from understanding code to making changes. Learn more about Flint

Flint has 39.8 billion total parameters, with approximately 2.5 billion active per token. Its 262K-token context provides room for substantial code and task history. Quantized versions are available for Apple Silicon and other GPUs hardware, and the model is released under Apache 2.0.

Coding benchmarks

In Recursive’s release evaluations, Flint scored 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified. These are publisher-reported results; evaluation setups and runtimes can affect scores. Source: Recursive’s evaluation

Recursive’s published BF16 and NVFP4 benchmark comparison.

Model details and evaluation methodology