Ripple publishes a dense 65B coding-model preview
Ricursive’s Ripple-65B Preview is a dense Transformer with published instruction-tuned weights in BF16 and NVFP4, plus a separate base checkpoint. Unlike a mixture-of-experts model, its dense design uses the full parameter network for inference. The company reports strong agentic-coding benchmarks and up to a million tokens of context, but those scores and deployment claims remain vendor results rather than independent validation.
Ricursive Intelligence
Why it matters
Available weights and architecture details let teams evaluate a different deployment trade-off instead of choosing only by a leaderboard score.
More context
The model card lists Apache-2.0 licensing and links the quantized checkpoint. Inspect serving compatibility and memory requirements before a small representative coding evaluation; hosted access is also listed by the vendor. The exact original publication date was not verified, so this card records a discovery rather than dating a new launch.
Keep in mind
This is a preview. Open weights do not cover inference hardware or hosted-service costs, and long-context quality or coding reliability on your workload was not established.
Original read