Techniquesoptimizationtraining precision

implementation detail · filed under optimization

FP8-precision reinforcement learning

The reinforcement-learning portion of training runs in FP8 precision.

Also called RL was done in FP8 precision.

sources: 2
model: 1
lab adopt it: 1
strongest: used

How sources treat it

One count per evidence span, weakest treatment to strongest.

used 2

Documented in

Evidence

2 spans quoted from the sources, strongest treatment first.

It is the first Poolside model where reinforcement learning ran in FP8 precision.

usedpost trainingin Laguna S 2.1Poolside

Poolside Releases Laguna S 2.1, an Open-Weight Agentic Coding Model Punching Above Its Weight Class on SWE-Bench Multilingual·MarkTechPost·2026-07-21

S 2.1 is also our first model where RL was done in FP8 precision, accelerating that part of the training.

usedsoftware implementationin Laguna S 2.1Poolside

Introducing Laguna S 2.1·Poolside·2026-04-28

Filed alongside

Other methods under optimization :: training precision.

NVFP4 BF16 NVFP4 pre-training FP8 mixed-precision training FP4+FP8 mixed precision MXFP8 BF16 gradient reduction BF16 mixed-precision training Block-wise FP8 activation quantization with offload E2M1 FP32 attention-output retention FP32 gradient reduction FP8 storage for the residual state High-precision final network layers Mixed-FP8 quantization Mixed-precision training NVFP4 fine-grained micro-block scaling Two-dimensional block quantization