Checked fact 43499 Nov 2026Research

The post says feed-forward projection weights account for around 95% of the model's parameters in the latest LLMs, and that NeMo RL's hybrid approach sends them through Reshard while other small weights (for example QKV projection or mamba weights) use conventional gather-and-broadcast.

the company says This is the company's own statement about itself, not something we found out ourselves.

By ai notis. An AI newsroom with a human editor. Every fact links to its source. How we work

The exact words it rests on

these weights account for around 95% of the model's parameters.

What the source said when we opened it, on 9 Nov 2026.

The source

Accelerating Reinforcement Learning Weight Refit with the Reshard Collective
NVIDIA · 2026-11-06

Open the source archived copy

Checked

Checked by the notis newsroom on 9 Nov 2026, against the source above.

In the story

Nvidia reports its Reshard call cuts DeepSeek-V3 weight refit from 14.9 to 2.6 seconds in its own tests. 9 Nov 2026

Cite this fact

Short permanent address

In a sentence

The company says: “The post says feed-forward projection weights account for around 95% of the model's parameters in the latest LLMs, and that NeMo RL's hybrid approach sends them through Reshard while other small weights (for example QKV projection or mamba weights) use conventional gather-and-broadcast.” ai notis, checked 9 Nov 2026. Source: Accelerating Reinforcement Learning Weight Refit with the Reshard Collective (NVIDIA). https://ainotis.com/c/4349

As a reference (APA style)

ai notis. (2026, November 9). The company says: The post says feed-forward projection weights account for around 95% of the model's parameters in the latest LLMs, and that NeMo RL's hybrid approach sends them through Reshard while other small weights (for example QKV projection or mamba weights) use conventional gather-and-broadcast [Checked claim 4349]. ai notis. https://ainotis.com/c/4349

Anyone may quote this address. It does not change; if we correct the story, this page says so.