Kimi K3: Open Frontier Intelligence · HackerPapers discussions

Kimi K3: Open Frontier Intelligence — mentioned by gpugreg

Fri, 31 Jul 2026 20:37:25 GMT

Notably, MXFP4 was introduced at the (much less costly) supervised fine-tuning stage after pretraining, so the number of B200/B300 GPUs could be relatively small in comparison to the number of H200 GPUs used during pretraining (or maybe not, who knows). (Kimi K3 tech report section 4.1.1 https://arxiv.org/pdf/2607.24653)

Paper and discussion history. A mention is not an endorsement.

Kimi K3: Open Frontier Intelligence — mentioned by gpugreg

Thu, 30 Jul 2026 10:23:54 GMT

From the Kimi K3 technical report: Kimi K3 supports a context window of up to 1 million tokens. We achieve this through extending the context window progressively as training proceeds, following a four-stage curriculum. The window grows from 8K to 64K tokens during pre-training, and from 256K to 1M tokens during the cooldown phase. https://arxiv.org/pdf/2607.24653 page 12

Paper and discussion history. A mention is not an endorsement.

Kimi K3: Open Frontier Intelligence — mentioned by janalsncm

Wed, 29 Jul 2026 22:54:41 GMT

Some companies are sharing their research. Mostly not the American ones Example: https://arxiv.org/abs/2607.24653 Exceptions to some like Deepmind, Nvidia, and Thinking Machines

Paper and discussion history. A mention is not an endorsement.

Kimi K3: Open Frontier Intelligence — mentioned by yorwba

Tue, 28 Jul 2026 17:42:02 GMT

The differences are explained in section 2.1.1 of the Kimi K3 technical report: https://arxiv.org/pdf/2607.24653#page=4

Paper and discussion history. A mention is not an endorsement.

Kimi K3: Open Frontier Intelligence — mentioned by senko

Tue, 28 Jul 2026 14:25:28 GMT

Old but relevant: if you read the recently-released Kimi K3 paper[0], you'll see that it's heavily based on Kimi Linear discussed here, scaling it up and adding a bunch more things (like native vision and RL improvements). [0] https://arxiv.org/abs/2607.24653

Paper and discussion history. A mention is not an endorsement.

Kimi K3: Open Frontier Intelligence — mentioned by throwa356262

Tue, 28 Jul 2026 13:38:55 GMT

For K3 read this instead: https://arxiv.org/abs/2607.24653 The main contribution of the K3 paper is Stable LatentMoE. Like some other models it compresses data sent between layers, which puts certain requirements on the router. K3 improves performance by using a more balanced expert selection strategy.

Paper and discussion history. A mention is not an endorsement.