[KongchangAI]
ModelKunon 3.8 FlashNext

Kunon 3.8 FlashNext

An extremely sparse MoE large language model with a 125B-parameter backbone containing 24,576 experts, activating only about 6B parameters per token, suited to local inference under low VRAM.

Timeline (last 90 days)

Nov 4

The Kunon 3.8 FlashNext model backbone has 125B parameters, with a total of 24,576 experts; of the 512 routed experts per layer, only 10 are activated per token, so the number of parameters actually involved in computation per pass is about 6B.

Unverified50%
Nov 4

The Kunon 3.8 FlashNext model comes with an Ngram table of about 51B parameters.

Unverified50%
Nov 4

When using Strata to run Kunon 3.8 FlashNext, it is recommended to use the IQ2 or IQ3 quantization level, prepare 48GB to 64GB of system memory, and leave 70GB to 80GB of NVMe SSD space.

Unverified50%

All Facts (3)

Source Articles