Flash Next
A large language model nominally comprising 125B parameters, built on a Mixture-of-Experts (MoE) architecture in which approximately 121B parameters constitute the MoE component. The model supports deployment with Q4 and Q2K quantization.
Timeline (last 90 days)
According to feedback from multiple users in the Reddit community, the Qwen4 family already carries a great deal of expectation before release, and Flash Next, as an "advance guard" version, has successfully paved the emotional road for subsequent products.
The Reddit community has recently been discussing a model called "Flash Next," with users full of anticipation for the upcoming Qwen4 family.
Flash Next was called an "awesome model" by the poster, and multiple users replied expressing similar positive impressions.
One user described Flash Next's performance as a "quality meme," saying they hadn't expected it to generate such high-quality content.
One user mentioned that "every major version update is a huge leap," implying that this model series maintains a steady pace of progress across versions.
Of this 125B model's backbone, approximately 121B is the Mixture-of-Experts (MoE) portion, with an additional approximately 51B PLE lookup table and approximately 4B MTP module, bringing the complete checkpoint close to 180B.
The model has 512 experts per layer, but only 10 are activated per token, meaning the weights actually participating in computation are approximately 6B.
With all three optimizations stacked, generation speed stabilizes at 9 to 10 tokens per second, but time-to-first-token for a 10,000-character long document still requires approximately 329 seconds.
Running the 125B-parameter Flash Next model on an RTX 3090 (24GB VRAM) with 24GB RAM, generation speed is stable at 9 to 10 tokens per second, with a peak of 18 for short texts.
All Facts (9)
The model has 512 experts per layer, but only 10 are activated per token, meaning the weights actually participating in computation are approximately 6B.
60%UnverifiedFlash Next was called an "awesome model" by the poster, and multiple users replied expressing similar positive impressions.
50%UnverifiedOne user described Flash Next's performance as a "quality meme," saying they hadn't expected it to generate such high-quality content.
50%UnverifiedOne user mentioned that "every major version update is a huge leap," implying that this model series maintains a steady pace of progress across versions.
50%UnverifiedAccording to feedback from multiple users in the Reddit community, the Qwen4 family already carries a great deal of expectation before release, and Flash Next, as an "advance guard" version, has successfully paved the emotional road for subsequent products.
50%UnverifiedRunning the 125B-parameter Flash Next model on an RTX 3090 (24GB VRAM) with 24GB RAM, generation speed is stable at 9 to 10 tokens per second, with a peak of 18 for short texts.
50%UnverifiedOf this 125B model's backbone, approximately 121B is the Mixture-of-Experts (MoE) portion, with an additional approximately 51B PLE lookup table and approximately 4B MTP module, bringing the complete checkpoint close to 180B.
50%UnverifiedWith all three optimizations stacked, generation speed stabilizes at 9 to 10 tokens per second, but time-to-first-token for a 10,000-character long document still requires approximately 329 seconds.
50%UnverifiedThe Reddit community has recently been discussing a model called "Flash Next," with users full of anticipation for the upcoming Qwen4 family.
50%