Ling-3.0-flash is 124B total and 5.1B active—the one-Spark discussion shows why both numbers matter
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
Ling-3.0-flash 模型被描述为“124B-A5B”，其总参数为 124B，但每个 token 仅激活 5.1B 参数。这个专家混合 (MoE) 模型有 512 个路由专家，每个 token 激活其中 8 个，这影响了其计算路径。这种区别至关重要，因为尽管在计算过程中只有 5.1B 参数处于活动状态，但机器仍需要存储所有 124B 参数。讨论强调了对于 MoE 模型，报告活动计算量、安装权重大小以及在指定硬件上的上下文/并发测量的重要性。
I first saw Ling-3.0-flash described as a “124B-A5B” model in an NVIDIA developer forum. It is a compelling headline, but the deployment discussion underneath it is a useful lesson in what “active parameters” does and does not mean.
The official specification is 124B total parameters and 5.1B activated per token. Its MoE has 512 routed experts and activates 8 of them per token. That helps explain the compute path. It does not mean the machine only needs to store 5.1B parameters.
The official single-DGX-Spark INT4 guide says the quantized weights occupy roughly 72 GB on a GB10 system with 121 GB of unified memory. The rest of the practical budget still has to absorb the runtime, KV and recurrent state, context length, concurrency, temporary allocations and the operating system.
That gives me a more useful way to read MoE headlines:
- Total parameters describe the model that must be represented in memory or storage.
- Active parameters describe how much of the routed network participates in each token.
- Quantization changes memory use and may change quality.
- Runtime and kernel support determine whether the theoretical efficiency appears on this hardware.
- Context and concurrency determine how much room is left after the weights load.
The forum thread showed all five layers interacting. The same checkpoint produced a retracted short benchmark, a better hard-mode score, long-prompt slowdown, an OOM report, and later more positive results after the software and quantization paths changed.
Would model releases be easier to evaluate if every MoE card reported three separate numbers up front: active compute, installed weight size, and measured context/concurrency on named hardware?