Open-Weight Models Put VRAM Limits in Focus
A report examines several reported open-weight releases: Kimi K3.1 Pro with 2.8 trillion parameters and a 1 million-token context window, ERNIE 5.5 Flash Base with 280 billion total parameters and 14 billion active per token, Axion 1.0 in 35B and 122B sizes under Apache 2.0, and Ripple-65B. The release details were not independently confirmed in the supplied material. Using executed Python calculations, the report estimates that ERNIE 5.5 Flash Base requires 305.5 GiB at 32K context in FP8 and 644.7 GiB at 128K context in BF16, assuming one concurrent request, 15% overhead and 95% of each modeled 80 GiB card available. Under the same assumptions, Axion 1.0-35B is estimated at 40.9 GiB in FP8 at 32K context and fits on one card. These are modeled figures rather than hardware measurements. In a five-run median comparison on one web-form task, Flash recorded 1.1 seconds per step and 15.4 seconds overall, while Pro recorded 3.2 seconds per step and 16.0 seconds overall. The small first-hand test was not independently reproduced.
The report argues that model selection should weigh deployment feasibility and task performance rather than leaderboard position alone. Its calculations show how precision, context length and KV-cache memory can materially change hardware requirements, while its task example shows why step latency must be considered alongside step count and success rate.
开源模型挤着发了一圈,先别比榜单,算算显存
上周四到周五,开源模型挤着发了一圈。月之暗面放了 Kimi K3.1 Pro 的完整权重(2.8T,100 万上下文),百度开源文心 5.5 Flash Base(280B MoE,每 token 激活 14B),字节开源 Axion 1.0(35B / 122B,Apache 2.0,训练代码和数据一起给),还有 Ricursive 的 Ripple-65B。
选题会上有人问该上哪个。我说先别比榜单,先算能不能装下。
这篇就是算的过程,代码都跑过。
总参数和显存之间隔了两个东西
最常见的误解是拿总参数量直接换算显存。
#Moonshot AI#Kimi K3.1 Pro#Baidu#ERNIE 5.5 Flash Base#ByteDance#Axion 1.0#Ricursive#Ripple-65B