ローカルLLMではこれまでQwen3.8 27BやQwen3.6 35B A3Bを使っていましたが、精度がいまいちだったのと、5070Tiだとあまりコンテキストを載せられないので今ひとつ主力になりきれていませんでした。
Il progetto Strata permette di eseguire un modello da 125 miliardi di parametri su una RTX 5070 con 12 GB di memoria video, distribuendo i dati anche tra RAM e SSD. Il modello Qwen3.8-Flash-Next ...
Previously, I measured the speed of local LLMs using Ollama on a slightly older gaming PC equipped with a Ryzen 5 3600 and a GeForce RTX 3060 12GB. At that time, I tested Gemma4 E4B, Qwen3.5 9B, and ...
Hosted on MSN
Who is Kid Flash?⚡ #theflash #dccomics
Who is Kid Flash? | Origin Story of Wallace West Explained How did Wallace West go from troublemaker to being The Flashs sidekick? @MoZoneMo tells us everything in between as he shares Wallaces back ...
@daluoseo @grok Could i really use normal Qwen 3.8 Flash on rtx 3090 with 90+ t/s usually?
模型：MiMo-V2.6-Flash-RL（309B 总参数 / 15B 激活，GGUF MXFP4） 硬件：8×RTX 3080 20GB + 双路 Xeon 8269CY + 768GB 内存 速度：纯 decode 23.6 tok/s，上下文 256K 测试备注： · 定稿配置为 48 层中 40 层上卡，不开 MTP ...
メーカーで研究職をしています 最近は生成AIと米国経済に興味があります Zennで主に生成AI関連のブログを書いてます https ...
DeepSeek-V4.1-Flash, ready to serve on two RTX PRO 6000 Blackwell cards (sm_120, 96 GB each) with vLLM at tensor-parallel 2. The routed experts are coolbho3k's calibrated EXL3 3.0 bpw quantization.
Measurement project for Qwen3.8-Flash-Next (MoE, 512 experts × 48 layers) on one desktop: RTX 5060 Ti 16 GB (PCIe 4.0 x4) + RTX 4070 SUPER 12 GB (PCIe 4.0 x16, display card — keep ≥2.5 GB free) + ...
Reasoning, tool calling and image input work with this configuration. --quantization modelopt The checkpoint is an NVIDIA ModelOpt mixed-precision checkpoint (NVFP4 experts, FP8 MTP experts and ...
Some results have been hidden because they may be inaccessible to you
Show inaccessible results