Strata-Qwen3.8-Flash-Next-512k-v100x2
Qwen3.8-Flash-Next 125B MoE on 2x Tesla V100 32G: 512K context, vision, 64-79 tok/s decode, 1.8K tok/s prefill. Build, memory arithmetic, tuning and measured results.
- 8
- Shell
- Apache-2.0
- 6 wk ago
ZackO2o
Hello World .
Built from ZackO2o's public GitHub profile, HongKong. Not on Take My Repo yet, so there are no posts or updates here.
Qwen3.8-Flash-Next 125B MoE on 2x Tesla V100 32G: 512K context, vision, 64-79 tok/s decode, 1.8K tok/s prefill. Build, memory arithmetic, tuning and measured results.
Measured serving recipes for DeepSeek-V4.1-Flash on 4x NVIDIA DGX Spark (GB10): 1M context on vLLM (CUDA graphs, vision, tools, DSpark) and a switchless-ring SGLang TP4 lane, plus a cross-project reference table. EN + 中文.
Measured serving recipe for DeepSeek-V4.1-Flash on 2x NVIDIA DGX Spark (GB10) with the TensorFold engine, TP=2: 512K context x 3 slots, FP8 KV pool, exact speculative decoding. Fit ladder, pool-vs-context failure mode, client-side vs engine-internal benchmark methodology. EN + 中文, MIT.
Three-layer anti-amnesia memory system for OpenClaw. SQLite FTS5, BM25 + CJK bigram search, temporal decay. 0-token search cost. No embedding API needed.
Practical UI/UX design principles skill for OpenClaw agents. Covers visual hierarchy, color systems, typography, micro-interactions, UX laws, conversion optimization, 99 anti-patterns, and 60+ UI style systems.
ALL-Relay-Service - 全模型API中转服务 / Universal AI API Relay Service
Apply Will Guidara's world-class hospitality principles to AI product design & development. An AgentSkill for OpenClaw.