Qwen3.8-Flash-Next -> 79.95 tok/s. 🔥 ONE DGX Spark. Full 262K cache configured. 🚀 Qwen3.8-Flash-Next EXL3 just got another major update. New measured default: MTP ndt=5 DSpark, dc=0.6 8-bit KV 262,144-token cache At an actual 240K-token prompt: 72.0 tok/s decode ~1,150 tok/s prefill Exact needle retrieval 𝗙𝗣𝟭𝟲 𝗞𝗩 → 𝟴-𝗕𝗜𝗧 𝗞𝗩 4K context: 66.6 → 69.8 tok/s 128K: 66.6 → 69.5 tok/s 240K: 65.1 → 72.0 tok/s 8-bit KV wins more as context grows, exactly as the memory math predicted. `EXL3_GR_INT8`, the int8 hyperconnection-mixer path, is now default-on in my ExLlamaV3 fork. PR #3 merged into master at `523ecd3`. And the real context ceiling is the MODEL, not the Spark. Caches up to 1,048,576 tokens load and decode, but Qwen’s trained window ends at 262,144. Needle retrieval is exact at 32K, 128K and 240K, then fails consistently at 300K+ at both KV precisions. One important correction: the earlier 103.8 tok/s result was a 32-token artifact. The defensible long-context result is 72 tok/s at 240K. Recipe PR #5 is also merged with the updated context findings, reproducible harnesses and raw logs under `tuning/`. Engine PR: github.com/vcruz305/exllamav… Updated recipe: github.com/vcruz305/Qwen3.8-… @Alibaba_Qwen @QwenDevs @turboderp_

BREAKING: Mike Pence officially announces he has "no plans" to run for president in 2028.

Confronted by political pressure, companies can behave like a swarm of AI bots. Credit to Meta CEO Mark Zuckerberg for injecting a note of common sense into the AI debate. wsj.com/opinion/mark-zuckerb… via @WSJopinion