Splash 1.4.0 hits 120 tok/s on M5 Pro with Qwen3.8-Flash-Next, plus open-source DFlash 2 drafter

HanGuo97 · x · 2026-10-22

• Inco AI shipped Splash 1.4.0: Qwen3.8-Flash-Next runs at up to 120 tok/s decode and 1.3K tok/s prefill at 32K on an M5 Pro (64GB) — over 30% faster prefill/decode than Strata, up to 1M context, bigger models in less memory.
• They also open-sourced DFlash 2 (Apache-2.0), a block-diffusion drafter for speculative decoding: predicts a whole block of tokens per pass, keeps top candidates at every position, and a lightweight selector traces a coherent path. Two-tap dynamic convolutions prevent draft decay near block end. Decoding is lossless — greedy output matches the target exactly.
• Repo includes SGLang launch commands and a target-feature capture patch (validated at TP1), with speculative config (DFLASH, 4 draft tokens). Evaluated against the official FP8 checkpoint.

Original post →

More from Infra

• Eden Deploys Vercel's Eve Agent to Cloudflare With One Command, Fully Open Source — ritakozlov · 2026-10-22
• Node.js ships emergency fix for high-severity Brotli decoder RCE in v22/24/26 — cramforce · 2026-10-22
• PyTorch's TorchTitan adds DeepSeek-V4, Kimi K3 support, RL training and big scale gains — PyTorch · 2026-10-22
• Epoch AI tracks 95 AI data centers with 14.2 GW — still just 44% of global AI compute — haider1 · 2026-10-22
• Two CVSS 9.2 RCE Vulnerabilities in PHP Push a European Government Site Offline — HankYeomans · 2026-10-22
• Old H100s Now Rent Above Launch Prices as Compute-per-Watt Value Keeps Climbing — abhiadesai · 2026-10-22

Infra channel →