Skip to content
Drix10 Blog

Slotstream runs Qwen3.8-Flash-Next on 16-64GB Macs, 20.91-23.69 tok/s benchmark

, 7 items in AI Consulting and Expertise, 4 min read

In this digest (7 items)

Slotstream streams the 105GB Qwen3.8-Flash-Next model from SSD and runs on Apple Silicon Macs with 16GB to 64GB of RAM. A 48GB M5 Pro achieved 20.91–23.69 tokens per second on short, long, and coding prompts at a 33GB memory budget. The engine is a single Swift binary, works offline after download, and requires no Python.

Key points

  • Runs Qwen3.8-Flash-Next from SSD on 16–64GB Apple Silicon Macs, one Swift binary, offline after download.

  • 48GB M5 Pro measured 20.91–23.69 tok/s across short, longer, and coding workloads at a 33GB memory budget.

Sources

bbtree Rust map released, 5x faster than std BTreeMap

bbtree is a Rust ordered map released by Xiangpeng Hao, benchmarked 5× faster than std's BTreeMap on one thread. It also runs 12× faster than crossbeam-skiplist at 96 threads, with mixed memory results.

Key points

  • Release: bbtree released last week, LLM wrote all implementation, Hao made every design decision.

  • Benchmark: 5× faster than std BTreeMap on one thread, 12× faster than crossbeam-skiplist at 96 threads.

Sources

Microsoft Authenticator bug lets Android app steal Teams tokens July CVE-2026-80097

Microsoft Authenticator bug let an Android app with zero permissions pose as Teams and steal a user's work-account tokens. In Enclave's tests the stolen tokens opened the victim's mailbox, files and calendar with no prompt. Microsoft fixed the issue in July and assigned CVE-2026-80097.

Key points

  • Bug allowed an Android app with zero permissions to impersonate Teams and obtain work-account tokens.

  • Tokens accessed the victim's mailbox, files and calendar without any prompt.

  • Fix released in July, identified as CVE-2026-80097.

Sources

BrowserCode launch – free, no-account dev environment

BrowserCode is a full dev environment in a browser tab, launched today, free and requires no account. It lets users share running projects instantly and experiment with agents in an isolated, safe playground.

Key points

  • Launch: Today we launch BrowserCode, a full dev environment in a browser tab.

  • Access: No account is required, and it's completely free.

Sources

Cloudflare adds 'Onboard your agent' button to dashboard

Cloudflare adds "Onboard your agent to Cloudflare" button on dashboard that copies a setup prompt for AI coding tools. It notes SaaS owners could add the same button.

Key points

  • Feature: "Onboard your agent to Cloudflare" appears first on dashboard.

  • Action: One click copies a setup prompt for your AI coding tool.

Sources

CapCut plugin for ChatGPT announced at CapCut World LA

CapCut announced a CapCut plugin for ChatGPT at CapCut World LA, enabling draft video creation directly in chat. Prompt format should be a shot list describing each beat, its duration, and music placement.

Key points

  • CapCut plugin is coming to ChatGPT.

  • Write prompts like a shot list: what each beat shows, how long it holds, and where the music lands.

Sources

Copilot chat free fails with spreadsheets, drops 25% of rows

Copilot chat (free) is now failing with spreadsheets. Users report it randomly drops about 25% of rows on small sheets.

Key points

  • Issue: Copilot chat (free) now fails miserably with spreadsheets.

  • Failure mode: Randomly dropping 25% of rows on small sheets.

Sources