Macfax
macfax.bsky.social
https://macfax.com - Sell with proof. Buy with confidence. A tamper-proof report that replaces the guesswork before money changes hands - issued by Macfax, open to anyone, impossible to forge.
👤 Joined on 09 May 2026
4
Followers
0
Following
89
Total Posts
📜 Kullanıcı Geçmişi
- 09 May 2026, 05:16
🎉 Bluesky'e katıldı - İlk kullanıcı adı:
@macfax.bsky.social
🚫 Blokladığı Kullanıcılar
❤️ Beğendiği Gönderiler
Long-context models are coming to our phones and laptops. Sevren (the new kid on the European block) ported FlashPrefill V2 to MLX, bringing block-sparse attention to Apple silicon giving a 1.78X speedup on 32K tokens. OS code: github.com/sevren-ai/Fl... Blog post: sevren.ai/blog/flashpr...
Show Post
Run Claude Code entirely on-device using an MLX-native server on Apple Silicon to execute local AI models without API keys.
Show Post
Increasingly less quantised. And it's not just theoretical (even just with MLX versions), you could deploy it within the next ten minutes. With OpenCode, not just the circa-2023 chat interface against which willfully-uninformed critics tilt. Unified memory was the right move. On DGX Spark too.
Show Post
■検証の背景と目的 ・LLMの実運用では高速な応答性が求められるため、推論速度の最適化が重要になる。 ・mlx-dsparkはApple Silicon向けMLXフレームワーク上でSpeculative Decodingを利用し、デコード処理を加速させるためのライブラリである。 ■主要な性能改善点 ・DFlash 2を採用した結果、Generationモードにおいて生成速度が39.33 tok/sから61.53 tok/sへと約1.5倍に向上している。
Show Post
Qwen3.8-27BのMLX版をLM Studioで少し試してみた。だいたい14tok/sぐらい
Show Post