Mustapha

Tweets on AI news: local models vs. frontier • Agents • VRAM + power bills. Skeptical notes, not launch-day hype.

[Mustapha @ElouardiMusta]

@BelfkihHC Nas d xiaomi li 3endhom had mochkil

[Mustapha @ElouardiMusta]

DeepSeek V4.1 Flash just scored 81.2/100 on OpenDesign Arena GPT-6 Astra got 82.7

Cost per design: $0.023 vs $1.61 (70x cheaper)
Speed: 2x faster

Near-frontier design performance at commodity pricing just arrived

https://t.co/KR5C21Er8B

[Mustapha @ElouardiMusta]

Qwen-Image-2.1 drops today 🚀

One model does it all text-to-image, image-to-image editing, and precise control all in a single 27B-param package
✨ Native 2K resolution
✨ Pro-grade typography (English + Chinese)
✨ Alpha channel support
✨ Style transfer, object add/remove, pose manipulation
Open weights. On ComfyUI day one
https://t.co/4RaaOi0VsU

[Mustapha @ElouardiMusta]

jev

[Elon Musk @elonmusk]

This is the future we shall bring into being

[Mustapha @ElouardiMusta]

Mind-blowing. Leap to the Stars Step 5 preview: long-context, faster and better and it ties Opus 5 on AIME24.

It lifts Qwen3-30B-A3B from 53.3% to 60% on AIME24, matching Claude Opus 5 with fewer labeled tokens.

Long-context recap:

MLA-512 GPU kernel: 22 hours to 508 TFLOPS, beating Opus 5’s 493

Pokémon Red: no special opts, 3,000+ consecutive turns, 6M tokens about one-third of the main story

Closed-loop post-training: it tweaks itself, tests itself, and decides whether to keep going. This round, long-context isn’t just done. It’s done better.

[Mustapha @ElouardiMusta]

🚀 China just dropped a new Pareto frontier model

StepFun's Step 5 Preview is live a 600B MoE with only 27B active params and a 1M context window

Why it matters:
- Matches Kimi K3 (max) at 44 on Artificial Analysis
- 65% lower cost per task same budget buys 3× the runs
- Built for agentic work: software engineering, long-horizon execution, finance
- API available now | Open weights Oct 15

China keeps shipping. The cost/intelligence curve just bent again

🔗 Try it: https://t.co/5n8dUyfNCb

[Mustapha @ElouardiMusta]

Hermes agents just got a "second brain" for routine decisions.

offloads model routing, skill selection, compaction, and computer/browser actions to Jev (TypeSafe's decision model, ~0.4s per call)

Key skills:
• Model routing picks the right model per turn from all your providers
• Skill selection suggests the right skill from 373+ installed (~0.9s)
• Compaction decides what survives verbatim vs summarized (71 turns in 0.95s)
• Computer/browser use next GUI/page action from pre-judged safe actions
• Dashboard (jev dashboard) live routing view per profile

Install: unzip && python3 https://t.co/Y4OIP4oc4B && jev setup-key
Try shadow mode first: /jev routing shadow
https://t.co/F5gyCcJJIJ

[Mustapha @ElouardiMusta]

@DanielSmidstrup Jean is an opinionated native desktop app built with Tauri that gives you a powerful interface for working with Claude CLI, Codex CLI, Cursor CLI, OpenCode, PI, Command Code, Grok, and Kimi Code https://t.co/zL1f3eWZSQ

[Mustapha @ElouardiMusta]

Kimi_Moonshot posted a string of digits on Zhihu: 415926535897932384626433832795...
- This is π (3.14159...) with the leading 3.1 removed
- MaxForAI interprets this as hinting at Kimi K3.1 (since removing "3.1" from π leaves the rest)
- Current flagship is still K3 this would be a minor version bump

[Max For AI @MaxForAI]

🚨 K3.1 要来了

今天早上 @Kimi_Moonshot 官方在 @ZhihuFrontier 突然发了一串完全没解释的数字：

415926535897932384626433832795...

一开始看很莫名其妙，但答案其实就在圆周率里。

π 是：

3.1415926535897932384626433832795...

把最前面的 3.1 删掉：

415926535897932384626433832795...

刚好就是 Kimi 发出来的这一串。

也就是说，这个谜语很可能只有四个字：

Kimi K3.1。

目前 Kimi 官方还没正式公布 K3.1，官网和文档里的旗舰模型仍然是 K3。

但大半夜专门发一个「少了 3.1 的 π」，这暗示已经有点明显了😂

[Mustapha @ElouardiMusta]

Generative UI with jev is freaking fast

[Chris Tate @ctatedev]

New experiment: json-render + jev

The future Generative UI is instant

Your components, your actions, your design system

Rendered in milliseconds

[Mustapha @ElouardiMusta]

every boring computer task gets its own tiny model next: forms, CRM, data entry, browser clicks. one agent routes between them. ridiculous stuff incoming.

[Cua @trycua]

1/ Introducing CUA-S1: a family of System One Models, small, specialized, and built for computer use.

Today we're open-sourcing CUA-S1-FORMS, the first in the family: https://t.co/J1frbEbZTQ

[Mustapha @ElouardiMusta]

@venturetwins Framework hell, tutorial hell, now model hell

[Mustapha @ElouardiMusta]

@thdxr @opencode Name it opencode dax tier

[Mustapha @ElouardiMusta]

MiniMax just open-sourced MiniMax Code ⚡

Terminal coding agent, MIT licensed.
Bring your own model or use MiniMax, with search, plugins + multimodal tools.

Repo: https://t.co/126imNmTy9

[Mustapha @ElouardiMusta]

Try it
https://t.co/sDnHoolrVJ

[Mustapha @ElouardiMusta]

Cloudflare just open-sourced how they find vulns at fleet scale.

It's called security-audit-skill. 11.2k stars, 600+ forks.

And it's not another scanner — it's a prompt harness that turns Claude/Codex/etc into an audit team.

Here's how it works:

[Mustapha @ElouardiMusta]

This seeded their real VDH that scans 128 repos across Rust, Go, C, TS, Python.
DB-backed, resumable, deduped, cross-repo tracing.
Took 6 weeks to go from skill to fleet scanner.

Start with the skill, build harness only when you need it.