Tarkhan

Local LLMs on Mac (MLX) · agentic coding field notes · building Snitt

[Tarkhan @otarkhan94]

I built Snitt — a Mac app that learns how you edit from your published videos, then rough-cuts your raw footage the same way.

Your pause lengths. The fillers you keep. Your pacing. Then it hands you a real Final Cut timeline, every cut restorable.

Free beta — 10 hours of processing included. Link below 👇

[Tarkhan @otarkhan94]

@grok @Grok
but Codex cli is opensource, I can use its prompts and create a similar environment in PI

[Tarkhan @otarkhan94]

@Grok can you answer this?

[Tarkhan @otarkhan94]

@Grok can you answer this?

[Tarkhan @otarkhan94]

Big improvement!! Did they forget to finish their post-training before launching the previous version? xD

[Qwen @Alibaba_Qwen]

🚀Qwen3.8-Max just got upgraded. Meet Qwen3.8-Max-0902!

2.4T parameters. 1M context tokens. Built for real world complexity.

Further post trained on Coding & Cowork, Qwen3.8-Max-0902 now delivers stronger performance across complex enterprise tasks, scientific research, and long horizon workflows.

💰Pricing per 1M tokens:
$2 input, $6 output.
$0.17 explicit cache hit, $0.25 implicit cache hit.

Now live via API on QwenCloud. Come try it! 🙌
API: https://t.co/dq3WgMk980

[Tarkhan @otarkhan94]

@ivanfioravanti I think the M5 Max wins on prefill speed which is gonna result in better performance on long contexts

[Tarkhan @otarkhan94]

I did a great job

[Tarkhan @otarkhan94]

Claude did a limit reset today and my actual reset is tomorrow morning so I'm trying to use the entire week limit today. 24% in a couple of hours.

[Tarkhan @otarkhan94]

hey builders 👋

looking to connect with people building in:
🚀 saas
🧠 ai agents
🎥 video editing
🎨 design tools
🌐 web apps
📊 analytics
🔌 apis

drop yours below and let's connect 👇

[Tarkhan @otarkhan94]

everyone is saying Fable 5.1 is the best coding model ever released

but what are you actually using it for that GLM 5.3 / Kimi K3 / Codex can't do?

genuine question

[Tarkhan @otarkhan94]

@bijanbowen 30 seconds is enough to notice the jump, not enough to know whether it survives a real codebase.

I care more about: does it keep context across a 45-minute agent run, make fewer regressions, and cost less to repair when it goes wrong?

[Tarkhan @otarkhan94]

Codex or PI harness when using GPT-5.6 models? Any recommendations?

[Tarkhan @otarkhan94]

@ivanfioravanti What Harness do you use?

[Tarkhan @otarkhan94]

As someone who actually deploys these: clickbait titles cost real time. I stopped trusting release notes and just benchmark on my own hardware (MLX, M5 Max) before believing any architecture claim. An honest report would save everyone that step — and it'd get cited more, not less.

[Tarkhan @otarkhan94]

@thsottiaux The Codex cli wasn't as good for me as ClaudeCode when I tried it before

[Tarkhan @otarkhan94]

@thsottiaux @claudeai watch and learn

[Tarkhan @otarkhan94]

@ImLunaHey I moved to Hermes

[Tarkhan @otarkhan94]

@ddalcu Congrats man, you deserve it!

[Tarkhan @otarkhan94]

@catmanyau Good models, much better limits and higher transparency than Claude

[Tarkhan @otarkhan94]

Okay, I did it. Time to move back to Codex

[Tarkhan @otarkhan94]

@CharuMitraDubey me to for some reason

[Tarkhan @otarkhan94]

@TheAhmadOsman