Aibrat (@Aibrat0) profile picture

Building a local AI setup and posting what it can actually do. Open models vs paid APIs, real results.

Joined June 2026
Photo by Aibrat (@Aibrat0)Photo by Aibrat (@Aibrat0)Photo by Aibrat (@Aibrat0)Photo by Aibrat (@Aibrat0)Photo by Aibrat (@Aibrat0)Photo by Aibrat (@Aibrat0)Photo by Aibrat (@Aibrat0)Photo by Aibrat (@Aibrat0)Photo by Aibrat (@Aibrat0)Photo by Aibrat (@Aibrat0)
People keep asking me to prove that open-source AI is good enough for real work. Fair challenge. I’m going to start showing what runs on my own hardware versus what requires paid cloud tools, with real outputs, costs, speed and limitations. Less hype. More proof.
2
1
4
375
First test of the continuity model, fully automated on local hardware. Same characters hold across the living room porch cuts. Slightly softer (720p) on one angle, but not bad for a fully automated V1.
12
Inkling 2 Mini runs on any Mac with 32GB RAM. Three things to know if you run models on a laptop. 1. It takes about 20GB total memory. Ollama support at launch covers any Mac with 32GB or more. 2. It was trained natively in 4 bit (NVFP4) for most layers, with the last 4 in 8 bit. The weights ship as trained, so nothing gets lost to quantizing. 3. Middle layers run 3 times per token. In their research version, 5 passes cut reasoning from 3,703 tokens to 1,661 on GPQA questions it got right both ways.
24
Inkling 2 Mini and Qwen3.8 27B both score 20.1 on AA Index. Tbh I'm taking Mini, since only 3.4B are active per token. Which one is on your laptop?
1
1
14
Last week I joked that Inkling 2 Small was carrying that name alone at 594B. Thinking Machines just shipped Inkling 2 Mini. 26B, open weights. Their post opens with "We heard you."
1
18
Mark Gurman reports that in macOS 27.3, turning on ChatGPT's computer use now takes 14 steps. It used to take 4. Apple says it's security. Its own agent could ship as soon as the end of 2027.
11
OpenAI is testing dots, ChatGPT assistants that message you first. Do you want that, or one that only speaks when asked?
16
Honestly, the best part of Gemini's new Mac computer use is that it runs in a background window. Every other agent so far takes over your mouse and you just sit there watching.
8
Gemini 4 Neon and Gemini 4 Hydrogen both scored 43 out of 43 on Plotly's data analysis test. Neon costs $17.03 per run. Hydrogen costs $6.36 and was faster. Neon's per token price is double Hydrogen's too. Plotly would still pick Hydrogen for data analysis.
1
17
Which one is doing your coding this week, Gemini 4 Neon, GPT 6.2 Astra or Claude Fable 5.2?
1
1
17
You can drop a new character into a Genie 4 world or change the scene while you're inside it, and the audio adapts live. I haven't tried it. As a creator, I'd still take this over another chatbot benchmark.
6
Alex Finn's own benchmark has GPT 6.2 Astra at 312.7 out of 500, and Claude Fable 5.2 led by 0.3. That gap is basically noise. Astra took a third of the time, a third of the price and a quarter of the tokens.
20
Quick one for the India folks, is GPT 6.2 Astra in your Codex model picker right now?
1
13
Theo just posted the best argument for local AI I've seen. T3 Code hit 31.7 trillion tokens in a day, which is over $10 million in API spend in 24 hours.
5
Anthropic guidance, via Ethan Mollick, says one agent often beats a pile of specialized subagents, and costs less. I have well over 100 custom agents installed. Creative directors, SEO agents, QC agents. I feel so called out 😭
7
Fable 5.2 and Opus 5.5 tied for the lead at 58 on the Coding Agent Index v2.1. Which one do you open first, Fable or Opus?
16
ChatGPT Wallet got one thing right. The agent pays and never sees your card details. Every agent tool that touches money should work like that.
18
To all those buying smart appliances like AC, Washing Machine, Refrigerator - Honestly, tell me what is the need for all that? Are they solving any problem? Or are we creating new ones when they can be easily hacked?
20
Love seeing llama.cpp and ggml hit 1.0 today, 3.5 years after it started. So many local AI apps quietly run on it. From now on, breaking C API changes get a major version bump.
1
1
20
Inkling 2 Small is 594B parameters, 17B active, and it still wants 500GB of VRAM. Small is carrying that name alone. Apache 2.0 open weights though, I'll take it.
14
You can self host Claude Managed Agents now, but the model still runs at Anthropic. Does that count as self hosted to you?
1
48