Exploring the limits of local AI. Running open-source LLMs on consumer hardware and sharing what works, what breaks, and what surprises me.
Nederland
Joined April 2018
- Tweets691
- Following134
- Followers53
- Likes59K
nando gaarenstroom retweeted
abacus 0.7.1 and Abacus Desktop 0.1.1 are out.
The big change: providers and profiles. Add any OpenAI-compatible or Anthropic API with its key and headers, pick from the models it serves, and save profiles with their own settings. First-run setup is now two questions.
nando gaarenstroom retweeted
Use kojo @ kojo.sh
We are planning on a secondary release concerning cybersecurity-oriented models, which might pique some of your interests.
gif made by kojo in ~3 minutes and 21 seconds.
nando gaarenstroom retweeted
It is time.
This is the thing i was talking about yesterday :D
GENERALLY AVAILABLE NOW
Kojo, our latest model, is out.
Chat with it for free, or build on it through an OpenAI-compatible API.
kojo.sh
nando gaarenstroom retweeted
Kojo, our latest model, is out.
Chat with it for free, or build on it through an OpenAI-compatible API.
kojo.sh
Proud to finally share what we’ve been working on at Empero: Kojo. 🚀
The model performed extremely well on the completed SWE-bench Verified test suite (461/500 bugs fixed = 92.2%). Claude Opus 4.8 achieved 88.6% in the published leaderboard; Sonnet 5 85.2%; and Mythos Preview achieved 93.9%. We achieved better performance than Opus 4.8 and Sonnet 5 and is within 1.7% of Mythos Preview.
The API prices are as follows:
$0.175 per million input tokens
$1.00 per million output tokens.
$0.02 per million cached input tokens.
It is also super exciting to be part of the team launching this product so we can get more people using Kojo!
Below is something that Kojo made in one-shot:
(Do know there are better things made with this model so definitely check out the others within EmperoAI!!)
Kojo, our latest model, is out.
Chat with it for free, or build on it through an OpenAI-compatible API.
kojo.sh
nando gaarenstroom retweeted
Our website just got an all new and fresh look!
empero.org/
nando gaarenstroom retweeted
You can now get an api key for the beta! (Requires an account)
kojo.empero.org
Say "Hi!" to our newest beta model!
kojo.empero.org/chat
I am proud to announce that I am currently an active contributer at @EmperoAI. I will be helping with training the models and providing datasets. Even though I am still a novice I am curious to see where my experience with AI will lead me!
The “qwen4” 35b we needed is here in the form of MiniCPM-V 4.7!!!
🚀 Introducing MiniCPM-V 4.7, our new open-source MLLM series!
Advancing the Performance–Efficiency Pareto Frontier.
It brings strong multimodal capabilities to iOS and Android devices with extreme efficiency.
Two new models:
🔹 MiniCPM-V 4.7 35B-A3B — Bring SOTA performance to local devices with newly introduced LES (Lookahead Expert Scheduling) technique, a new milestone of on-device multimodal AI.
🔹 MiniCPM-V 4.7 1B — Lightweight, efficient multimodal AI. Easier to deploy.
⚡ What's new?
SOTA performance: Competitive results across 34 benchmarks spanning multimodal understanding, STEM reasoning, OCR and long-video comprehension.
Fast on-device inference: LES predicts and preloads expert weights to reduce inference stalls, boosting 5× speedup on iPad Pro M4 (16GB) with byte-for-byte identical outputs. It enables MiniCPM-V 4.7 35B-A3B to achieve 17+ tokens/s on iPad Pro M4 (16GB) and 13+ tokens/s on iPhone 17 Pro Max, while costing only 3GB memory 🤯
Efficient visual processing: 4× and 16× hybrid visual token compression. Even at 16× compression, 35B-A3B scores 90.4 on OCRBench, 5.2 points ahead of Qwen3.6 35B-A3B.
Easier customization: We introduce Self-Replay Fine-Tuning, a simple way to customize new target capabilities at low cost while maintaining general performance.
Broad hardware support: Day-0 adaptations for Apple, Intel, AMD, Arm and Rockchip, with many more platforms supported through FlagOS!
Download, deploy, and build something we haven't imagined yet. 👀👇
🔗 GitHub: github.com/OpenBMB/MiniCPM-V
🤗 Hugging Face: huggingface.co/collections/o…
Mistral might actually be back in the game as this model is actually pretty good if the benchmarks are trustworthy.
Meet Mistral Large 4, aka Le Chonk.
• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.
Open weights release end of October.
I haven't been that active with X and that is for 3 good reasons:
1. I have been developing my AI knowledge especially with certificates like AB-900 / AI-901 and currently learning for GH-300
2. I want to see new content and it is now pretty quite as for the big releases.
3. I have been getting quite tired and not really focused to go onto X. It might also be since I am doing this aside from my full time job. If anyone else does this how do you do it?
This is of course not a post that I am giving up as I am really excited to meet more people as posting have let me open up a lot more!
NOOOOOO @ClaudeDevs Please more plushies!
Claude.dev is our new home for developers building with Claude.
You'll find engineering deep dives, Claude Code and API guides, tips from the teams building Claude, and some fun easter eggs.
This is an interesting watch if you have the time! It is an hour long so make sure it is nice and raining!
Build Visual AI Agents From a Prompt With NVIDIA Cosmos and VSS 3.3 teapaw.org/i/broadcasts/1qGvveVbD…
I am going to install Omarchy on one of my desktops which is quite old now. It is an old HP prodesk but the good thing is it has 32gb of ddr3 ram. Does anyone have tips for me?
Google is back in the game!! And wow if the benchmarks are to be trusted this is one hell of a model!
A frontier LLM which boast about having a 10 million context window. This is starting to get insane. Like, where do they get the vram from??
