WEIRD AI TOOLS
The unhinged side projects and apps that shouldn't exist but do.
Do not upgrade Strata! Tons of breaking changes!
Devs forcing desktop apps on CLI users: the ultimate 'why are you like this' move.
Old server machine with 2x 3060 12GB for MoE models, anything I’m missing?
Old workstation + used 3060s = the poor man's AGI rig.
Deepseek released DeepInfer - Open Source high-performance, large-scale inference framework for DeepSeek models!
DeepSeek just dropped their secret sauce in a bucket and handed us the spoon.
HermesConsole: A demo showing off VoiceManager, my low latency streaming STT / TTS server. MIT, client docs, and python client.
Local Her is here and it sounds like Star Trek without the cloud latency tax.
2x R9700 layer split side effect
Local LLM enjoyers discover the secret to silent, multi-GPU bliss.
I got tired of Apple's configurator, so I built one for local AI
Finally, a way to calculate how much the Apple Unified Memory tax will hurt your wallet.
Budget upgrade for a 31B model? Dual GPUs over PCIe x1 or another approach?
Poverty-tier hardware vs. SOTA weights: the eternal struggle of the local LLM dev.
24GB is plenty if your engines take turns
Hot-swapping models like it's 1999 so your 4090 actually listens to you.
Open GPU kernel engineering series: naive GEMM to FlashAttention, H100 and Blackwell (44 sessions, code public) [P]
Why wait for Blackwell when you haven't even saturated your memory bandwidth yet?
Taco: cross-platform LLM compression that doesn't break when you change your GPU or CPU. [P]
Finally, an LLM compressor that doesn't freak out when you switch from Mac to Linux.
I built an app to efficiently run decision models on the Apple Neural Engine with Shortcuts actions for OS automations
Your iPhone's Neural Engine is finally doing something besides tagging your cat photos.
GameSummoner: ease of local use update
Local dungeon masters just got a one-click install for their AI hallucinations.
Jev Based Adaptive Thinking for LLMs
Infinite loops, but make it intelligent: JEV-gated adaptive thinking is here.
Replika’s founder built the AI assistant she pitched at Disrupt 12 years ago
The 'Her' prophecy is 12 years late but finally shipping.
I checked BitNet and Bonsai on CPU against references: bitnet.cpp main still gives wrong output, Bonsai 1.7B is correct wherever it loads, and ik_llama.cpp with -rtr is usually fastest
Your 1-bit LLM is hallucinating because of a one-line SiLU bug, stay winning CPU bros.
Strata 2 beta running at half speed?
Beta update turns 4090s into expensive space heaters.
Has anyone else here found a place for DiffusionGemma in their local work?
Local speedrunners are turning DiffusionGemma into an image-analysis demon.
TorchNeuron: running HF models on AWS Trainium is easier than ever
Jensen isn't the only one selling shovels anymore.
People are using AI to make their smart TVs dumb again
Spending $2,000 on a TV just to have an AI agent act as an ad-blocker is a mood.
Maxwell: a new model that speeds up AI models like Whisper and LTX-2 by rewriting their binaries [D]
Your AI is now optimizing its own binary code at runtime, what could possibly go wrong?
older Nvidia Maxwell running Strata
LocalLLaMA wizard brings zombie Maxwell GPUs back to life with 100% VRAM hit rate.
Nvidia launches new open models to run agents in the enterprise and on PCs
Jensen is done selling just shovels; he's giving you the dirt too.
Harness Engineering 101: The LLM is the brain, the harness is the body
Stop prompt tuning and start building the exoskeleton your model actually needs.
Show HN: Qwen3-4B with the AI writing tics trained out
Finally, a bot that doesn't start every sentence with 'In the rapidly evolving landscape of...'
I used an AI coding agent to solve one of gaming's "millennium problems": what the Mario 64 plaque really says
Agentic workflows just solved a 28-year-old Nintendo conspiracy theory.
Strata with multi gpu explanation
Throwing more GPUs at the problem doesn't always make the RAM go poof.
What to expect from the Intel D270 GPU?
Intel enters the chat with 480GB VRAM, because size actually does matter.
Qwen3.8-27B and Flash-Next on consumer hardware: engine numbers I could verify, and what they cost with 2026 RAM prices
LocalMaxxing is the new meta for budget GPU hoarders.
DeepSeek V4.1 Flash at 2.61 bpw + Engram: 26 t/s on dual RTX 3090s with dsx
Local AGI go brrr: Running a massive MoE on consumer GPUs at usable speeds.
incoai/Qwen3.8-Flash-Next-DFlash2
Local inference just found another gear while you were sleeping.
Author-in-the-loop Agentic Reproducibility Auditing for a Machine Learning Conference
AI agents are now grading your homework—and checking the math in your research papers.
Kapitsa: open-source Apple Silicon optimized Metal engine built for Qwen and Nemotron. See benchmarks of Qwen 27B, Nemotron + 8 other models on an M1 Pro (16GB), M4 Pro (48GB) and M5 Max (128GB)
Local models on Mac go brrr: Kapitsa drops custom Metal kernels for peak performance.
Strata - Windows - working dual AMD GPU setup with VISION !
Local-first AMD users are finally eating, even if they have to cook the drivers themselves.
Data dump repository and browser extensions for Open LLMs
Crowdsourcing the 'vibes' to kill the closed-source moat once and for all.
Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling
Your terminal multiplexer just became the AI agent's favorite playground.
Arrow detection with ConvNext pretrained on DINOv3 [P]
Labeling 8,000 images by hand is the modern labor equivalent of building the pyramids.
SteerTTS: 23M-param English TTS you steer with numbers (pitch, rate, loudness, etc. as z-scores), runs real-time on CPU [P]
Mamba-powered speech that doesn't sound like a depressed toaster or melt your CPU.
If you don't run your own LLM proxy, how do you know whats going on?
Raw-dogging external APIs is a devops disaster waiting to happen.
Google Open Sources Harness2 for Self Improving AI Agents
Google just dropped the 'auto-tune' for AI agents and the weights didn't even move.
Little Sweet Room Challenge | Strata Fork | Qwen 3.8 FN NVFP4 | DeepSeek Harness
Local models are now spatial architects for people who can't use Blender.
Qwen 3.8 Flash Next optimized for two CMP 170hx GPUs
Extracting insane throughput from crypto-mining relics through sheer engineering spite.
llm-echo 0.5
The smartest dumb model just got a massive brain upgrade for your CI pipeline.
I messed up. Buying Z9PE-D8 WS with 4 x16 Slots and 256GB DDR3 RDIMM
Local LLM enjoyer tries to run 2024 models on 2012 e-waste, fails miserably.
PSA: Power limit your 5090s!
Local LLMs are finally solving the energy crisis by literally burning your house down.
Embedding model can be used to search text inside image
Local OCR without the OCR: just vibes, embeddings, and 330MB of Gemma.
Is there a Strata for lfm 24b a2b
LocalLLaMA users: squeezing a 24B whale into a 16GB bucket.
Show HN: Sphinx – A hook that quizzes you on your AI's code before shipping
Pop quiz, hotshot: do you actually know what that Cursor tab-complete just did?
[P] asmsense: 228k stripped x86-64 functions with source-aware labels, and a map of the data we still lack. [P]
Reversing the un-reversable: naming stripped functions just got a lot less painful.
Made sum crap that beats Jev
LocalLLaMA users are now shipping 'crap' that allegedly cooks the pros.
Buying Strix Halo right now?
POV: You're trying to buy hardware from a future that hasn't happened yet.
New quant types suggested: IQ2_K, IQ3_K, IQ4_K, IQ5_K and IQ6_K
Squeeze your weights harder: new K-quants just dropped for the local LLM enjoyers.
Built a search engine for debugging which is not a wrapper [P]
Local-first RAG that actually finds the needle in the codebase haystack.
GPT-2 KV cache from scratch: correctness, latency, and memory measurements [P]
Back to basics: proving the math behind KV caching with raw PyTorch and cold hard metrics.
told Qwen3.8 27b to hack my router
Claude says 'illegal', Qwen says 'hold my firmware update'.
Inline rendering with document.currentScript
Local-first React is just <script> tags and a dream.
Openspline - host a realtime talking avatar on a commercial GPU (90+ FPS on RTX4090)
Your AI girlfriend just got a 90 FPS upgrade and a lot less latency.
Basalt: NVFP4 + faster prefill for Qwen3.8 Flash-Next on Blackwell (single 5090)
Local LLM bros are already redlining the 5090 with NVFP4 native support.
Muse Spark on M3U 256 GB
Local LLM enjoyers are now running 800B models on Mac minis at the speed of a slow reader.
2 gaming PCs + an M4 Pro MacBook over Ethernet + 2 iPhone 17 Pro Maxes over USB-C: Qwen3.8 Flash Next at 37-39 tok/s at 60k context (262k layout in the works)
Frankenstein’s LLM: Stitching five random devices together to run a model that shouldn't fit.
if you're on Strata before 0.1.45 and your client doesn't set sampling, you're running greedy
Your model isn't dumb, it's just stuck in a greedy loop—update your Strata builds.
GET THE DAILY CHAOS
The only newsletter for people who read AI news at 3am and feel things. One email a day.