Executive Summary
The dominant story of the last 24 hours is the release of Kimi K3 by Moonshot AI—a 2.8T parameter Mixture-of-Experts model with open weights and a full technical report, now available on Hugging Face and Ollama's cloud. The release sparked widespread discussion about open-weight licensing models, local AI feasibility on consumer hardware, and the competitive landscape against closed labs like Anthropic. Meanwhile, SSI (Ilya Sutskever's startup) is reportedly raising $5 billion from Nvidia, signaling continued massive capital flows into frontier AI. On the hardware and tooling front, developers demonstrated increasingly capable local AI workflows—running 27B parameter models on 8GB consumer GPUs—and Grok 4.5 showed impressive CAD/engineering generation capabilities.
Key Events
- Kimi K3 open weights released — Moonshot AI published a 2.8T MoE model with weights, technical report, and day-0 integration recipes for vLLM, SGLang, and Ollama. The license includes a novel commercial clause requiring profit-sharing from cloud providers. → link
- SSI reportedly raising $5B from Nvidia — The funding round signals Nvidia's deepening investment in frontier model development beyond its chip business. → link
- Flux 3 announced — A single model jointly trained across images, video, and audio, with video prediction accounting for 95%+ of compute costs. Early testers report impressive results. → link
- Grok 4.5 demonstrates advanced CAD generation — Users showed Grok 4.5 building complex parametric OpenSCAD models (full Starship stack with Mechazilla) with accurate spatial reasoning. → link
- Latent Space expands as media org — Hired Richard MacManus as Head of Editorial, opening sponsorships, and hiring writers/producers. → link
- Nvidia's first "Made in USA" GB300 AI chips rolling off production line — Reported via Kalshi Finance. → link
- Cohere released multiple open-source models — Transcribe, Command A+, and North Mini Code all under Apache license this year. → link
- Apple Core AI on iPadOS 27 beta — Developers running FLUX.2 [klein] 4B INT4 locally on iPads with sketch-to-image LoRA, fully offline. → link
- Meta confirmed new open weights release — Adding to the week's flood of open-weight model drops. → link
Analysis
The past 24 hours mark a significant inflection point for open-weight AI. Kimi K3's release—combined with Meta's confirmed upcoming open weights and Cohere's existing Apache-licensed models—represents a coordinated industry push that isolates Anthropic as the lone major holdout against open weights. The Polymarket odds (99.5% NO on Anthropic signing an open-weight letter) confirm this is priced in.
A second major pattern is the democratization of local AI on cheap hardware. Multiple developers demonstrated 27B parameter models running agentic loops on 8GB consumer GPUs (RTX 3060 Ti), with efficiency metrics showing ~70x cost advantage per token over frontier APIs. This challenges the narrative that meaningful AI requires datacenter-scale infrastructure. The open-weight license model from Kimi K3—free for self-hosting but requiring profit-sharing for commercial cloud deployment—is a notable innovation in sustainable open-weight business models.
What to watch next: Whether Kimi K3 distills and fine-tunes proliferate in the local AI community; Anthropic's response to industry-wide open-weight pressure; Meta's open-weight release details; and whether Flux 3's unified multimodal approach pressures OpenAI's Sora strategy.
Tweet Feed
Kimi K3 Release & Open Weights
@Kimi_Moonshot (via RT) · 2026-07-27T15:16
Releasing the model weights and technical report of Kimi K3. Kimi K3 is our most capable model: a 2.8T MoE model with n… → link
@Teknium · 2026-07-27T15:21
Kimi made it to Hugging Face and is now open weights, with a paper too! → link
@TheAhmadOsman · 2026-07-27T15:32
Kimi K3, aka Fable 5 at home, weights are now fully available to download for free. Comes with a technical report as well where they share their latest research findings and how they built the model. Needs ~2TB storage for cold backups. → link
@ollama · 2026-07-27T15:58
Kimi K3 is now available on Ollama's cloud. To use it with Claude Code, run:
ollama launch claude --model kimi-k3:cloud. Currently requires a Pro or Max subscription. → link
@Ex0byt · 2026-07-27T18:23
KIMI-K3 PRISM-DQ (Dynamic Quant) Recipe Analysis is out. A 1- or 2-bit Kimi-K3 isn't happening at any quality worth shipping. The MXFP4 weights model has no dead weight. → link
@Ex0byt · 2026-07-27T18:53
Kimi's flooding the zone today: weights, tech report, MoonEP, FlashKDA, AgentEnv, PerceptionBench, plus day-0 vLLM, SGLang and TokenSpeed recipes. → link
@tinygrad · 2026-07-27T15:53
This is a great sustainable business model for open weights. It's free if you run it yourself, but if you are running a cloud providing it to others for money, you should have to share profits. (from Kimi K3 License) → link
@victormustar · 2026-07-27T15:46
Can't run kimi K3 locally? doesn't matter that much imo. This model is going to thrive in the local AI community because you'll be able to train your own distills for cheap: one trained on your own traces, one for 3D generation, one distilled on math reasoning, whatever you want. → link
@TheAhmadOsman · 2026-07-27T12:08
Today we make sure Kimi K3 weights are downloaded and backed up across several nodes. They will never be able to take away my Fable 5. → link
SSI / Nvidia Funding
@TrungTPhan · 2026-07-27T18:57
SSI reportedly raising $5 billion from Nvidia. → link
Grok 4.5 / CAD & Engineering
@sudoingX · 2026-07-27T17:35
Grok build with grok 4.5 is genuinely insane at CAD now. I pointed grok build at openscad and asked for the starship stack, and it built the whole thing—super heavy and the ship and mechazilla, 121 meters, one unit per meter, to scale. One parameter file drives every module. → link
@sudoingX · 2026-07-27T18:55
grok 4.5 is really really really good at cad man, for what it costs. → link
@sudoingX · 2026-07-27T18:40
reporting back. turns out grok build is a great rocket engineer. → link
Local AI on Consumer Hardware
@sudoingX · 2026-07-27T14:34
A 3.9gb local ai model built a working rotary engine on a $200 used rtx 3060ti gpu card, and i filmed the whole thing. Bonsai 27b, a full dense 27b crushed to about a bit a weight, 3.9gb, running on a single rtx 3060 ti with 8 gigs of vram. This is episode one of the local arena. → link
@sudoingX · 2026-07-26T19:21
The 3060 ti isn't just faster—42 tokens a second to the 1660's 20—it's more efficient per watt too. 0.26 tokens/sec per watt vs 0.16. About seven million output tokens for one dollar of electricity vs a frontier API giving you a hundred thousand for that same dollar. → link
@sudoingX · 2026-07-27T00:39
Everything moving in this clip, the epitrochoid housing, the rotor turning at exactly a third of shaft speed, the chambers firing in sequence, every line of it written by bonsai 27b dense model, running on one used 3060 ti with 8 gigs of vram. Own your cognition. → link
@sudoingX · 2026-07-27T00:30
Hot take: telling beginners in local ai space they need a 24gb card is an upsell in 2026. An 8gb card runs a 27b agent today. → link
@sudoingX · 2026-07-27T01:12
Look at this. 50 tok/s on qwen 3.6 27b, mobile 5090, with MTP on. For reference I get about 35 on that same card without it. MTP is multi token prediction. → link
@TheAhmadOsman · 2026-07-27T17:19
RTX 3090 owners trying to run Kimi_K3_3T_Q_0.0001_K GGUF → link
@sudoingX · 2026-07-27T10:17
I love laguna s 2.1. Running it on my dgx spark and it just works. 38 tok/s at nvfp4. 128gb unified in a box that fits on the desk, dead quiet. Hermes agent streams the whole run to Telegram—I can steer it mid-run by just replying to the message. → link
Flux 3 & Multimodal Models
@ivanfioravanti · 2026-07-27T10:08
Reviewing Flux 3 blog post I understood why OpenAI abandoned Sora: "FLUX 3 is one model, jointly trained across images, video and audio from the beginning. The most demanding part of that training - accounting for over 95% of the total compute costs - is video prediction." → link
@ivanfioravanti · 2026-07-27T14:51
Flux 3 + Martin Scorsese 🚀 (BTW I've got access, testing, stay tuned!) → link
Open Source AI Advocacy & Debate
@jezell · 2026-07-26T21:54
Free to distribute does not equate to open. If you have the weights from kimi, you can't rebuild kimi. None of these models are open source. They just let you make copies for your friends. Wake me up when they release the training data and the full training pipeline. → link
@sudoingX · 2026-07-27T02:11
Polymarket: will anthropic sign the open-weight letter? YES 0.5% · NO 99.5%. priced in half a second. → link
@sudoingX · 2026-07-27T01:52
Closed AI is a car that only starts on the dealer's wifi, and anthropic is the dealer calling it a safety feature. → link
@sudoingX · 2026-07-27T01:42
The open models are exactly the flood I've been asking for. The biggest player shipping open weights is the frontier coming down to hardware people own. But watch the incentive—an independent open team and a trillion dollar company build harnesses for very different reasons. → link
@Teknium (via RT) · 2026-07-27T16:55
At Nous Research we believe that open model sovereignty will make the world a safer place. → link
Developer Tools & Frameworks
@sqs · 2026-07-27T18:22
Just reset all Amp subscribers' orb usage. Also just cut @AmpCode orb prices by 20% for everyone. It's orbin' time! → link
@steipete · 2026-07-27T15:45
My agent reported a bug, their agent fixed it. [in the same night] @jarredsumner's robobun setup is future. → link
@levelsio · 2026-07-26T19:00
Gf didn't ask me anything and made an entire app in Claude Cowork locally. Then Claude Code deployed it on Netlify with Supabase without her knowing anything about that and I didn't help her. → link
@mraleph (via RT) · 2026-07-27T18:51
The official Flutter and Dart websites now run on Jaspr! 🐶 → link
@jezell · 2026-07-27T10:21
Flocker running WASM apps in a browser via web workers via a plan9 terminal powered by libghostty. Zig, C, Rust, Dart, and WebGPU all playing nicely together in the same process. → link
@jezell · 2026-07-27T05:00
@pavanpodila is porting ProseMirror to Flutter. Nice to see a lot of the top tier JS packages getting dart ports. → link
@MilksandMatcha · 2026-07-27T17:58
The hard part of agent loops is not starting them! It's knowing when to move up a layer, and when to come back down. That flexibility (not token volume for its own sake) is the key to productive loops. → link
Model Intelligence & Reasoning
@kunchenguid · 2026-07-27T15:36
Bigger models are "wiser"—better intuition, connects the dots, creative ideas. Reasoning effort makes a model more "diligent"—assesses each option, thinks through consequences. "Wisdom" and "diligence" are orthogonal. If the problem requires a genius, use a bigger model. If it requires a pen and lots of paper, use higher reasoning effort. → link
@kunchenguid · 2026-07-27T01:04
By popular demand, I made a new video of me building a real full stack app end to end using agents. To make it fun and challenging, I chose to use only gpt-5.6-luna :) → link
@TheAhmadOsman · 2026-07-27T01:15
Opus 5 is SoTA in Game Designing. HELLA IMPRESSIVE and this is not easy for me to say. → link
@MengTo · 2026-07-27T03:57
Opus 5 is getting eerily good at creating product videos with camera moves like zooms, perspective shifts, and seamless transitions. → link
@KingBootoshi · 2026-07-27T15:31
THE NEW CHATGPT VOICE MODE IS ABSOLUTELY CRACKED. I just had it autonomously 3d scan and print an object using my macbook ON THE FIRST TRY. → link
Apple / iPadOS AI
@ivanfioravanti · 2026-07-26T19:42
Pushing Core AI on iPad OS 27 beta 4 like crazy! JustDraw is a work in progress that turns sketches into images entirely on an iPad. FLUX.2 [klein] 4B + sketch-to-image LoRA, compressed to INT4 and running through Apple Core AI on iPadOS 27 beta 4. No server. No connection. Fully local and offline. → link
Hardware & Chips
@jezell (via RT) · 2026-07-27T08:39
BREAKING: $NVDA first "Made in USA" GB300 AI chips are rolling off the production line → link
@FrameworkPuter · 2026-07-27T16:14
Framework Laptop 13 Pro reviews are going live today, and as usual, Phoronix is one of the first out, going through the experience with Linux. → link
@sama · 2026-07-26T22:52
agreed feels big, i want a new kind of computer → link
Open Source Ecosystem & Tools
@victormustar (via RT) · 2026-07-26T20:19
Cohere has released open-source models Transcribe, Command A+, and North Mini Code so far this year, all available under Apache… → link
@nummanali · 2026-07-26T20:09
Tips for creating an Open Source portfolio: Always start with a problem you have yourself. Start with the smallest project you can maintain with one hour a week. Treat your project like a product. Ensure you have agentic practices in place to make maintenance easier. → link
@louszbd · 2026-07-27T08:02
I've seen more and more devs run GLM-5.2 locally over the past week. Sharing a few impressive ones. → link
@MengTo (via RT) · 2026-07-27T17:15
I'm open-sourcing my Three.js game dev skills. Build an isometric action RPG with camera controls, VFX, audio, monster assets. → link
Latent Space / AI Media
@swyx · 2026-07-27T16:52
Somehow without me noticing, @latentspacepod has quietly overtaken some of my own podcast heroes. We are seriously ramping up LS as a new tech media org. @ricmac joins us as our first fulltime Head of Editorial. We are hiring writers and show producers. → link
Coding & Agent Workflows
@MilksandMatcha · 2026-07-27T17:21
Rust has no exceptions, errors are values. Enums + pattern matching make states explicit. Result and Option handle failure and absence. Part 3/4 of the Independent Studies series w/ @shreyas4_ at @KernelLabs_ai 🦀 → link
@jxnlco (via RT) · 2026-07-27T17:19
"do i fix this slop now or just wait for the next model to do it" is the modern day wait equation → link
@thdxr · 2026-07-26T21:14
I am way more accepting of slop on the frontend vs the backend. Frontend slop could result in very user visible problems, but it feels much easier to fix. → link