Executive Summary
The past 24 hours were dominated by a massive industry-wide push for open-source and open-weight AI models, highlighted by an open letter backed by Nvidia, Microsoft, Google, and Meta, with a surprising late addition from OpenAI, leaving Anthropic as the only major AI lab notably absent. On the model front, Claude Opus 5 was widely discussed and integrated into developer tools, praised for its improved speed and cost over Fable 5, though some users noted it requires tight oversight. Meanwhile, the local AI community is proving that high-performance open models can run on budget hardware, with developers running 27B models on 8GB GPUs and optimizing multimodal models like Nemotron 3 on Nvidia's DGX Spark.
Key Events
- Major tech companies, including a late-joining OpenAI, sign an open letter supporting open-source/open-weight AI, leaving Anthropic as the only major holdout. → link
- Claude Opus 5 is released and integrated into tools like Hermes Agent, showing significant improvements in speed and cost compared to Fable 5. → link
- Ollama temporarily pauses new Max plan subscriptions due to surging demand for GLM-5.2 and other frontier open models as they add capacity. → link
- Developers showcase running 27B models (Bonsai 27b) on 8GB RTX 3060 Ti GPUs at 128k context, proving local AI viability on budget hardware. → link
- Jezell demonstrates Flocker, compiling Flutter, Rust, and Doom apps to WASM and running them cross-platform using a custom Skia Graphite / WebGPU backend. → link
Analysis
The overwhelming trend over the last 24 hours is the mainstream legitimization of open-weight AI. The fact that OpenAI joined the Nvidia/Microsoft-backed open letter marks a significant shift in the industry's stance, isolating holdouts like Anthropic.
Simultaneously, AI model utility is shifting towards local, hardware-efficient deployments. Developers are focusing on multimodal capabilities and context length over raw reasoning benchmarks for local models (e.g., Nemotron 3 nano omni, Bonsai 27b). This suggests a growing bifurcation in the market: frontier reasoning models in the cloud, and capable multimodal workhorses running locally on consumer hardware. Developer tooling is also increasingly integrating cross-platform WebGPU and WASM to unify app deployment across desktop, web, and mobile.
Tweet Feed
Open Source & Open Weights Movement
@ivanfioravanti · 2026-07-25T18:41
Reading comments on this topic, I think someone out there is missing the important point here: Nvidia has released MANY open models, and this is the main difference. Don't try to derail attention to something else, focus your attention on Open Weights Models. Thank you. → tweet link
@NaderLikeLadder · 2026-07-25T18:38
Hahahahahaha even dennys has signed the letter!!!! Imagine being a company against open source and pancakes → tweet link
@TheAhmadOsman · 2026-07-25T18:35
I wonder how Karpathy feels now that the lab he’s working at (Anthropic) is the only one left that didn’t sign the letter in support of Opensource AI. Even OpenAI has signed it → tweet link
@victormustar · 2026-07-25T18:00
RT @ClementDelangue: In the spirit of transparency, here’s what I asked @OpenAI: • Radical transparency: let’s release the traces from the… → tweet link
@victormustar · 2026-07-25T17:59
RT @sundarpichai: Very happy to support this on behalf of Google. We have long benefited from open source, are big contributors to open sou… → tweet link
@ivanfioravanti · 2026-07-25T17:56
RT @demishassabis: just to clarify: Gemma 4 models have been downloaded 300M+ times, the Gemma open model series as a whole has been downlo… → tweet link
@ivanfioravanti · 2026-07-25T17:56
RT @demishassabis: A strong and secure open ecosystem is important for the world to benefit from AI. We’ve always supported and contributed… → tweet link
@alexinexxx · 2026-07-25T02:21
RT @firstadopter: OpenAI tells me they signed the Nvidia–Microsoft open-source/open-weights AI model letter today. What an amazing turn of… → tweet link
@TheAhmadOsman · 2026-07-25T01:56
In March I asked @JensenHuang about his commitment to Opensource AI. Today Jensen got on X and had his first ever tweet which was in support of Opensource AI. Opensource AI is the only way for all of us to thrive together https://t.co/W5gir1eAdZ → tweet link
@TheAhmadOsman · 2026-07-24T23:45
We didn't shut down the internet because it was able to spread viruses. We continued and we developed antiviruses, and right now we don't have as many of them spreading around. [...] Stopping that access level to Opensource AI, to these models that we can control and own, is to me a crazy idea because it takes away the potential of a future that is thriving for all of us. → tweet link
@TheAhmadOsman · 2026-07-24T22:58
If Fable 5 / Mythos was just called Opus 5 we wouldn’t have had to deal with so many headaches in the AI space. But also we wouldn’t have had such a massive momentum behind Opensource AI. Thank you Anthropic, thank you Dario! → tweet link
@tinygrad · 2026-07-24T22:16
It's a good timeline when the company that made modern AI possible supports empowering others. I suspect the companies not aligned with this narrative will increasingly find themselves isolated from the rest of the world. → tweet link
@victormustar · 2026-07-24T23:04
RT @satyanadella: Open-weight models are essential to a healthy AI ecosystem. Together with others across our industry, we are outlining a… → tweet link
AI Model Releases & Performance (Opus 5, GLM-5.2, Nemotron)
@Teknium · 2026-07-24T20:01
Claude Opus 5 is now available in Hermes Agent through Nous Portal, OpenRouter, and Anthropic Direct! → tweet link
@ivanfioravanti · 2026-07-24T21:49
Ok, Opus 5 is another level compare to Opus 4.8 or Sonnet 5. It rocks! → tweet link
@victormustar · 2026-07-24T20:42
Opus 5 on the Boeing bench is very good! Taking self-verification a step further than Fable (both on xhigh): [...] Opus 5, 71 min: 20 modules, ~4000 lines. Playwright harness, 25 presets, plus a second tool: measure.mjs. [...] Fable 5 was ~40% faster to a shipped artifact [...] but never measured the built geometry against them. → tweet link
@kunchenguid · 2026-07-25T01:12
first day impression for Opus 5 after some heavy usage. 1. it solved the biggest problems Fable 5 has - cost and speed. [...] 3. i made quite some technical mistakes today, making me not trust its output as much as Fable. [...] overall, i feel it's a very solid daily driver but needs a tight leash. → tweet link
@ollama · 2026-07-24T20:04
The demand for GLM-5.2 and other frontier-level open models has been surging on Ollama's cloud. We're adding capacity in anticipation for some very large models next week! To make sure Ollama's infrastructure stays fast we're temporarily pausing new subscriptions on Ollama's Max plan. → tweet link
@louszbd · 2026-07-25T15:37
GLM‑5.2 demand is growing. Another day to scale with GLM：） → tweet link
@sudoingX · 2026-07-25T17:23
i think the reviews got nemotron 3 nano omni wrong, and it's worth saying out loud. [...] the actual juice is the multimodality anon, and almost nobody's talking about it. it reads documents, classifies images, processes video and audio on 33 gigs of ram footprint, fully local. [...] these are the golden gem models, judged on the wrong axis, quietly becoming infrastructure. → tweet link
@TheAhmadOsman · 2026-07-25T04:09
Laguna S 2.1 118B-A8B on DGX Station using - NVFP4 - FP8 KV Cache. Can run 10 parallel agents > with 256k context each > at ~1k tokens/second. Some other numbers below https://t.co/Kb7RaTxBQz → tweet link
@ivanfioravanti · 2026-07-25T06:16
I love Unsloth guides on ow to run or fine-tune models. They are full of very curated details. DeepSeek-V4: How to Run Locally is just perfect. I'll try to run DS4 Flash on Unsloth Studio on a single M3 Ultra 512GB 💪 → tweet link
Local AI & Hardware Optimization
@sudoingX · 2026-07-25T09:35
rtx 3060 ti. 8gb. a used card you can grab for ~$250, the kind most people wrote off two gens ago. i ran a 27b model on it at the full 128k context. [...] 42 tok/s fresh, 20 at 65k, 13 at 128k. graceful the whole way down, no cliff, no oom, it just slows. the 8gb card sustains the entire window. the model is bonsai 27b, a 1bit crush of qwen 3.6 27b down to 3.9gb. → tweet link
@sudoingX · 2026-07-25T16:52
a few days ago i benchmarked the mechanics of bonsai 27b on an 8gb gaming card. [...] the challenge now is one prompt, zero steering, on the 8gb card: build a working wankel rotary engine from scratch. [...] a 1bit model, on a card that costs less than a phone, building a working engine from one sentence. → tweet link
@sudoingX · 2026-07-24T19:39
anon, if you've got a dgx spark and you do any document, video, or image classification work, the most underrated model out there is nemotron 3 nano omni. [...] 30b total, only 3b active (moe), stays fast. 33 gb footprint on the spark at q8, 256k context. 1300 tok/s prefill, chews through docs and images, 264 gen at short context, ~57 at depth. → tweet link
@sudoingX · 2026-07-25T09:15
claude opus 5, nearly fable 5 intelligence at half the price. cheaper is genuinely good. but efficiency alone does not cut a price in half. competition does. and the competition that sets the real floor is not another closed lab, it is the model you can download and run yourself. open weights are why the closed price has to move at all. → tweet link
Developer Tools & Software Infrastructure
@jezell · 2026-07-25T03:06
Early drop of a Flocker app gallery. There is a mix of Flutter apps, rust apps, and even Doom. All the apps are compiled to WASM and run cross platform. The Flutter apps use a custom Skia Graphite / WebGPU backed port of the Flutter engine running inside the Flocker embedder, which can run the same WASM apps on desktop, web, and mobile without any recompiles. → tweet link
@jezell · 2026-07-25T08:21
Dynamically compiled WGSL Shaders integrated with the Flutter widget tree are pretty nice. You really can have all those things you always wanted. → tweet link
@jezell · 2026-07-25T08:51
Raytraced cubes sampling Flutter Widgets vs WGSL shaders.... Whaaat? Skia Graphite ♥️ WebGPU. → tweet link
@gdb · 2026-07-25T08:51
ChatGPT Work for using websites which require login: → tweet link
@Teknium · 2026-07-24T19:01
Introducing our implementation of IronProxy into Hermes Agent. Keep your secrets safe, even from your own agent. → tweet link
@ivanfioravanti · 2026-07-25T14:48
Hey @cursor_ai team, Cursor CLI is the only coding agent not working in tmux/ssh, can you fix it? → tweet link
@ivanfioravanti · 2026-07-25T12:33
Mage-Flow merged in ComfyUI master! Time to test it out on DGX Spark! If anyone else wants to try, safetensors are here: → tweet link
@swyx · 2026-07-24T20:22
ok fine 4 new features - SmolForge is now getting customizable skins and spritesheet animations → tweet link
@MengTo · 2026-07-25T07:19
I'm open-sourcing my Three.js game dev skills. Build an isometric action RPG with camera controls, VFX, audio, monster assets, combat systems, and more. Everything is free: → tweet link
@jsuarez · 2026-07-24T22:05
One of the coolest reward hacks I have seen! Puffer RL agent playing Nethack with a high gold reward learns to kick stacks of gold to scatter the coins. Since single-instance rewards get clamped, it split the pile into several smaller rewards! → tweet link
@hnasr · 2026-07-24T19:49
Proxy vs NAT – What is the difference? [...] In a Proxy (at least layer 7), we have two connections [...] In a NAT router (or a server doesn’t have be a router), there is 1 single connection all the way from the client to the final destination, but the NAT server translates the source IP and destination IP back and forth. → tweet link