Executive Summary
The past 24 hours in tech and AI have been dominated by a wave of new open-weights model releases and significant advancements in local inference hardware and developer tools. Notable releases include LiquidAI's LFM2.5-2.6B agentic model, DeepSeek V4 Flash, and MiniMax H3 for video generation, all pushing the boundaries of on-device capabilities. Developer workflows are rapidly evolving with the adoption of agentic environments like Amp's "orbs" and local-first tools like Nativ. Meanwhile, the community is heavily debating hardware bottlenecks, GPU operating systems, and the competitive landscape of open-source versus proprietary AI.
Key Events
- LiquidAI releases LFM2.5-2.6B, an on-device agentic model capable of tool calling and multi-step workflows, running locally on Apple Silicon with impressive benchmarks. → link
- OpenCode Go users are spending $130K daily on DeepSeek V4 Flash, causing capacity issues and highlighting massive adoption of cost-effective open models. → link
- MiniMax H3 video model gains traction with local execution support via ComfyUI and MLX, enabling text-to-video generation on Apple Silicon. → link
- Amp introduces "orbs," significantly changing developer workflows by offloading agent tasks and enabling multi-threaded agentic programming. → link
- Firecrawl launches "anydoc," a Rust-based document parsing engine capable of sub-5ms markdown parsing for PDFs and Word docs. → link
- Airtable is acquired by Bending Spoons at 2.7x ARR, marking a significant consolidation in the no-code SaaS space. → link
Analysis
There is a clear acceleration in the "local AI" movement, with developers actively benchmarking open-weights models (DeepSeek, Qwen, LiquidAI, MiniMax) on consumer hardware like Mac Studios and RTX 4090s. The focus is shifting from sheer model size to practical inference metrics: capacity, memory bandwidth, and software stack efficiency. Agent frameworks like Amp's orbs and Nous Research's Hermes harness are becoming the standard interface for executing these models, emphasizing multi-step reasoning and tool calling over simple chat. Expect to see further optimization in local inference stacks and increased competition among hardware vendors (Nvidia, Apple, AMD, Intel, Tenstorrent) to capture the developer market.
Tweet Feed
AI Model Releases
@victormustar · 2026-08-04T18:37
RT @CChadebec: 📢 New @heyjasper release ! 📢 MONET 🌸 : An Apache2.0 deduped and recaptioned dataset of 105M samples unlocking reproducible… → tweet link
@victormustar · 2026-08-04T18:32
RT @bfl_ai: FLUX 3 Video is here. Serious, fun, creative, real, cinematic, whatever you need it to be. Native audio, Text to Video, Image… → tweet link
@Prince_Canuma · 2026-08-04T18:15
What a release! Congratulations @liquidai You can try out the latest LFM2.5 models on @Nativ_AI https://t.co/yWzsrYLud7 → tweet link
@Prince_Canuma · 2026-08-04T18:06
RT @Nativ_AI: LFM2.5-2.6B now runs locally on your Mac 🚀 Great release by @LiquidAI — and Nativ supports it Day 0. Full BF16, no quantiza… → tweet link
@Prince_Canuma · 2026-08-04T18:04
How fast is @liquidai LFM2.5-2.6B on Apple Silicon? We benchmarked it — from a single request up to 16 concurrent streams, all on one Mac. M5 Max (48GB) · Nativ v0.2.2 · full bf16 Get started 👉 https://t.co/yWzsrYLud7 → tweet link
@Prince_Canuma · 2026-08-04T17:56
Congrats to @liquidai on LFM2.5-2.6B! Excited to have partnered with them for Day 0 support in Nativ 🎉 Built for agentic + coding workflows — and it’s fast. On an M5 Max (48GB) with Nativ v0.2.2 — full bf16, no quantization: ⚡ 11,231 tok/s prefill ⚡ ~84 tok/s decode 🧠 Full 128K context in just 8.5GB 📈 476 tok/s aggregate decode at batch 16 Local inference doesn't get much better than this. Get started today👇🏽 https://t.co/JoWC2hLXYA → tweet link
@victormustar · 2026-08-04T15:30
RT @AntLingAGI: Today, we’re releasing the open weights for Ling-3.0-flash. 🎉 Official BF16 and FP8-quantized versions are now available,… → tweet link
@victormustar · 2026-08-04T14:15
RT @liquidai: Today we release LFM2.5-2.6B, an agentic model that runs entirely on-device. It plans, calls tools, and works through multi-s… → tweet link
@Teknium · 2026-08-04T14:40
RT @liquidai: We trained it inside the real agent harnesses people use. > Four stages of post-training: SFT, expert specialization, multi… → tweet link
@victormustar · 2026-08-04T13:33
Qwen3.8-Max result on the THREEJS Boeing benchmark: https://t.co/o6O0xcA5OI → tweet link
@Teknium · 2026-08-04T16:40
RT @NousResearch: Qwen 3.8-Max, the latest model from @Alibaba_Qwen, is now available in Hermes Agent at 20% off. > hermes update → tweet link
@Teknium · 2026-08-04T04:18
Be sure to take advantage of 90% off DeepSeek v4 Flash on https://t.co/bpIDTDlUh0 - a truly epic model for the price → tweet link
@ivanfioravanti · 2026-08-04T13:48
Atomic Chat released some interesting GGUF for DeepSeek V4 Flash 0731! Let me try them on their chat! → tweet link
@TheAhmadOsman · 2026-08-03T22:45
So far we've gotten - Kimi K3 - MiniMax H3 - DeepSeek V4 Flash Coming up - Qwen 3.8 Max 2.4T - Qwen 3.8 27B - GLM 5.3 By the time we get to October, Opensource AI will have won → tweet link
@TheAhmadOsman · 2026-08-03T23:47
Built my own MiniMax H3 Studio for local video generation Thanks for the SoTA model, @MiniMax_AI https://t.co/4RjwZaWJNj → tweet link
@ivanfioravanti · 2026-08-04T11:09
Ostris (the best person to follow if you are into LoRA for image and video models) has added support for MiniMax H3 to AI Toolkit!!! 😱 → tweet link
@ivanfioravanti · 2026-08-04T12:29
Testing MiniMax H3 on Apple Silicon, I don't expect super speed honestly, but we'll see. https://t.co/lPetEz0DQM → tweet link
@ivanfioravanti · 2026-08-04T13:14
Kudos to @ComfyUI for the amazing work on MiniMax H3 model! Blog posts have all details, including the explanation of pruned models. https://t.co/ZL4hZO9KMa → tweet link
@ivanfioravanti · 2026-08-04T13:31
Look at this detailed repo: MiniMax-H3-MLX by @AIBizarrothe full of great details on the conversion! https://t.co/8WfnGUuDLL → tweet link
@ivanfioravanti · 2026-08-04T13:35
Did you know you can integrate MiniMax H3 online video creation in your agents? Here it's running in Hermes Agent! https://t.co/YxibvLSY6y → tweet link
@ivanfioravanti · 2026-08-04T14:02
Another single DGX Spark experiment with MiniMax H3, minimax_h3_fl2va_int8_convrot model not the pruned one here. 1376 x 768 - 10 seconds took 50 minutes. We need some speed-up tricks here. I bet they'll arrive from the community. https://t.co/9uxapnQEd2 → tweet link
@ivanfioravanti · 2026-08-04T09:42
RT @MiniMax_AI: Claims that MiniMax H3 “cannot legally be used” in certain regions are incorrect.😅 MiniMax H3 can be licensed for deployme… → tweet link
@LinusEkenstam · 2026-08-04T11:42
RT @maxescu: There are fun AI video models and serious production models. MiniMax H3 is both! Featuring @LinusEkenstam @techhalla and @BL… → tweet link
@gdb · 2026-08-03T22:25
GPT-Live is a new architecture and stack for realtime audio: → tweet link
Developer Tools & Frameworks
@sqs · 2026-08-04T09:57
You can now attach videos, PDFs, datasets, log files, and more, so Amp can use and see them in orbs. https://t.co/3GieLpFvlk https://t.co/ImNIXEFQ1C → tweet link
@sqs · 2026-08-03T21:44
The way we work on the Amp team has changed more in the last 6 weeks than all of last year. We need to do a better job of showing this, will do → tweet link
@sqs · 2026-08-04T16:08
RT @thorstenball: Spent the whole day interviewing Amp team (recordings out soon). Everybody said they switched to orbs. Everybody said th… → tweet link
@sqs · 2026-08-04T12:21
"even when I'm sitting right at my desk, having a job run in [orbs] lightens my local machine" → tweet link
@sqs · 2026-08-04T07:40
RT @thorstenball: Everything has changed for us with orbs. My biggest struggle right now is figuring out how to make you see what we see.… → tweet link
@sqs · 2026-08-04T02:36
"Amp with Orbs feels significantly above the current state of agents in the cloud, much like the jump from Opus 4.1 to 4.5 from late last year" Great review of Amp vs. Claude Code Cloud vs. Codex Cloud → tweet link
@badlogicgames · 2026-08-04T13:29
getting to see how the @AmpCode sausage is made. @thorstenball has a little waveshare esp32 thing with touchscreen. just tells it to build breakout for it. doesn't care about the stack (arduino instead of esp-idf). concerning. (try Amp! it's orbin time) → tweet link
@badlogicgames · 2026-08-04T12:58
i just saw an @AmpCode person submit a prompt, like it's 2025. midly disgusted, ngl. → tweet link
@Prince_Canuma · 2026-08-03T23:49
Nativ v0.2.1 is here 🚀 Talk to your Mac. Record your meetings. Fully offline. 🎙️ Global voice dictation & meeting recording — anywhere in the app 📌 Pinned chats, session folders & drag-and-drop sidebar organization 🧠 Embeddings support + per-model configuration profiles 🛠️ Pluggable tool registry, new Artifacts page & Open Interpreter integration ✨ Settings persistence, Japanese/Chinese input & dozens of fixes No cloud. No API bills. Nothing leaves your Mac. Download for macOS 👇 https://t.co/JoWC2hLq92 Star the repo ⭐ https://t.co/ZXxv58RR6W → tweet link
@Prince_Canuma · 2026-08-03T20:15
Nativ v0.2.0 is here 🚀 Talk to your Mac. Record your meetings. Fully offline. 🎙️ Global voice dictation & meeting recording 📌 Pinned chats, session folders & drag-and-drop sidebar organization 🧠 Embeddings support + per-model configuration profiles 🛠️ Pluggable tool registry, new Artifacts page & Open Interpreter integration ✨ Settings persistence, Japanese/Chinese input & dozens of fixes No cloud. No API bills. Nothing leaves your Mac. Download for macOS 👇 https://t.co/JoWC2hLq92 Github repo 🌟 https://t.co/ZXxv58RR6W → tweet link
@Prince_Canuma · 2026-08-03T20:07
Why type when you can talk? We added support for voice dictation. Your audio never leaves your Mac. Talk and it types. Download Nativ v0.2.0 for macOS: https://t.co/yWzsrYKWnz → tweet link
@jezell · 2026-08-04T16:22
RT @firecrawl: Today we're launching anydoc, our new Rust-based doc parsing engine 🔥 sub ~5ms markdown parsing for PDFs, Word docs, slide… → tweet link
@Teknium · 2026-08-04T05:07
How to get photon up and running with Hermes Agent to get iMessage all setup! → tweet link
@levelsio · 2026-08-03T23:33
Okay so today I worked on the coolest part of is my AI video editor: the agent! It's a Cursor-like sidebar and you can just tell it to edit your video with your clips and library for you. It's still very basic but it made this edit all by itself! It sends the current state to @xAI and then asks it to edit it based on your story. Live now for everyone on my site Photo AI 😊 Tomorrow I'll try make it just multi-lanes and become more smart... → tweet link
@RayFernando1337 · 2026-08-04T06:12
I finally cracked it! There is a better way to run Codex agents without the slop. As bonus it saves you tokens! Engineers at top FANG companies are using this workflow to ship production software. https://t.co/lLCDRrxnJT → tweet link
Hardware & Local Inference
@TheAhmadOsman · 2026-08-03T21:38
Local AI hardware = capacity * bandwidth * software stack. - Capacity tells you what fits - Bandwidth tells you how hard the box can breathe - The software stack tells you how much of the spec sheet you can actually cash out. Hardware by Memory Bandwidth... The only mental model that matters: 1. What must fit? 2. What bandwidth tier do I need? 3. What software stack can actually deliver it? → tweet link
@tinygrad · 2026-08-04T15:56
Hey @intel, want a shot at relevancy in AI? I know you still have those 5k DC Max 1450 cards. Sell the lot to us for $1M and we'll build $5,000 DeepSeek V4 Flash boxes for the people. Good cultural test for Intel. → tweet link
@tinygrad · 2026-08-04T15:14
GPUs need a new operating system. We're stuck in this paradigm of launching kernels when really we have 256 independent processors with various synchronization and communication primitives. → tweet link
@sudoingX · 2026-08-04T08:47
my laptop has two gpus and it absolutely blasts when my agents get to work. right now one's ffmpeging the shit out of a batch of videos, watch these nvtop spikes. rog scar 18, 5090 24gb, 64gb ram, gifted to me by someone on here, and this thing does not stop producing. → tweet link
@sudoingX · 2026-08-04T05:21
the loudest voices in ai spent two years lobbying to ban the best chips from china. they got exactly what they asked for, and it built china a second ai industry from scratch. huawei's atlas 950 is the receipt. an exaflop of fp8 on 256tb of unified memory, all ascend silicon... → tweet link
@sudoingX · 2026-08-04T05:05
i don't think anyone on x is more excited to run qwen 3.8 27b dense than me, and it's days away now. i am biting my nails. this is going to add at least 5 more years to my rtx 3090, my 24gb vram tier cards. → tweet link
@sudoingX · 2026-08-04T07:28
i benched the 3.6 27b on the 3090 until it was the king of the 24gb tier, so i know exactly what that card does with a 27b, tok/s, context, what it can build. when 3.8 drops at the same size, i run the same bench and measure the exact jump. a smarter model on the same 24gb for free. i'll have the numbers the day it lands. → tweet link
@alexocheema · 2026-08-03T20:50
big big big launch coming! let us know if you're interested to help. we spent way too much money making this good (close to $1M). why? we built it for ourselves... we didn't' find anywhere else to help us choose the best open weight models to run locally, and the best hardware to buy. we've tested every popular model / quantization / hardware. TPS is not enough... → tweet link
@alexocheema · 2026-08-03T21:01
We're bringing rigor and transparency to the local hardware debate. Make buying decisions based on real data from real agent tasks. https://t.co/nyfrArLJMm → tweet link
@TheAhmadOsman · 2026-08-03T00:54
He thanks me for inspiring him to build this 4x RTX PRO 6000 rig but I am actually jealous of how beautiful his build is. Very well done. → tweet link
@sudoingX · 2026-08-04T18:02
the people who say local models "aren't there yet" last ran one in 2024 and are describing a museum. → tweet link
Software Development & Programming
@jezell · 2026-08-03T22:33
How about some WebGPU compute shaders running in Flutter via three_flocker? WASM guest, WebGPU compute running natively on the gpu via dawn wire. https://t.co/yc1NTT6ro4 → tweet link
@jezell · 2026-08-03T20:22
three_flocker materials / clearcoat test. WebGPU really delivers the goods. Rendered into a native texture and displayed via a Flutter widget here. Original sample ported from three.js. https://t.co/23GsXy2rmy → tweet link
@jezell · 2026-08-04T08:52
Early build comparisons, flocker skia wasm comparison vs main flutter. Flocker's skia graphite wasm is fairly close to Flutter's skwasm build. The big difference though, Flocker skia graphite exposes 1,672 skia functions that can actually be used, and skwasm exposes 284. I'm sure it can still come down quite a bit from here, but I think 0.2kb is worth it for 6x the API surface... → tweet link
@MatejKnopp · 2026-08-04T15:32
Apparently what "trailing_commas: preserve" really means is 'sometimes I will add a trailing comma if I feel like it, who are you to disagree'. I still don't understand how we got to the point where a formatter can modify anything else than whitespace. → tweet link
@louszbd · 2026-08-04T16:18
General purpose coding agents still have limits. They may not know the ugly code is to work around upstream limits which hasn’t been fixed... In this case, it’s better to have a reviewer on the team. A good start is to take time to solve problems which your workflows keeps failing. May feel slow but somehow speed up everything that follows. → tweet link
@kunchenguid · 2026-08-04T04:37
i'm on my macbook, ssh'ed to my mac mini where my firstmate is, and firstmate ssh'ed to my macbook controlling crewmates here, some of which ssh'ed to my mini to run compute heavy tasks. i don't know how many roundtrips there are now, and at this point i'm too afraid to ask → tweet link
Startups & Industry News
@RealGeneKim · 2026-08-04T16:59
RT @deedydas: Wow. Airtable, founded in 2012 and once valued at $11.7B, is getting acquired by Bending Spoons, founded 2013, at 2.7x ARR.… → tweet link
@thdxr · 2026-08-04T18:07
OpenCode Go users are spending $130K a day on deepseek - that's $47M a year this is why their API keeps going down, they've now limited our traffic we're working through this but very frustrating as we're their #1 customer → tweet link
@thdxr · 2026-08-04T05:29
RT @opencode: deepseek flash is currently having capacity issues from the unprecedented volume you may see errors - we're working on a fix → tweet link
@swyx · 2026-08-04T17:39
we are huge airtable fans - aie runs on airtable - so seeing this number might seem surprising, but don't count @howietl out! if you were at WF26 you would already have seen @hyperagentapp which is the next chapter of team Airtable! https://t.co/yQhb4VUue4 → tweet link
@jack · 2026-08-04T04:13
RT @ProjectLoupe: We’ve been busy scanning 24 bitcoin repos for two months. Between them, we flagged 643 vulnerabilities, 70+ of which have… → tweet link