Executive Summary
The AI landscape experienced a massive shift toward open-weight frontier models, highlighted by the release of Moonshot AI's Kimi K3 (3T parameters, 1M context window) and Thinking Machines' Inkling (975B MoE), both directly challenging closed models like GPT-5.6 and Fable 5 on agentic benchmarks. Extreme quantization breakthroughs by PrismML demonstrated that 1-bit compression of large models not only fits them onto consumer GPUs (like the RTX 3090) but actually increases inference speed by reducing memory bandwidth bottlenecks. Developer tooling evolved rapidly around autonomous agents, with new CLIs for real-world commerce (DoorDash) and OpenAI introducing GPT-Red for automated prompt injection red-teaming. Meanwhile, the software ecosystem saw Rust and WebAssembly maturing into serious platforms for high-performance cross-platform applications, including running full browsers and databases inside WASM.
Key Events
- Moonshot AI launched Kimi K3, a ~3T parameter open-weight model with a 1M context window, beating GPT-5.6 Sol Max and Fable 5 Max on the BrowseComp benchmark. → link
- PrismML showcased 1.125-bit quantization of Qwen 3.6 27B, shrinking it to 3.5GB and running at 67.9 tok/s on a single RTX 3090—1.69x faster than the full q4 model due to memory bandwidth optimizations. → link
- Thinking Machines released Inkling, its first open-weight model: a 975B total / 41B active multimodal MoE trained on 45T tokens with a 1M context window. → link
- OpenAI introduced GPT-Red, an internal automated red teamer designed to find prompt injection vulnerabilities at scale. → link
- xAI open-sourced the Grok Build CLI and reset usage limits in an effort to rebuild community trust. → link
- Ollama reported that over 85% of Fortune 500 companies are already using open models locally for specific tasks, emphasizing ownership and data privacy. → link
- A new version of Postgres is being written in Rust, aiming to become the "LLVM of databases." → link
Analysis
Patterns & Trends: The open-source AI community is aggressively pursuing scale to match closed frontier models, as seen with Kimi K3's 3T parameters. However, because these models are too large to run natively on standard hardware, we are seeing a parallel surge in extreme quantization research. PrismML’s 1-bit "bonsai" technique proves that pushing models below 4-bit doesn't necessarily destroy them; in fact, it can enhance inference speed on consumer GPUs by optimizing for memory bandwidth rather than compute.
Simultaneously, AI agents are transitioning from code-generation tools to general-purpose computer users and real-world operators. The introduction of the DoorDash CLI for agentic food ordering and OpenAI's GPT-Red for automated red-teaming indicate that the industry is moving toward agents that interact with external APIs and operate autonomously, which brings security and prompt injection risks to the forefront.
In software engineering, Rust and WebAssembly are experiencing explosive growth. Running Firefox entirely inside a WASM sandbox, or rendering Flutter apps via WebGPU shaders at 120fps in a WASM environment, signals a future where OS-level dependencies are abstracted away in favor of unified, portable, and secure execution layers.
What to watch next: * Intelligence retention audits: Whether 1-bit or sub-4-bit quantized models actually retain their reasoning capabilities on complex agent tasks, or if the speed gains come at a severe cognitive cost. * Local AI hardware: Apple's next generation of Silicon (M5/M7 Ultra) and how unified memory capacities (up to 1.5TB rumored) will democratize running massive local models. * Agentic security: How the industry will mitigate prompt injection risks now that automated tools like GPT-Red are scaling vulnerability discovery.
Tweet Feed
AI Models & Open Weights
@TheAhmadOsman · 2026-07-16T15:38
Kimi K3
- 2.8T Parameters
- 1M Context Length
Benchmarks
GDPval-AA v2: 3rd place, ranks directly below GPT 5.6 Sol Max & Fable 5 Max
AA-Briefcase: 2nd place, beats GPT 5.6 Sol Max, right below Fable 5 Max
BrowseComp: 1st place, beating GPT 5.6 Sol Max & Fable 5 Max → tweet link
@Ex0byt · 2026-07-16T15:57
I had extremely high hopes for Kimi K3, my verdict is, regrettably, a disappointing one. For a model reportedly weighing in somewhere north of 2T params, K3 feels less like a frontier leap than an exercise in bloat. Inference is slow, agentic behavior is overly cautious and clumsy, and the outputs read like a heavy distillation of Claude Opus. → tweet link
@alexinexxx · 2026-07-16T16:00
just got access to kimi k3. honestly, it feels like another big step forward for open-weight models. they’re getting a lot closer to closed frontier models. → tweet link
@alexinexxx · 2026-07-16T05:05
RT @eliebakouch: first open weight thinking machine model!! 975B total, 41B active trained on 45T tokens, 1M context, multimodal in → tweet link
@victormustar · 2026-07-16T10:49
RT @julien_c: For a first model release, @thinkymachines Inkling is quite a good model. Realize, it's infinitely more difficult to ship yo… → tweet link
@jezell · 2026-07-15T21:29
RT @whoisanku: xAI has open-sourced Grok Build. The release includes the Git repo for the Grok Build CLI and a reset of usage limits for a… → tweet link
@kunchenguid · 2026-07-15T21:39
open sourcing grok build is a good move from xAI to start rebuilding trust → tweet link
@jxnlco · 2026-07-16T02:25
RT @OpenAI: Introducing GPT-Red An internal automated red teamer on a mission to find our models’ prompt injection vulnerabilities at scal… → tweet link
@sama · 2026-07-15T20:17
RT @OpenAI: Introducing GPT-Red An internal automated red teamer on a mission to find our models’ prompt injection vulnerabilities at scal… → tweet link
Local AI, Quantization & Hardware
@sudoingX · 2026-07-16T18:28
i said i'd make it prove itself. here are the numbers.
prismml took qwen 3.6 27b, the model i've been calling king of the 24gb tier all month, and quantized it to 1.125 bits. every weight is a single sign bit now, the whole 27b packed into 3.5gb. it should be broken. sub 4 bit is where models turn to mush. it isn't broken. it's faster. → tweet link
@sudoingX · 2026-07-16T18:47
quantize a model this hard and the instinct is that it limps. broke to a quarter of the precision, surely it pays for it.
it's the opposite. i measured it on my 3090. the 1-bit bonsai runs at 67.9 tok/s. the full q4 model, same card, does 40.1. that's 1.69x faster, and prefill is 3.4x. → tweet link
@alexocheema · 2026-07-16T01:02
Most inference will run locally.
Internalise this, play out the implications, and plan accordingly. → tweet link
@ollama · 2026-07-16T02:11
Open models are already being used in the enterprise.
Over 85% of the Fortune 500 companies already use Ollama to fulfill specific tasks. @jmorgan → tweet link
@alexocheema · 2026-07-15T23:51
Why I'm optimistic:
Apple hosted their first ever official Local AI event on June 23rd. It was a private event at Apple Park, @tim_cook / John Ternus / other execs speaking there about local AI, ~300 in attendance. → tweet link
@alexocheema · 2026-07-16T00:48
Apple accidentally made the best hardware for Local AI.
What can they do when it's intentional and core to their strategy? → tweet link
@alexocheema · 2026-07-16T16:42
We're starting to roll out early access for local dot ai.
We're 1% through the early access list. → tweet link
@tinygrad · 2026-07-15T23:51
On a car or mobile robot, you don't have the money for bandwidth or time for latency to stream your cameras to the cloud. eGPUs are going to be the top choice of smart robots everywhere. Why limit your robot to some tiny chip, stick a 5090 or 9070XT on it with tinygrad! → tweet link
Developer Tools & Agents
@jxnlco · 2026-07-16T13:33
i’m jason. i work on developer experience for Codex at OpenAI. before that, i built Instructor and spent a lot of time helping people use language models to build real things. → tweet link
@MilksandMatcha · 2026-07-16T18:45
The underrated Codex pattern: stop asking one thread to hold the whole universe.
OpenAI's @jxnlco talks about thread managers, threads that talk to other threads, Slack context, hourly loops, and feedback IDs. → tweet link
@RealGeneKim · 2026-07-16T05:00
RT @andyfang: Today we're opening up the DoorDash CLI in limited beta.
dd-clilets you order DoorDash directly from your agent: search s… → tweet link
@jxnlco · 2026-07-16T04:49
Hey codex.
Set an automation where at 8:30 when my free DoorDash benefit comes in. I rotate across my teams doordash auth and order a full Korean bbq meal to maximally extract the late night meal benefit to feed my family. → tweet link
@Teknium · 2026-07-16T07:11
Hermes Agent is fully supported in Raft 1.0 (Previously Slock) - Give it a try! → tweet link
@thdxr · 2026-07-16T13:37
i regret to inform all the people who get mad for some reason whenever people build a TUI, OpenTUI is at 300,000+ weekly downloads → tweet link
@jxnlco · 2026-07-16T01:00
RT @aidenybai: Introducing ReactBench
A benchmark for coding agents on real React work
Models write bad React code - useEffect, slow perf… → tweet link
@jsuarez · 2026-07-16T14:50
Awesome work! Valo has been porting OSRS content to high-perf C RL envs in PufferLib. Inferno and Colosseum are both end-game challenges that take the top few % of players days or weeks to complete for the first time. Here's a tiny model taking down Sol with no supplies. → tweet link
Software Engineering & Frameworks
@jezell · 2026-07-16T17:35
RT @glcst: I am excited to announce that we are officially writing a new version of Postgres. In Rust - and creating the LLVM of databases… → tweet link
@jezell · 2026-07-16T17:03
Firefox running in WASM inside Chrome inside my custom WASM embedder, rendering via Skia Graphite. Who needs iframes and webviews? → tweet link
@jezell · 2026-07-16T00:12
How about iOS, MacOS, and Web all running the same WASM app. Custom host embedder, custom Flutter engine on top of Skia Graphite mixing custom shaders / textures and Flutter widgets? → tweet link
@jezell · 2026-07-15T22:15
MacOS host running WASM skia graphite / flutter app with dynamic WebGPU shader at 120 fps with smooth resizing. Yep. → tweet link
@jezell · 2026-07-16T18:02
RT @astuyve: Nobody is talking about Rust's explosive growth thanks to AI
The Rust Foundation is expected to serve 18PB of rust crates thi… → tweet link
@juliarturc · 2026-07-16T02:59
Here's an ethical dilemma.
Find useful-looking SaaS (say a CRM) Sign up for a free trial "Hey Fable, clickety-clack to figure out minimal feature subset I need & implement it" Cancel free trial → tweet link