This Week in All Things AI covers key developments in models, agents, tools, infrastructure, and policy curated from discussions in the All Things AI Telegram group.
If you follow AI for work, research, investing, or just to understand where the technology is heading, this weekly brief is a concise way to scan the most important launches, risks, and resources in a few focused minutes.
The week of 5th July to 11th July 2026 was defined by a wave of frontier model releases and infrastructure moves, with Grok 4.5, OpenAI’s GPT‑5.6 family, Meta’s Muse Spark 1.1, and Tencent’s Hy3 all pushing toward longer context, stronger coding, and more agentic workloads, while Meituan’s LongCat‑2.0 went fully open source and MiniMax paired its proposed 2.7T‑parameter M3 Pro with a fresh $2 billion fundraise and AGI-linked commitments. Infrastructure and efficiency stayed in focus as pxpipe-style token reduction, SambaNova’s $1B raise at $11B valuation, Microsoft’s reported move toward in-house MAI models in Office, and Qualcomm’s interest in Modular and Tenstorrent pointed to a broader push for better unit economics and greater control beyond Nvidia’s CUDA moat.
In parallel, culture and adoption signals remained more uneven: US enterprise surveys continued to show modest AI uptake, while filmmakers such as Gareth Edwards and designers responded strongly to tools including Seedance and Fable 5. Voices including Dylan Field and Chamath Palihapitiya highlighted the widening tension between rapidly improving capabilities, rising inference costs, and the slower, more complicated work of turning AI into durable products, workflows, and careers.
The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interests.
How ByteDance is making Hollywood inroads with its Seedance video generator, thanks to low pricing, striking realism, and features like timeline-based prompting
via Steve IP
cut Fable 5 token usage by rendering text context as images
GitHub - teamchong/pxpipe: cut Fable 5 token usage by rendering text context as images
cut Fable 5 token usage by rendering text context as images - teamchong/pxpipe
via Robby Yung
LongCat-2.0 is now fully open-source — MIT licensed, no restrictions.
LongCat-2.0 is now fully open-source — MIT licensed, no restrictions.Since our launch a few days ago, the response from the community has been incredible.
Thank you for all the feedback, discussions, and interest.
Today, we’re releasing the model weights and inference code
1.6T parameters · MoE with ~48B active · 1M context
The full model behind Owl Alpha on @OpenRouter — now available.
Built for agentic coding from the ground up:
◆ LongCat Sparse Attention (LSA) — scales efficiently for 1M-context tokens
◆
Jurassic World Rebirth’ and ‘Rogue One’ director Gareth Edwards on AI + filmmaking
It has no taste whatsoever. It is a fucking genius at helping you. I view it like having a second-unit director who is a billionaire on acid. Like, it’ll do anything you ask, not a problem. Sometimes, it’ll [go] batshit crazy. And you’ll give it notes, and it’ll be like, ‘I don’t do notes. I’ll just do something totally different.’ But it’s worth it.
Gareth Edwards Is Excited About AI Filmmaking - Even Though It's Like a "Second-Unit Director Who Is a Billionaire on Acid"
The 'Jurassic World Rebirth' and 'Rogue One' director said he has been experimenting with diffusion models and eventually wants to create a hybrid AI film.
'Like a billionaire on acid': Star Wars director Gareth Edwards comes out in favour of AI
Speaking at Amazon's AI on the Lot event, the Rogue One film-maker Gareth Edwards said 'it'll do anything you ask' and 'it's going to be better than CGI'
In response to Ming Guang Yong question about Qualcomm reportedly paying US$4B for Modular, the following view was shared in the group:
el-GOOG paid USD 2.7 billion for CharacterAI which was in effect an acqui-hire for Noam Shazeer who is credited in many ways for the rise of Gemini to be in LLM conversations when it was massively behind the curve.
Modular has Chris Lattner of LLVM, Swift fame. Modular ultimately built a unified, hardware-agnostic AI software stack and maybe Qualcomm thinks that they need to attack nVidia's moat which is CUDA
Modular's MAX platform is essentially vendor-neutral vLLM: it runs interchangeably on AMD or Nvidia hardware already, and because it's built on Mojo (not CUDA), it doesn't require the hand-tuning that porting CUDA kernels demands. For Qualcomm, this is the key — MAX can now natively target Qualcomm's silicon
Qualcomm reportedly has entered advanced talks to acquire Tenstorrent (Jim Keller's AI chip startup) for $8–10 billion, meaning Qualcomm is potentially committing over $14 billion to reshape its AI silicon portfolio within weeks.
Personally, I'm keeping an eye on the disaggregated inference space and it could be that a new architecture might be needed for future inference demand
Claude Opus 4.8 and Sonnet 5 seem worse at tool calling than older models, likely due to post-training that assumes Claude Code-like harnesses as the targets
Better Models: Worse Tools
About an aggravating tool-calling regression in newer Claude models.
Tencent Releases Hy3: Efficient Open-Source AI Rivaling Top Models
Hy3 features 295 billion total parameters but activates only 21 billion during use, plus a 3.8 billion-parameter layer for better reasoning and a massive 256,000-token context window
It excels on benchmarks like SWE-Bench Multilingual at 75.8% and Terminal Bench at 71.7%, often matching or beating much larger rivals such as GLM-5.2 and GPT-5.5.
Refined from April 2026 preview feedback and now under permissive Apache 2.0 license, it shines in real-world workflows per blind human evals by 270 experts, earning praise for agent reliability and easy deployment on 8x H200 GPUs.
https://x.com/TencentHunyuan/status/2074148098876768478
It is focused on cost-effective agentic use, and particularly strong on coding, tool-calling reliability, reasoning, and 256K long-context tasks.
https://x.com/ZixuanLi_/status/2074017499792756949
via Huzefa
US builds AI, citizens lag 🤖
The US makes most leading AI models but ranks 24th in adoption, with just 28% of working-age Americans using AI regularly. Countries like the UAE and Singapore have higher use after embedding AI in government services and education.
Microsoft reportedly looking to reduce AI costs is starting to replace models from OpenAI and Anthropic with its MAI models in products like Excel and Outlook
Microsoft is reportedly ditching OpenAI's and Anthropic's AI models in favor of its own to cut costs - SiliconANGLE
Microsoft is reportedly ditching OpenAI's and Anthropic's AI models in favor of its own to cut costs - SiliconANGLE
via my high-school classmate Abhi Ingle, the Chief Product and Strategy Officer of Sambanova whom I recently met after 38 years
Chip startup SambaNova raises $1 billion at an $11 billion valuation, underscoring investors’ conviction in rising demand for infrastructure tied to AI
$1Bn at a post money of $11bn. So incredibly proud and excited to share the news of our Series F funding round closing. It's been a long journey for the world to see that 'Inferencing is the... | Abhi I. | 35 comments
1Bn at a post money of $11bn. So incredibly proud and excited to share the news of our Series F funding round closing. It's been a long journey for the world to see that 'Inferencing is the defining workload in AI now and in the future'.
AI chip maker SambaNova raises $1B at $11B valuation, 5 months after last mega round | TechCrunch
AI chip maker SambaNova has raised at an $11 billion valuation months after Intel was rumored to be trying to buy it for about $1.6 billion.
Some links about Sambanova's disaggregated inference approach
The First Disaggregated Inference Demo for AI Agents Is Live
SambaNova's disaggregated inference uses dedicated GPUs for prefill and SN40 RDUs for decode, delivering 2x the speed of GPU-only setups. See the live demo.
Disaggregated Inference Explained for Enterprise AI
A practical guide to disaggregated inference: the prefill and decode phases, KV cache transfer, and how to scale AI infrastructure. Learn more.
China’s MiniMax Plans to Launch 2.7-Trillion Parameter M3 Pro Model
Chinese AI developer MiniMax is preparing a 2.7 trillion parameter large language model called M3 Pro, which is over six times larger than its current flagship M3 model with 428 billion parameters. The model is expected to be released as open source as early as the third quarter. It would become the largest Chinese AI model by parameter count if launched.
tl;dr MiniMax is preparing a 2.7T parameter open-source model, potentially launching as early as Q3.
That would make it far larger than any Chinese model currently on the market, and over 6x bigger
China's MiniMax plans to launch giant 2.7 trillion parameter model - The Economic Times
Chinese AI firm MiniMax is developing a 2.7 trillion-parameter model. This new model could become the world's largest open-weight AI system. Cheaper Chinese AI models are gaining traction as alternatives to US systems. MiniMax will also launch a multimodal video generation model soon. The company is planning a second listing on Shanghai's STAR Market.
via David An
What is the CURRENT STAGE OF AI ENTERPRISE ADOPTION ? LOWER THAN THE HYPE SUGGESTS. We at Provocation Lab analyzed the US enterprise market:
-> 18% average AI adoption
--> 56% say they won’t use AI in the next 6 months
---> 60% say it’s not even applicable to their business
----> ~96% report no significant change in employment due to AI.
In this video, I explain the data & findings:
Deck is here
via Just-LDA
Do any of you have any recommended tools / game development repos or skills, in particularly for isometric 2D/3D pixel sprite games?
response from Naor W
If I recall correctly Hermes has Godot skill and I've seen MCP for it. There are a couple of repos around it in GitHub.
SpaceXAI announced Grok 4.5 as its first model specialized for coding and agentic tasks, trained in collaboration with Cursor on tens of thousands of NVIDIA GB300 GPUs using RL for multi-step engineering workflows.
Grok 4.5 is now accessible via SpaceXAI console, Grok Build, and Cursor with limits reset for free starts, though unavailable in the EU until mid-July 2026.
x.ai/news/grok-4-5
It’s our most powerful model yet and the first we've built for more than software engineering.
It's an Opus-class model that's fast and low cost. It's a significant step up over any model we've developed so far, including Composer 2.5, and has become the daily driver for many on our team.
First of many releases. More soon.
It’s our most powerful model yet and the first we've built for more than software engineering.
Grok 4.5
x.ai/news/grok-4-5
It’s our most powerful model yet and the first we've built for more than software engineering.
It's an Opus-class model that's fast and low cost. It's a significant step up over any model we've developed so far, including Composer 2.5, and has become the daily driver for many on our team.
First of many releases. More soon.
It’s our most powerful model yet and the first we've built for more than software engineering.
Grok 4.5
Some more tweets from Elon wrt Grok 4.5 including guidance on future updates
Grok 4.5
Users should notice a meaningful improvement in the usefulness of the Grok Build harness with our V9 foundation model (aka Grok 4.5) every week.
Pareto dominant for coding by the numbers.
We will see on the all-important vibes.
Instinct is the benchmarks are likely directionally accurate given the stated focus on real world utility.
One of the hardest parts of working with AI is iterating on a project or task that you're working on.
As the models have gotten smarter (and more expensive), it's taking longer
Next month’s release will be another step-change improvement, as we close the loop on solving real-world engineering problems at Tesla, SpaceX, Neuralink and Boring Company.
It feels like Opus 4.8 at 2x the speed at a much cheaper price point.
I tasked it to brainstorm > plan > implement a big feature for my game (this act 1 boss fight) and it did
Meta releases Muse Spark 1.1 agentic and coding AI model through new Meta Model API
Meta launched Muse Spark 1.1, an upgraded multimodal AI model excelling in agentic performance, tool use, computer use, coding, and multimodal reasoning, featuring a 1M token context window and support for parallel sub-agents. The model is available in the Meta AI app and through the new Meta Model API in public preview, marking Meta's first paid AI model for developers with pricing of $1.25 per million input tokens, $0.15 for cached input, and $4.25 for output.
available now through the new meta model api and in meta ai.
Introducing Muse Spark 1.1
Today, we're excited to introduce Muse Spark 1.1, the latest model from Meta Superintelligence Labs and a significant upgrade from Muse Spark. Muse Spark 1.1 is a multimodal reasoning model built for agentic tasks, with major gains in tool and computer use, coding, and multimodal understanding.
via OpenAI
We’re launching the GPT‑5.6 family of models for general availability following our limited preview: our new flagship, Sol, alongside Terra, a balanced model for everyday work, and Luna, our most cost-efficient model.
GPT-5.6: Frontier intelligence that scales with your ambition
More intelligence from every token, stronger performance per dollar, and more capability on demand for your hardest work.
MiniMax Closes $2 Billion Funding Round as CEO Yan Junjie Commits to Forgo Salary Until AGI Achievement and Donate 5% Equity
MiniMax secured HK$16 billion (approximately $2 billion USD) in a new funding round that was seven times oversubscribed, attracting over 100 investors including global sovereign wealth funds and Chinese institutions. CEO and founder Yan Junjie announced he will receive no salary until the company achieves AGI and will allocate shares equivalent to 4% of the company's total equity from his personal holdings to reward long-term employees, plus 1% to support the open-source comm
MiniMax has just closed a new $2B funding round.
At the same time, our CEO, IO, shared three long-term commitments with the team:
• No salary until we achieve AGI.
• Over the next four years, he will dedicate shares equivalent to 4%
Markets will fluctuate, and external noise will come and go, but our direction remains unchanged.
Being at the forefront of this industry, we have a clear understanding of the pace of technological evolution and the long-term value we are building
MiniMax has just closed a new $2B funding round.
At the same time, our CEO, IO, shared three long-term commitments with the team:
• No salary until we achieve AGI.
• Over the next four years, he will dedicate shares equivalent to 4%
Dylan Field, Figma CEO, responds to designer Gal Shir's viral post claiming an AI tool called Fable 5 created a better logo than he could, prompting Shir to quit design and focus on human consciousness.
Field describes a repeating pattern where new AI models trigger hype and existential crises in creative fields like design, followed by users discovering limitations and stabilizing, amplified by designers' longstanding imposter syndrome.
He argues design's value is rising due to the attention economy, where standout creativity in branding and experiences differentiates winners, urging designers to build confidence, take risks, and educate stakeholders instead of fearing AI replacement.
The psychological journey people go through with AI is quite fascinating to
A client of mine just created a logo with Fable 5, and the result left me speechless.
It understood the brand story, values, audience, strategy, and turned all of it into a smart, minimal symbol. A genuinely brilliant concept. The kind of idea
via Sam Lai who posted about ChatGPT Voice which is powered by the GPT-Live model
This is very natural voice interaction.
Introducing GPT-Live
A new generation of voice models for natural human-AI interaction, now powering ChatGPT Voice.
Haven't finished watching this video in its entirety as yet but thought that this might interest some of the folks here.
Tanay Kothari , CEO of Wispr Flow on his small sales team using AI-powered sales systems to close 1,000 enterprise clients monthly by streamlining outreach, discovery, conversion, and post-sales processes.
At least in the brief time that I watched the video , I saw Tanay mentioning Hex Tech as something his data team utilises
PS: I use Wispr Flow extensively
via Chafik Naceri
Hey all i built an ai native platform for mobile gaming studios to turn ideas into production ready mobile games in hours rather than weeks. Check out demo
tell me your feedback
Platform was built to help studios prototype more games ar a fraction of the costs
Chamath Palihapitiya on The All-In Podcast where Chamath reveals his company's AI inference token costs are doubling every 45 days with only about 5% productivity improvement. Chamath explains that further model iterations now demand far more tokens due to hitting diminishing returns and an asymptote in gains, prompting a strategic pause at his firm. He advises AI-related companies to exit or raise funds at current high valuations before widespread cost pressures affect the sector in the next 3-4 years.
Vibing isn’t tolerated because these are the systems that run western society - banking, power, healthcare, insurance etc.
Over the last few quarters, the gains that we got from using frontier models
Below is my personal website which aggregates links to many of my socials as well as the various content and community that I curate. Feel free to share this link to others who you think may find this content/community useful to them
The cover image of this newsletter via generated via the Seedream 5.0 model within the Krea tool via the following prompt
3d rendering, isometric, mountain park, paper work for a tactile style, select a color palette with vibrant colors, blue, and white with green, emphasize textures that rough-textured colored paper, with lighting that creates a warm, golden-hour glow, handmade