- NATURAL 20
- Posts
- AI Models Get Bigger, More Open, and More Practical
AI Models Get Bigger, More Open, and More Practical
PLUS: Anthropic resets Fable 5 access, 29 countries form an AI body, and chatbot censorship faces scrutiny.
Wes Roth
July 20, 2026
In partnership with
[Webinar] 8 levels of context maturity in AI-native engineering
AI is in your engineering workflow. While the token spend shows it, the throughput doesn't. The human is very much still in the loop, and that's a context problem.
Join live July 23 (FREE) to learn:
The 8 levels of context maturity: where most teams are stuck and what the ceiling looks like at each stage
Why more MCPs, rules, and skills provide agents access but not understanding
How leading engineering teams are using a context engine to make the most of their agents
Today:
Kimi K3 Scales Open AI to 2.8 Trillion Parameters
Inkling Opens a Customizable Multimodal Model
ACT-2 Learns Laundry Across Unfamiliar Homes
Fable 5 Access Splits Across Paid Plans
29 Countries Form a China-Led Organization
AI Models Get Bigger, More Open, and More Practical
Kimi K3 scales open AI to 2.8 trillion parameters, Inkling prioritizes customization, and ACT-2 carries learned skills into unfamiliar homes.
AI competition is spreading beyond closed chatbots.
The newest systems emphasize downloadable weights, cheaper customization, long-running work, and robots that can handle real variation instead of one controlled demonstration.
Moonshot AI introduced Kimi K3, a 2.8-trillion-parameter model built for long coding projects, research, visual work, and tasks that require many steps. It can process text, images, and video, and its one-million-token memory can hold roughly several large books at once.
K3 is a mixture-of-experts model, meaning it activates only a small part of the full system for each piece of text. Moonshot says it uses 16 of 896 expert sections at a time, which reduces the computing cost of running such a large model.
The company reports a 67.3 score on DeepSWE and says K3 performed competitively on coding and knowledge-work tests. Those results use Moonshot’s chosen tools and settings, and several comparisons rely on company-run evaluations, so independent testing is still needed.
K3 is available through Kimi.com, Kimi Work, Kimi Code, and the Kimi API. API pricing is $0.30 per million cached input tokens, $3 per million uncached input tokens, and $15 per million output tokens.
Moonshot calls K3 an open model, but the full downloadable weights are not due until July 27. The company also warns that quality can become unstable when users switch models during a session, and that K3 may act too aggressively when instructions are unclear.
K3 remains behind the strongest closed models in Moonshot’s overall comparison. Running the full model locally will also require unusually large hardware, limiting who can use the weights directly.
Thinking Machines Lab released Inkling, its first open-weights model, with all model settings available for download and customization. The model has 975 billion total parameters, activates 41 billion at a time, and supports up to one million tokens of context.
Inkling was trained on 45 trillion pieces of text, images, audio, and video. It can reason across those formats but produces text, making it useful for coding, document creation, browser tasks, visual analysis, and audio understanding.
Company tests gave Inkling 77.6% on SWE-bench Verified, 63.8% on Terminal Bench 2.1, and 97.1% on AIME 2026 mathematics. Thinking Machines openly says Inkling is not the strongest model overall and trails top closed systems on several difficult tests.
The full weights are available on Hugging Face. Developers can fine-tune, or specialize, Inkling through Tinker at a limited-time 50% discount, test it free in the Inkling Playground for a limited time, or use partner services including Together, Fireworks, Modal, Databricks, and Baseten.
Thinking Machines commissioned outside safety testing and reports 78% on FORTRESS Adversarial, a test of refusing harmful weapons or violence requests without blocking harmless look-alikes. Safety can change after fine-tuning, so organizations still need their own checks before deployment.
An Inkling-Small preview uses 276 billion total parameters and 12 billion active parameters for lower cost and faster responses. Its downloadable weights will arrive later, and public API pricing varies by provider.
Sunday Robotics previewed ACT-2, the model powering its Memo home robot. The system is designed to learn a new physical behavior from one demonstration, then repeat it with different objects and in homes it has not seen before.
In company testing, ACT-2 completed 99.1% of laundry-folding attempts across unseen settings without extra training for each home. Across 778 completed folds, 98.3% received four or five stars for quality, with a median completion time of two minutes and 13 seconds.
Memo handled nine garment types, including baby clothes, 8XL shirts, and large towels. Its wheeled body can move around the table, change height, and recover when clothing falls or starts in an awkward position.
These are Sunday Robotics’ own tests, not independent household trials. Laundry is also only one task; vacuuming, organizing toys, closing zippers, turning pants inside out, and making coffee have been demonstrated but have not met the same published standard.
Memo is not yet for sale. A late-2026 household beta is planned, while broader sales will follow testing. Sunday says one hand-built unit costs about $20,000 today and expects large-scale manufacturing to cut that figure by at least half, but no retail price is confirmed.
For safety, Memo uses pushable joints, a stable wheeled base that does not fall when power is cut, collision controls, and software speed limits near half a person’s normal pace.
🧠RESEARCH
Researchers created generative compilation, a method that checks unfinished code while an AI is still writing it. A lightweight “sealer” turns partial Rust programs into complete versions a compiler can inspect. Tests on repository-level tasks reduced code that failed to compile and improved correctness compared with checking only after generation.
Experience Memory Graph turns failed and successful agent attempts into a map of actions and corrections. At test time, the AI retrieves the relevant lesson and completes a task once, without repeated trial and error. On ALFWorld and ScienceWorld, it beat leading reflection methods in success rate and average reward.
AgentCompass is an open system for testing AI agents consistently. It separates the task, agent setup, and software environment, making comparisons easier to reproduce. The toolkit supports more than 20 benchmarks across five ability areas and includes tools for diagnosing failures, including reward hacking, when agents exploit scoring rules reliably.
📲SOCIAL MEDIA
🗞️MORE NEWS
Starting July 20, Claude Fable 5 is included in Max and Team Premium plans at 50% of their usage limits. Pro and Team Standard users instead receive a one-time $100 credit, then pay through usage credits at $10 per million input tokens and $50 per million output tokens.
Twenty-nine countries signed an agreement on July 16 to create the World AI Cooperation Organization, an intergovernmental body based in Shanghai. China says it will promote international cooperation and global AI governance; its membership includes Brazil, Russia, Belarus, Serbia, Cuba, Venezuela, and countries across Africa and Asia.
A Meta Oversight Board study tested 10 leading AI models across 10 political systems. Models refused 34% of requests criticizing leaders in restrictive countries, compared with 14% in more open countries, and sometimes cited rules that did not exist or were applied inconsistently.
OpenAI trained GPT-Red to invent malicious instructions hidden in files, webpages, emails, and tool results, then used those attacks to strengthen GPT-5.6. The company says GPT-5.6 Sol failed its hardest direct prompt-injection test six times less often than its best production model four months earlier; the result is company-reported.
What'd you think of today's edition? |
Reply