Daily AI Briefing — July 16, 2026: Inkling Open-Weights, Japan's National AI Infrastructure, and China's Companion Rules

A busy 48 hours in AI: Thinking Machines Lab opens the weights on its first production model, Japan and NVIDIA announce the world’s first national AI infrastructure, DeepSeek accelerates toward IPO, China’s companion-AI rules take effect, and state-level AI safety regulation gains momentum.

1. Model Releases

Thinking Machines Lab Unveils Inkling — A 975B Open-Weights Multimodal MoE

The biggest story of the week: Thinking Machines Lab — the AI startup founded by former OpenAI CTO Mira Murati — released its first production model on July 15. Inkling is a mixture-of-experts transformer with 975 billion total parameters and 41 billion active per forward pass, supporting text, image, and audio inputs with a context window of up to 1 million tokens. Pretrained on 45 trillion tokens, Inkling is distributed under the Apache 2.0 license, and the lab is positioning it as a customizable base model for fine-tuning rather than a leaderboard-chasing flagship. Early independent benchmarks from Artificial Analysis place it as the leading U.S. open-weights model at position 41 on their Intelligence Index. Weights are available immediately, and the model can be fine-tuned via Thinking Machines’ Tinker platform. Source: Bloomberg | Source: Artificial Analysis | Source: TechCrunch

PrismML Ships Bonsai 27B — A 27B Model That Fits on a Phone

On July 14, Pasadena-based PrismML released Bonsai 27B, a 1-bit binary and 1.58-bit ternary compression of Qwen3.6-27B. The ternary variant consumes just 5.9 GB and retains 94.6% of FP16 performance; the 1-bit variant fits in 3.9 GB (89.5% retention) and can run on an iPhone 17 Pro. Both variants ship under Apache 2.0. At 134–163 tok/s on an RTX 5090, this is the first 27B-class multimodal model capable of genuinely running on-device. Source: PrismML | Source: MarkTechPost

Kimi K3 Leaks Hint at Moonshot’s Next Flagship

Leaked promo pages and beta sightings suggest Moonshot AI is preparing to launch Kimi K3, reportedly a 2.5-trillion-parameter MoE model with a 1M-token context window, native 4-bit inference, and a new “Agent Swarm” orchestration layer. Early benchmark leaks suggest coding performance near the Fable-class level. Source: explainx.ai | Source: Times of AI

2. Industry Moves

Japan and NVIDIA Launch the World’s First National AI Infrastructure

In a major development announced today, the Japanese government — backed by the Ministry of Economy, Trade and Industry (METI) — partnered with NVIDIA and Noetra Corp. to build a 140-megawatt AI factory powered by 27,500 next-generation NVIDIA Rubin GPUs and 13,750 Vera CPUs. Dubbed the world’s first national AI infrastructure for physical AI, the initiative targets manufacturing, logistics, healthcare, and telecommunications. Jensen Huang, visiting Japan this week, also announced that Japanese enterprises will build specialized AI models using NVIDIA’s Nemotron open model family. Source: NVIDIA Newsroom | Source: Tom’s Hardware | Source: WSJ

PwC and OpenAI Launch Agentic Customer Service Solutions

PwC US announced on July 15 the launch of agentic contact and service solutions built in partnership with OpenAI. The offering combines OpenAI’s latest models with PwC’s systems integration capabilities and a dedicated Center of Excellence, targeting enterprise contact centers. Source: PwC

Gemini 3.5 Pro Aims for July 17 GA After Full Rebuild

Google DeepMind is targeting tomorrow — July 17 — for the general availability of Gemini 3.5 Pro, a model the company rebuilt entirely from scratch after its original base model failed to meet internal benchmarks on math and reasoning. Source: Tech Times | Source: HackerNoon

3. Funding Rounds

DeepSeek Seeks $1.5B at $71B Valuation, Eyes 2027 IPO

Chinese AI lab DeepSeek is reportedly in talks to raise $1.5 billion at a pre-money valuation of approximately $71 billion, according to Bloomberg and TechCrunch — just weeks after closing a $7.4 billion round at a ~$50 billion valuation. The company is also preparing for an IPO on the STAR Market as early as 2027, and is on track to generate nearly $500 million in annualized revenue. Founder Liang Wenfeng has become the highest-valued AI founder globally, per the Bloomberg Billionaires Index. Source: TechCrunch | Source: Bloomberg | Source: aiweekly.co

Other Notable Rounds

Bolna AI raised $6.3M led by General Catalyst to build multilingual voice AI agents for Indian enterprises Source: EntrepreneurLoop
Arq disclosed a $1M pre-seed round for an undisclosed AI application Source: AI Funding

4. Open-Source News

Inkling’s Apache 2.0 Release Reshapes the Open-Weights Landscape

The open-source community is digesting the implications of Thinking Machines Lab’s Inkling release. As the first major U.S. open-weights model at the ~1T-parameter class, it offers a domestic alternative to Chinese open models like DeepSeek-V4 and Qwen3.6. Source: Thinking Machines Lab

Beijing Considers Restricting Overseas Access to Chinese AI Models

Reuters reported on July 7 that China’s Ministry of Commerce has held meetings with Alibaba, ByteDance, and startup Z.ai about potentially restricting foreign access to China’s most advanced AI models — including open-weight releases. Source: Reuters | Source: The Next Web

Anthropic Expands Claude for Open Source Program

Anthropic expanded its “Claude for Open Source” program on July 8, offering up to 10,000 qualifying open-source maintainers and core contributors six months of Claude Max 20x free. Source: explainx.ai

5. Policy & Regulation

China’s AI Companion Rules Take Effect

July 15 marked the effective date of China’s Interim Measures for the Administration of AI Anthropomorphic Interactive Services — the world’s first comprehensive national framework specifically governing AI companion and emotional chatbots. Issued jointly by five agencies including the Cyberspace Administration, the rules require providers to disclose the AI nature of their services, file with regulators, implement safeguards against psychological harm, and maintain liability for noncompliance. Source: IAPP | Source: Lexology

Illinois Enacts AI Safety Measures Act

On July 6, Illinois Governor JB Pritzker signed SB 315, the Artificial Intelligence Safety Measures Act, into law. Modeled after California’s SB 1047 and New York’s earlier legislation, the law requires “large frontier developers” to publish annual AI safety frameworks, conduct pre-deployment testing, and submit transparency reports to the state. Source: Capitol News Illinois | Source: Norton Rose Fulbright

FTC Proposes Policy on AI Accuracy and Output Steering

The Federal Trade Commission is seeking public comment (until July 31) on a proposed policy statement that would classify the covert steering of AI system outputs toward undisclosed ideological objectives as a deceptive practice under Section 5 of the FTC Act. Source: FTC | Source: Federal Register

The takeaway: July 16, 2026, is a day of juxtapositions — open-weights philosophy (Inkling, Bonsai) runs headlong into sovereign AI infrastructure (Japan’s Rubin factory) and tightening regulatory guardrails (China’s companion rules, Illinois’ safety act).

See how it compares → /comparisons/

CodeIntel Log — code quality, debugging, and software engineering benchmarks
NiteAgent — AI agent development, frameworks, and production patterns

Cross-links automatically generated from None.

← Back to all posts