Kimi K3, Inkling, and Gold Eagle: AI Model News, July 2026

Three significant AI model developments landed in the same week of July 2026: Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-source model claiming the largest-ever open weights release; Mira Murati’s Thinking Machines open-sourced Inkling, a 975-billion-parameter multimodal model under Apache 2.0; and the Trump White House launched “Gold Eagle,” a program that gives the federal government a say in who can access new frontier AI models.

Kimi K3: The Largest Open-Source Model to Date

Moonshot AI, the Beijing-based startup backed by Alibaba, released Kimi K3 on Thursday, July 17, 2026 — a 2.8-trillion-parameter model the company says is the largest open-source AI model ever built. According to VentureBeat’s coverage, benchmark results place Kimi K3 neck-and-neck with top proprietary systems from Anthropic and OpenAI, a striking result for an open-weights release.

The model’s parameter count is roughly 75% larger than DeepSeek’s V4 Pro, which Moonshot’s own technical documentation charts at approximately 1.6 trillion parameters. Kimi K3 features a 1-million-token context window and native visual understanding — capabilities that push it into territory previously occupied only by closed, subscription-gated systems.

Full model weights are scheduled for release on July 27, according to details shared by researchers who reviewed Moonshot’s technical documentation. The model is already accessible via kimi.com with a free account — no credit card required. The release was timed to coincide with the 2026 World Artificial Intelligence Conference in Shanghai, giving Moonshot a high-profile stage for what amounts to a comeback after roughly 18 months of market-share erosion following DeepSeek’s rise.

The scale of the engineering effort drew notice across the AI community. David Sacks, founder of Craft Ventures, commented on the model’s performance relative to U.S. frontier systems after Kimi K3 outperformed them in at least one independent benchmark.

Thinking Machines Releases Inkling Under Apache 2.0

Thinking Machines — the AI startup founded by former OpenAI CTO Mira Murati — released Inkling on July 17, 2026, its first major open-weights model, licensed under Apache 2.0. The 975-billion-parameter Mixture-of-Experts system handles text, images, and audio natively, and is available immediately on Hugging Face and through the company’s own API, Tinker.

Benchmark Performance

According to Thinking Machines’ release announcement, Inkling posts competitive but sub-state-of-the-art scores among open-weights models:

• 77.6% on SWE-bench Verified (software engineering), beating Nvidia Nemotron 3’s 71.9%
• 91.4% on VoiceBench (voice understanding), compared to 94.4% for Gemini 3.1 Pro on high reasoning effort

Design Priorities

Two design choices distinguish Inkling from most frontier-class open models. First, a “controllable thinking effort” mechanism lets operators dial compute up or down depending on workload cost targets — a departure from fixed-compute inference pipelines. Second, Thinking Machines explicitly designed the model, per its announcement, “to answer directly on topics that may be subject to censorship,” positioning Inkling for enterprises that need factual outputs on sensitive subjects without content filtering.

Alongside the flagship, Thinking Machines previewed Inkling-Small, a 276-billion-parameter variant optimized for lower-cost deployments.

John Schulman, co-founder of Thinking Machines, posted about the release on X, underscoring the team’s focus on openness and practical utility for developers.

White House Launches “Gold Eagle” AI Access Program

The Trump administration this week launched a program called Gold Eagle that gives the federal government explicit approval authority over which companies and organizations can access new frontier AI models, according to CNBC, which cited sources familiar with the initiative.

Previously, Anthropic and OpenAI independently decided who received early access to new models — through their respective Project Glasswing and Daybreak cyber initiatives. Under Gold Eagle, future partner rollouts will require explicit government sign-off, CNBC reported.

The White House also launched an AI clearinghouse this week with private-sector partners to address cyber vulnerabilities tied to frontier model releases. In a statement, the administration said AI companies still control whether to release their models and are not required to participate in government testing or meetings — framing Gold Eagle as coordination rather than mandate.

The shift nonetheless represents a meaningful change in how access to the most capable AI systems is governed. Anthropic and OpenAI had previously operated their early-access programs with significant autonomy; Gold Eagle inserts a federal checkpoint into that process for the first time.

What This Means

The three developments this week, taken together, sketch the current state of the AI model market with unusual clarity.

Kimi K3’s release at 2.8 trillion parameters — with full weights dropping July 27 — sets a new ceiling for open-source model scale and hands researchers and enterprises a model that benchmarks suggest rivals closed systems costing thousands of dollars per month to access. That Moonshot AI, a company that had lost significant ground to DeepSeek, produced it signals that the competitive dynamics in Chinese AI remain fluid and capable of rapid reversals.

Inkling’s release, meanwhile, shows that the open-source push isn’t limited to Chinese labs. Thinking Machines, with Murati’s backing and a focus on enterprise-grade licensing and censorship resistance, is positioning open weights as a serious alternative to proprietary APIs for companies that want control over their AI stack.

Gold Eagle is the most structurally significant of the three. If the administration’s approval authority over frontier model access becomes durable policy, it would represent the first formal U.S. government checkpoint in the AI release pipeline — a development that could affect how quickly models reach researchers, foreign partners, and commercial customers, and one that both Anthropic and OpenAI will need to navigate carefully.

FAQ

What is Kimi K3 and when can I access the full model weights?

Kimi K3 is a 2.8-trillion-parameter open-source large language model released by Beijing-based Moonshot AI on July 17, 2026. The model is already accessible via kimi.com for chat, and full weights are scheduled for public release on July 27, 2026.

What makes Inkling different from other open-source AI models?

Inkling, released by Thinking Machines under an Apache 2.0 license, combines multimodal capabilities (text, image, and audio) with a “controllable thinking effort” mechanism that lets users adjust compute cost at inference time. It is also explicitly designed to resist content censorship filters, targeting enterprises that need unfiltered factual outputs.

What is the White House’s Gold Eagle program and what does it do?

Gold Eagle is a Trump administration initiative that gives the federal government approval authority over which organizations can access new frontier AI models, according to CNBC. It replaces the previous arrangement in which Anthropic and OpenAI independently controlled early-access decisions through their own programs, Project Glasswing and Daybreak.