Thinking Machines Releases Inkling Open-Weight Multimodal Model
Moonshot AI Releases Kimi K3, a 2.8T Parameter Coding Powerhouse. OpenAI Introduces Codex Site for Agentic Coding Workflows.
Thinking Machines Releases Inkling Open-Weight Multimodal Model
Mira Murati's Thinking Machines Lab dropped its first model, Inkling, on July 15 — open-weight, multimodal (text, image, audio), and explicitly built for customization rather than benchmark supremacy [1][2]. This is a ~200-person company making a clear bet: the frontier-scale, one-size-fits-all race isn't the only game, and there's real value in models companies can bend to their own workflows.
The framing matters more than the specs. Thinking Machines isn't trying to out-GPT OpenAI. It's betting that the next wave of value comes from teams who can fine-tune, fork, and adapt a capable base model to their own judgment calls — which is exactly the muscle post-code companies need to build.
Reaction on X has focused on Inkling's Chinese training influences and its practical customization potential, with builders already sketching out what they'd fork it into [2]. Open weights plus multimodal plus a credible pedigree is a combination worth watching closely, not just admiring.
Moonshot AI Releases Kimi K3, a 2.8T Parameter Coding Powerhouse
Beijing's Moonshot AI released Kimi K3, a 2.8 trillion parameter model that's reportedly topping coding benchmarks and going toe-to-toe with frontier Western models. Open weights are scheduled for July 27 — a date now circled on a lot of calendars [1][2].
The scale here (2.8T parameters) combined with a genuine open-weights commitment is notable: this isn't a lab teasing capability and gatekeeping access, it's a direct challenge on both performance and openness simultaneously. If K3 delivers on the benchmark claims once weights are public, expect a fast wave of self-hosted deployments from teams tired of API lock-in.
X sentiment is high-energy and a little nervous — excitement about a legitimate open challenger to closed frontier labs, mixed with the usual scramble to figure out what "competitive with frontier models" actually means once independent testing starts.
OpenAI Introduces Codex Site for Agentic Coding Workflows
Alongside GPT-5.6, OpenAI quietly launched a new Codex site — an agentic coding product built on the new model family, aimed at developer tooling and rapid app-building through ChatGPT Sites [1][2]. This isn't a chatbot wrapper; it's a workflow surface designed for AI agents to build, ship, and iterate on software with a human steering rather than typing.
It's a small feature next to the model launch itself, but it's the practical expression of the same shift: fewer humans writing code line-by-line, more humans setting direction and reviewing output. The "vibe coding" language showing up in hackathon and personal-project chatter on X is the informal name for what's becoming a formal product category [4].
What This Means For Your Business
Every story today points at the same inflection: the bottleneck in software is no longer writing code, it's deciding what to build, how much to spend getting there, and which model tier deserves the job. GPT-5.6's three-tier pricing structure and Codex's agentic workflows are OpenAI operationalizing that shift — they're selling you a routing problem, not just a smarter assistant. Meanwhile Inkling and Kimi K3 make clear that the "just use the frontier API" era is ending; open, customizable, self-hostable models are now credible options for teams who want control over cost, data, and behavior.
For companies still evaluating AI as "which model is smartest," you're asking the wrong question. The real question is architectural: do we build an orchestration layer that routes tasks across cheap and expensive models, open and closed weights, based on judgment about what each task actually needs? That's a capability gap most organizations haven't closed yet, and it's exactly where competitive advantage is moving. The code to call any of these models is trivial and free. The judgment to architect when, how, and which one to use — that's the product now.
Key takeaway: The industry just handed you three tiers of intelligence, an open-weight challenger, and an agentic coding surface — in one week. The differentiator isn't access to models anymore, it's the orchestration judgment to use the right one, at the right cost, for the right task.
Sources
- https://openai.com/index/previewing-gpt-5-6-sol/
- https://openai.com/index/gpt-5-6/
- https://www.coderabbit.ai/blog/gpt-5-6-sol-and-terra-benchmark
- https://techcrunch.com/2026/07/15/thinking-machines-amps-up-its-bet-against-one-size-fits-all-ai-with-its-first-open-model-inkling/
- https://fortune.com/2026/07/15/what-is-mira-murati-thinking-machines-first-ai-model-inkling/
- https://thinkingmachines.ai/
Stay ahead of AI
No spam. Unsubscribe anytime.
Want to go deeper?
Reading the news is one thing. Exploring the frontier is another. See what we're building.