This Week in AI

Model & Product Releases: Kimi K3, Inkling, Gemini Flash

By Scott Andersen, Head of AI Services and Co-Founder Leave a comment
AII Model Releases

Moonshot AI releases Kimi K3 (2.8T parameters)

  • K3’s scale pushes China into the frontier tier of open-weight model development.
  • Its top performance on coding benchmarks signals a shift: Chinese labs are now competing directly in agent-builder ecosystems rather than just chat models.
  • The open-weights posture is strategically important — it positions Kimi as a foundation for domestic enterprise adoption and sovereign AI stacks.

Thinking Machines ships a new open-weight flagship

  • Continues the trend of regional labs producing competitive open models.
  • Reinforces the global pivot away from closed-model dependence for enterprise workloads.
  • Adds pressure on Western vendors to justify closed-weight pricing.

Google releases three new Gemini Flash variants focused on agents and efficiency

On July 21, Google launched:

  • Gemini 3.6 Flash: Workhorse model with better coding/knowledge/multimodal performance, ~17% fewer output tokens than 3.5 Flash (up to 65% in some benchmarks), and lower cost.
  • Gemini 3.5 Flash-Lite: Ultra-fast/high-throughput option (~350 tokens/sec) for agentic workflows.
  • Gemini 3.5 Flash Cyber (with CodeMender agent): Specialized for vulnerability detection/patching; found unique issues in tests (e.g., 55 in V8 engine) at competitive frontier performance and lower cost. The flagship 3.5 Pro remains in partner testing. These target scalable, cost-effective agent deployments.

Model Releases & Delays

  • Gemini 3.5 Pro’s third delay — Gemini 3.5 Pro reportedly missed its July 17 target for the third time after falling short on coding and complex reasoning in testing, following Google scrapping the original base model in June and restarting pretraining. Google is reportedly exploring a stopgap Gemini 3.6 Flash release, and has published no official model card, pricing, or benchmarks.
  • Kimi K3’s open-weight shock — Moonshot AI’s model stunned the industry after the 2.8-trillion-parameter open model took the top spot on a major coding leaderboard days earlier, beating Claude Fable 5 on that benchmark. The caveat: K3 ranks around ninth on general chat, so it is a coding and agent specialist rather than an all-around frontier replacement. Moonshot AI has promised Kimi K3’s open weights by July 27, 2026, about eleven days after the July 16 API launch.
  • Claude Fable 5’s free window closed — Anthropic’s free access period for Fable 5 expired July 19, forcing a decision on what’s next — an Opus 5 release, a fourth extension of free access, or a shift to a credits-based model. The timing landed awkwardly alongside Kimi K3’s free-weights announcement, though Anthropic reportedly had a strong week otherwise, including a confidential IPO filing.
  • Concentrated open-weight release window ahead — DeepSeek’s V4 stable release lands July 24, ending the preview-build churn that has kept cautious enterprises from moving production workloads onto it, and Kimi K3’s open weights go free on July 27. DeepSeek’s roughly $0.44 per million output tokens is already the price floor the industry gets measured against.

AI Product releases

  • Thinking Machines launched Inkling. The 975-billion-parameter open-weight model is the company’s first general-purpose release and performed competitively on agent-related evaluations. Reuters
  • Moonshot AI released Kimi K3. The 2.8-trillion-parameter open-weight model achieved strong coding and agent results, further strengthening China’s open-model ecosystem. ODSC Weekly Review
  • Cadence launched an AI “super agent” for circuit-board and chip-package design. The release shows agentic AI moving into specialized engineering and industrial workflows. Reuters
  • Apple Intelligence moved closer to release in China. Its localized architecture will incorporate Chinese models and comply with national registration requirements.

Leave a Reply