Kimi K3 arrives at 2.8 trillion parameters — the largest open model ever released

🧠 LAUNCH

Kimi K3 arrives at 2.8 trillion parameters — the largest open model ever released.

Kimi K3 lands with 2.8T total parameters but only 50B active per token via Mixture of Experts, delivering Opus 4.8-class performance at Sonnet 5 pricing. Moonshot AI just proved that MoE economics can give you frontier intelligence without frontier costs — and they open-sourced it. If you're budget-constrained on inference but need top-tier reasoning, this is your new baseline to benchmark against. Read more →

Thinky's Inkling gets the Latent Space deep-dive — and the open model landscape shifts again. Inkling (975B total, 41B active) with its Apache 2.0 license already turned heads on Jul 16, but Latent Space's independent benchmarking puts it in direct comparison with K3. Two massive MoE models in one week — the open-source frontier just leapfrogged. Read more →

💡 INSIGHT

Cursor joins the Fable 5 production club — and the pattern is now undeniable.

Cursor publishes a "Working at the frontier" case study documenting how they deploy Fable 5 for their hardest 1% of coding problems. This is the third major dev tool company (after Base44 and Cognition) publicly vouching for Fable 5 in production within a single week. When your top-tier coding agent, your app builder, and your autonomous dev platform all converge on the same model, that's not marketing — that's a market signal. Read more →

Anthropic publishes a CISO's guide to agentic AI — selling to security teams, not around them.

Anthropic drops a practical risk-assessment framework for enterprise security leaders evaluating agentic AI deployments. The title says it all: "Zero risk isn't the job." Instead of pretending agents are risk-free, they're giving CISOs a structured way to evaluate acceptable risk — a mature play that signals Anthropic is going after enterprise procurement by winning over the people who usually say no. Read more →

The state of open source AI report lands — with perfect timing. A comprehensive analysis of the open-source AI ecosystem drops the same week K3 and Inkling redraw the frontier. With 353 likes on HN, this is the data-driven counterpart to the vibes-driven hype cycle. (353 likes | 260 RTs) Read more →

🔧 TOOL

Claude Code gets /fork — split conversations into parallel background sessions.

Claude Code v2.1.212 ships /fork, which copies your current conversation into a background session (its own row in claude agents) while you keep working in the original. This is the parallel-task workflow developers have been requesting — investigate a rabbit hole without losing your main context. Also adds session-wide WebSearch rate limits to prevent runaway tool use. Read more →

CrewAI promotes Skills Repository to stable — reusable skill abstractions for agent teams. CrewAI 1.15.4 graduates the Skills Repository from experimental, giving multi-agent workflows a first-class primitive for packaging and reusing agent capabilities. Combined with last week's interception hooks, CrewAI is assembling a real production framework, not just a demo harness. Read more →

llama.cpp adds OpenCL MoE kernel — just in time for the giant MoE models. Build b10066 ships a new OpenCL kernel for Q6_K quantized MoE inference. With K3 at 2.8T and Inkling at 975B both being MoE architectures, efficient inference on non-NVIDIA hardware just became urgent — and llama.cpp is already there. Read more →

🔬 RESEARCH

AI finds real bugs in a zero-knowledge VM — where correctness is the entire product. AI-powered code analysis discovered genuine vulnerabilities in OpenVM's zkVM, a system where a single bug can compromise cryptographic guarantees. This isn't "AI catches a typo" — it's finding flaws in code that formal verification was supposed to guarantee correct. Concrete evidence that AI bug-finding adds value even in the highest-assurance codebases. (80 likes) Read more →

Simon Willison stress-tests Kimi K3 — the reality check the launch hype needs. Willison runs his pelican benchmark and adversarial instruction-following tests against K3, providing grounded independent evaluation against the official numbers. If you're deciding whether to deploy K3, start here before trusting the launch blog. Read more →

Newer models, same advantage? What happens to your fine-tuning investment across generations. Dharma AI analyzes whether optimizations built for one model generation (fine-tuning, prompt engineering, custom evals) transfer to newer models or need to be rebuilt from scratch. If you've invested months tuning prompts for a specific model, this tells you whether that work survives an upgrade. Read more →

📝 TECHNIQUE

NVIDIA and Hugging Face team up to make multi-GPU diffusion fine-tuning accessible. NeMo Automodel wraps multi-GPU training orchestration behind the Diffusers API you already know — scaling video and image model fine-tuning without learning a new framework. If you've been doing single-GPU LoRA and hitting quality ceilings, this is your on-ramp to full fine-tuning at scale. Read more →

Train a generative kick drum model on 6GB VRAM — diffusion audio hits consumer hardware. A step-by-step guide to training a diffusion model for audio synthesis on an old Linux desktop with 6GB VRAM. Generative audio was a "rent 8x H100s" problem six months ago — now it runs on hardware you already own. (86 likes | 53 RTs) Read more →

🏗️ BUILD

KlaatCode: open-source coding agent with smart model routing claims 10x cost reduction. Routes queries to the right model tier automatically — simple completions go to cheap models, complex reasoning gets the frontier. The cost-optimization angle differentiates it from the crowded "open-source Claude Code alternative" space. (104 likes) Read more →

Simon Willison builds an LLM cliché highlighter — spot the "delves" before you ship. A practical tool that flags LLM verbal tics (delve, tapestry, landscape, nuanced) in any text. If you're editing AI-generated content — and in 2026, who isn't — this catches the tells before your readers do. Read more →

ReasonGate: a prompt injection firewall that explains why it blocked you. Open-source security middleware that doesn't just binary allow/deny — it tells you why a prompt was flagged. Explainability in security tooling is rare and invaluable for debugging false positives without disabling protection entirely. Read more →

🎓 MODEL LITERACY

Mixture of Experts (MoE) — Active vs. Total Parameters: Both K3 (2.8T total, 50B active) and Inkling (975B total, 41B active) launched this week as massive MoE models, so understanding the distinction matters. Total parameters represent everything the model knows — its full capacity spread across dozens of specialized "expert" sub-networks. Active parameters are the small subset that fires for any given token — a learned router picks 2-4 experts per input. This is why K3 can have 2.8 trillion parameters but run at Sonnet-tier pricing: you're only paying inference compute on the 50B that activate, while benefiting from the knowledge capacity of the full 2.8T. Think of it like a hospital with 200 specialists but only 3-4 consult on your case.

⚡ QUICK LINKS

• OpenAI Python SDK v2.46.0: Project-scoped service account API keys for isolated credentials per project. Link
• OpenAI Node SDK v6.48.0: Mirrors Python's new service account key endpoints. Link
• Anthropic Google Cloud TypeScript SDK v0.0.6: Bumps google-auth-library to fix Vertex AI auth compatibility. Link
• Vercel AI SDK v7.0.31: Fixes denied tool approval state and input event ordering for human-in-the-loop flows. Link

🎯 PICK OF THE DAY

K3's real disruption isn't the benchmark scores — it's what MoE economics means for the frontier. Moonshot AI releasing a 2.8-trillion-parameter model that runs at Sonnet pricing proves something the industry has been theorizing for two years: MoE architectures can deliver Opus-class intelligence without Opus-class costs. The implication is profound — if frontier intelligence is now achievable at mid-tier pricing and open-sourced, then the moat is no longer model quality. It's ecosystem lock-in: tool integrations, agent frameworks, enterprise compliance, developer muscle memory. Anthropic, OpenAI, and Google aren't competing on IQ anymore — they're competing on switching costs. For builders, the action is clear: abstract your model dependency now, because the commodity era of frontier AI just started, and loyalty to a single provider is about to get very expensive relative to what's available for free. Read more →

Until next time ✌️