Moonshot's newly released Kimi K3 is a 2.8 trillion parameter model that marks their most capable release to date. This article explores its performance benchmarks, pricing structures compared to Anthropic’s Sonnet series, and how it handles complex tasks like SVG generation. Through the lens of the pelican benchmark, the author examines the relationship between reasoning token consumption, cost, and a model's spatial awareness.
Main points:
- K3 is described as an open 3T-class model with high performance on long-horizon knowledge work.
- The pricing structure for K3 represents a significant increase over previous Moonshot models.
- Testing the pelican prompt reveals high reasoning token usage and substantial costs per task.
- While useful for checking spatial awareness, simple benchmarks fail to test agentic tool calling.
Main points:
- K3 is described as an open 3T-class model with high performance on long-horizon knowledge work.
- The pricing structure for K3 represents a significant increase over previous Moonshot models.
- Testing the pelican prompt reveals high reasoning token usage and substantial costs per task.
- While useful for checking spatial awareness, simple benchmarks fail to test agentic tool calling.
OpenAI has released the GPT-5.6 model family, comprising three sizes: Luna (smallest), Terra, and Sol (largest). These models feature a one million token context window, 128,000 maximum output tokens, and a knowledge cutoff of February 16th, 2026. Key updates to the API include programmatic tool calling via JavaScript orchestration, native multi-agent support for parallel task execution, explicit prompt cache breakpoints, and an option to receive unresized images in requests.
* Three model tiers: Luna, Terra, and Sol
* Improved performance in long-running agentic professional workflows
* New API features including programmatic tool calling and multi-agent orchestration
* Three model tiers: Luna, Terra, and Sol
* Improved performance in long-running agentic professional workflows
* New API features including programmatic tool calling and multi-agent orchestration
Simon Willison explores his latest approach to running untrusted Python code safely within applications by utilizing MicroPython inside a WebAssembly (WASM) sandbox. The project addresses the security risks of plugin systems where code normally executes with full privileges, potentially leading to data leaks or system compromise. By leveraging wasmtime and an alpha package called micropython-wasm, Willison demonstrates how to enforce memory and CPU limits while providing controlled access to host functions through a custom thread-based request queue for persistent interpreter state.
Main topics:
- Security challenges in Python plugin systems
- Advantages of WebAssembly as a sandboxing technology
- Building the micropython-wasm alpha package
- Implementation details for persistent state and host functions
- Integration with Datasette Agent to execute code via LLMs
Main topics:
- Security challenges in Python plugin systems
- Advantages of WebAssembly as a sandboxing technology
- Building the micropython-wasm alpha package
- Implementation details for persistent state and host functions
- Integration with Datasette Agent to execute code via LLMs
* **Rapid Model Competition:** The title of "best model" shifted frequently between Anthropic (Claude), OpenAI (GPT), and Google (Gemini) during November 2025.
* **Advancements in Coding Agents:** Using Reinforcement Learning from Verifiable Rewards, coding agents transitioned from being unreliable to becoming dependable "daily-driver" tools for professional work.
* **Rise of Personal AI Assistants ("Claws"):** The emergence of highly popular local personal assistant projects like OpenClaw (formerly Warelay), leading to increased demand for hardware like Mac Minis to run them locally.
* **Gemini 3.1 Pro Release:** Google released an updated model with improved capabilities in visual/SVG generation.
* **Google Gemma 4 Series:** The release of highly capable open-weight models from a US company.
* **GLM-5.1 Release:** A massive, 754B parameter (1.5TB) open-weight model released by the Chinese lab GLM.
* **High-Performance Local Models:** Small, laptop-runnable open-weight models like Qwen3.6-35B-A3B began wildly outperforming expectations and competing with much larger frontier models in specific tasks.
* **Advancements in Coding Agents:** Using Reinforcement Learning from Verifiable Rewards, coding agents transitioned from being unreliable to becoming dependable "daily-driver" tools for professional work.
* **Rise of Personal AI Assistants ("Claws"):** The emergence of highly popular local personal assistant projects like OpenClaw (formerly Warelay), leading to increased demand for hardware like Mac Minis to run them locally.
* **Gemini 3.1 Pro Release:** Google released an updated model with improved capabilities in visual/SVG generation.
* **Google Gemma 4 Series:** The release of highly capable open-weight models from a US company.
* **GLM-5.1 Release:** A massive, 754B parameter (1.5TB) open-weight model released by the Chinese lab GLM.
* **High-Performance Local Models:** Small, laptop-runnable open-weight models like Qwen3.6-35B-A3B began wildly outperforming expectations and competing with much larger frontier models in specific tasks.
Explores how to use the llm CLI tool within a script's shebang line by utilizing the env -S pattern. This approach allows plain text files to become executable via large language models, treating file contents as prompts. The article covers using flags like -f for fragments and -x for code extraction, integrating tools such as llm_time, and leveraging YAML templates with parameters or embedded Python tool functions to create highly functional scripts.
Main topics:
- Using env -S with the llm CLI
- Making text files executable via LLM
- Integrating tools within shebang lines
- Utilizing YAML templates for prompts
- Embedding Python functions in templates
Main topics:
- Using env -S with the llm CLI
- Making text files executable via LLM
- Integrating tools within shebang lines
- Utilizing YAML templates for prompts
- Embedding Python functions in templates
Simon Willison discusses why requesting HTML rather than Markdown as an LLM output format can significantly enhance technical explanations. While token constraints previously favored Markdown, modern models benefit from the ability of HTML to incorporate SVG diagrams, interactive widgets, and improved navigation. The article provides prompt examples for reviewing pull requests via HTML artifacts and showcases a GPT-5.5 generated explanation of a Linux security exploit that uses CSS and JavaScript to create a rich documentation experience.
An exploration of the new Qwen3.6-27B open weight model, which claims flagship-level agentic coding performance that surpasses previous larger MoE models while being significantly smaller in size. The author tests a quantized version using llama-server and demonstrates its impressive ability to generate complex SVG graphics locally.
Key points:
- Qwen3.6-27B outperforms the older Qwen3.5-397B-A17B on coding benchmarks.
- Dramatic reduction in model size from 807GB to approximately 55.6GB for the base version.
- Successful local execution using a 16.8GB quantized GGUF version via llama.cpp.
- High-quality SVG generation capabilities for complex prompts like a pelican riding a bicycle.
Key points:
- Qwen3.6-27B outperforms the older Qwen3.5-397B-A17B on coding benchmarks.
- Dramatic reduction in model size from 807GB to approximately 55.6GB for the base version.
- Successful local execution using a 16.8GB quantized GGUF version via llama.cpp.
- High-quality SVG generation capabilities for complex prompts like a pelican riding a bicycle.
Google DeepMind has released four new open-source, vision-capable LLMs under the Apache 2.0 license – Gemma 4, with sizes ranging from 2B to 31B parameters, and a 26B-A4B Mixture-of-Experts model. The models are notable for their intelligence-per-parameter ratio, with the smaller models (E2B and E4B) utilizing Per-Layer Embeddings to maximize efficiency.
The models support both vision and audio input, although audio functionality is not yet fully implemented in tools like LM Studio or Ollama. Testing with LM Studio showed varying results, with the 31B model experiencing output issues. The author also experimented with the models through the llm-gemini API, generating SVG images of a pelican riding a bicycle to assess their visual capabilities.
The models support both vision and audio input, although audio functionality is not yet fully implemented in tools like LM Studio or Ollama. Testing with LM Studio showed varying results, with the 31B model experiencing output issues. The author also experimented with the models through the llm-gemini API, generating SVG images of a pelican riding a bicycle to assess their visual capabilities.
Simon Willison explores "vibe coding" - building macOS apps with SwiftUI using large language models like Claude Opus 4.6 and GPT-5.4, without extensive coding knowledge. He successfully created two apps, Bandwidther (network bandwidth monitor) and Gpuer (GPU usage monitor), demonstrating the potential of this approach. The process involved minimal prompting and iterative development, leveraging the LLMs' capabilities for both code generation and feature suggestions.
While acknowledging the need for caution regarding the apps' accuracy, Willison highlights the efficiency and accessibility of building macOS applications in this manner.
While acknowledging the need for caution regarding the apps' accuracy, Willison highlights the efficiency and accessibility of building macOS applications in this manner.
Starlette 1.0 has been released, and Simon Willison explores its new features by leveraging Claude’s skill‑building capabilities. He demonstrates how Claude can clone the Starlette repository, generate a comprehensive skill document with code examples, and even create a fully functional task‑management app complete with database, API endpoints, and Jinja2 templates—all generated and tested by Claude itself. The article highlights the practical benefits of integrating an LLM as a coding agent, showcases the new lifespan mechanism, and reflects on the growing popularity of Starlette as the foundation of FastAPI.
