Hi there, this is your daily ☕️ Devshot.
In today's Devshot:
🤖 xAI launches Grok 4.7
🏆 Xiaomi's MiMo tops open-weight AI models
🧑💻 AWS open-sources new coding agent
🌊 Browsers gain native HTML streaming
🐍 Cloudflare Python Workers hit general availability
Plus: 🎁 8 other news you might like, 🧰 6 tools, and 📚 5 papers.
🤖 xAI launches Grok 4.7 LINK
- xAI shipped Grok 4.7 on Monday, its strongest model so far, live now in the Grok app, Cursor, Grok Build, and the xAI API at the same $2/$6 per million input/output tokens as Grok 4.6.
- The model runs 2.1 trillion parameters, up 40% from 4.6, spends longer on hard problems and double-checks its answers more often, and folds in SpaceX Starlink telemetry and failure logs to reason better about hardware.
- Benchmarks place it a rung below the frontier: 1695 on GDPval versus Claude Fable 5.1's 1735, and on Cursor's CursorBench 4.0 it costs more per coding task than GPT-6 Astra and Claude Sonnet 5.
🏆 Xiaomi's MiMo tops open-weight AI models LINK
- Xiaomi's new MiMo-V2.6-Pro became the top open-weight model on Artificial Analysis' Intelligence Index with a score of 46, edging out xAI's Grok 4.6 and Google's Gemini 3.8 Flash while tying the newly released Grok 4.7.
- The MIT-licensed model is free to download from Hugging Face, ships a 1M-token context window with text, image, audio and video input, runs at ~134 tokens/second, and costs $0.435/$0.87 per million input/output tokens via API.
- Xiaomi trained it with a single large reinforcement-learning run mixing coding, visual, computer-use and security tasks, spending ~$2.6M and adding groupwise reward grading that curbed reward hacking and produced smaller, cleaner code patches instead of brittle workarounds.
🧑💻 AWS open-sources new coding agent LINK
- AWS released Strands Harness, an open source general-purpose AI agent built atop its Strands Agents SDK that gives developers a preconfigured agent runnable locally or deployable to any cloud provider they pick.
- It ships with file, shell and web tools plus built-in context management, memory, persistent sessions, prompt caching and delegation, all running on the local machine except model calls, which default to Bedrock but swap to Anthropic, OpenAI, Google or Ollama in one line.
- AWS says Strands Harness ran 45% cheaper than Claude Code and Codex with comparable accuracy across six benchmarks, crediting its context-management defaults that truncate large tool outputs and compact context once the window crosses a threshold.
🌊 Browsers gain native HTML streaming LINK
- Browsers are getting native out-of-order HTML streaming, letting a page render and stay interactive while slow parts load, with Chrome and Edge 150 shipping the declarative markup and Safari and Firefox signaling support.
- A new template attribute defines insertion points or fallback content, and when the server streams a matching template later, the parser removes the markers, drops intermediate nodes, and inserts the template's contents into that DOM position.
- To block cross-component injection, a template can only patch its immediate parent's siblings and descendants unless placed directly under the root for global scope, while a companion suite of static and streaming DOM methods advances through separate standardization.
🐍 Cloudflare Python Workers hit general availability LINK
- Cloudflare has moved Python Workers to general availability, making Python a fully supported, first-class language on its Developer Platform after roughly two years in preview.
- Under the hood, Python code is compiled to WebAssembly via Pyodide and executed inside Cloudflare's V8-based workerd runtime, with some documented limitations where certain features remain non-functional in the WebAssembly VM.
- For local development, the pywrangler tool, packaged as workers-py on PyPI, runs a full local simulation of the stack, executing code through Pyodide in WebAssembly inside V8 via a 123MB binary.
Other news you might like
- Grafana Alerting: Scale alert routing without scaling complexity using multiple notification policiesLINK
- Transformers now runs llama.cpp quantsLINK
- Changesets v3: Bumps Peer Dependents by Patch and Ships ESM Only with an 88% Smaller InstallLINK
- Uber Redesigns M3DB Sharding with Subclusters to Limit Failure ImpactLINK
- DuckDB Now Ships inside dbt v2LINK
- Encore for RustLINK
- Developers Complain After Z.ai’s ZCode Sends Data To CloudLINK
- Python 3.15 Preview: frozendictLINK
🧰 Trending tools
Prefactor: an evaluation layer that scores agent runs in real time, catching quality regressions and drift before they reach production users.LINK
SKI: adds voice control to AI coding agents like Claude Code and Codex, letting you dictate commands and hear responses hands-free while codingLINK
Die With Me: a Mac app that tracks your Claude and Codex usage, compares it with friends, and offers a chatroom when your allowance runs out.LINK
Dograh: open-source voice agent platform with visual flow builder, telephony, human transfer, and self-hosted QA monitoring across 30-plus model integrationsLINK
Jev: converts unstructured input into typed probabilistic decisions like Choice, Score, and Bool with calibrated probabilities, delivering fast, cheap responses your code can act on directly.LINK
AgentSky: runs managed AI coding agents with auto-recovery, letting you control remote development from WhatsApp, Telegram, Slack, web, API, or CLILINK
📚 Trending papers & reports
Grid control room assistant lets a language model suggest actions for power operators while a rulebook it cannot bypass blocks unauthorized moves, hitting ~94% task success with safety rules holding across every one of 1,416 runs.LINK
Coding fingerprints reveal that different AI models write code in measurably distinct styles, giving buyers a way to tell rival models apart and pick one even when they all pass the same performance tests.LINK
Coding-agent token bills break into three separate savings levers, and the one everyone assumes matters, compressing file reads, saves only ~2% per turn upfront but compounds to overtake fixed savings after about 6 turns.LINK
Coding assistant bias shows that telling an AI your nationality, gender, or experience level changes the code you get back, dropping one model's accuracy by up to ~3.6% while nudging another's up ~3.4 to 5.7%.LINK
Coding agents for official statistics often run flawlessly yet still return wrong answers, and this benchmark shows accuracy comes from clear task specs and retry attempts, not from feeding the model its error messages.LINK
See you tomorrow for a new dose of ☕️ Devshot!