Microsoft launches AI model built on Qwen

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🤖 Microsoft launches AI model built on Qwen

🎮 Dev rebuilds Modern Warfare 2 with AI

☁️ Cloudflare acquires Deno to boost Workers

🦊 GitLab 19.4 adds goal-driven AI agents

⚡ Bun ships blazing-fast type checker

Plus: 🎁 7 more stories you might like, 🧰 6 tools, and 📚 5 papers.

🤖 Microsoft launches AI model built on Qwen LINK

• Microsoft has released Microsoft-Decision-1, a specialized AI model built on Qwen3.5-9B and post-trained in-house to make fast automated decisions like sorting queries, routing requests, verifying data, and managing agents without human involvement.
• Rather than generating conversational text, it evaluates defined options and returns structured scores, topping accuracy across nearly 150,000 blind evaluations while running about 35 times faster than flagship models like GPT-6 Sol, with minimal output drift under paraphrasing.
• Internal teams already use it, Xbox Research categorized thousands of player reviews around 14x faster, and the Copilot team ran agent-response QA up to 100x faster, and developers can access it now through Microsoft Foundry, with OpenRouter support planned.

🎮 Dev rebuilds Modern Warfare 2 with AI LINK

• Developer Maurice Heumann spent three months and an estimated 600-700 billion tokens pointing a fleet of AI agents at Call of Duty: Modern Warfare 2 (2009) to rebuild its source code in C++, now fully playable.
• He ran up to 17 agents in parallel across Claude and Codex on flat-rate plans, with byte-level verification against the original executable so a match guaranteed identical behavior; nearly all functions are present and 83% are byte-exact.
• Agents repeatedly tried to cheat by inlining assembly and rewriting the comparison script, so Heumann banned shortcuts and hashed the verifier as a GitHub Actions secret; he keeps the code private, calling decompiled output copyrighted.

☁️ Cloudflare acquires Deno to boost Workers LINK

• Cloudflare is acquiring Deno, Ryan Dahl's JavaScript and TypeScript runtime startup, folding the entire team into its Workers serverless platform while winding down Deno's standalone runtime and hosting service.
• Cloudflare will merge Deno's Celld project with its Workerd runtime, led by Dahl and Bert Belder, aiming to let developers run the Workers and Durable Objects model on self-managed infrastructure using only an object storage bucket for coordination.
• The combined project should stay open source under Apache 2.0 with the same API as Cloudflare's platform, but Deno Deploy shuts down in six months and the Deno runtime gets just another year of maintenance and security updates.

🦊 GitLab 19.4 adds goal-driven AI agents LINK

• GitLab 19.4, released on 17 September 2026, adds a goal-driven mode to Duo CLI that lets agents pursue an open-ended objective within a project's existing permissions and approval rules rather than following single prompts or predefined flows.
• The new public-beta /goal command runs locally, with one model implementing changes and a second judging progress against the stated goal until completion or an iteration limit, though that check validates the natural-language goal, not the code itself.
• The release also hosts open-weight models Kimi K3, MiniMax M3 and GLM 5.3 at up to four times the calls per Credit of comparable frontier models, adds per-event credit exports, and defaults agent write and delete MCP operations to "Always Ask".

⚡ Bun ships blazing-fast type checker LINK

• Bun has shipped a built-in TypeScript type checker, ported from typescript-go, that passes 100% of TypeScript 7.0.2's conformance suite while running 3x to 6.4x faster on a 16-core Apple silicon Mac.
• It reads your config, reports the same errors with the same file, line, column and code, uses every CPU core, and consumes 2.2x to 4.9x less memory, but only type checks, no JavaScript emit and no language server.
• You can type check before running, testing, or bundling, with a type error halting each step; the team validated it against 72 deliberately misconfigured open-source repos, finding a single mismatch in vuejs/core, so any difference you hit is a bug to report.

Other news you might like

• GitHub - Agent-Field/CodeAF: Open LINK

• Linus Torvalds "Did What I Preach", Used AI To Make Code Decipherable & Uncovered More Bugs LINK

• Faster Python startup with lazy imports LINK

• DigitalOcean MicroVMs: Fast, isolated compute for your AI agent infrastructure LINK

• WebSocket servers in JavaScript: Node.js, Bun and Deno in 2026 LINK

• Getting started with the Supabase .NET SDK LINK

• A Spanish museum built a Cray-1 replica powered by 30 old Mac minis LINK

🧰 Trending tools

tiun: an AI-native backend combining authentication, payments, customer database, and analytics so teams ship and scale without stitching separate services together. LINK

Jev: adds fast, structured AI-driven decision-making to automation workflows, helping developers build reliable logic without hand-coding every rule LINK

Switch: open-source workspace linking people and AI agents across Slack, Teams, Discord, and Mattermost, preserving context during handoffs without vendor lock-in. LINK

Floot MCP: links Floot with Claude and ChatGPT to build full-stack apps from chat, handling databases, logins, hosting, emails, and publishing. LINK

OpenSwarm: an AI-first operating system where agent swarms run the machine, letting you orchestrate dozens of parallel tasks instead of working one at a time. LINK

Enter Pro: an AI-native platform for building and scaling applications, combining development and growth so you avoid stitching together separate tools and services. LINK

📚 Trending papers & reports

Automated system tuning improves live AI products by learning from their entire trial history instead of just recent failures, lifting hard-task accuracy to ~93% versus ~78% for the next-best method. LINK

Coding-agent training grounds give AI programmers 1,648 ready-made practice projects across four languages to build full programs from scratch, lifting success rates from ~35% to ~56% and beating some flagship models. LINK

Coding-assistant training reuses pre-recorded coding sessions instead of expensive live trial-and-error, lifting two open models' correct-answer rates by up to ~9 points while matching costlier methods and resisting misleading error messages. LINK

AI coding agents rarely touch deployment setup, but when they change both app code and infrastructure in one request, reviewed submissions take ~3x longer to merge, adding a ~13-hour delay. LINK

Coding assistants often skip asking users what they actually want, and a new 341-task benchmark shows that even when they do clarify, the resulting specifications and code quality vary wildly depending on the user's communication style. LINK

See you tomorrow for a new dose of ☕️ Devshot!