- unwind ai
- Posts
- China's New Reasoning Model Trained in Just $535k
China's New Reasoning Model Trained in Just $535k
PLUS: New Gemini 2.5 Flash model, Infinite canvas for vibe-coding apps
Shubham Saboo & Gargi Gupta
June 18, 2025
In partnership with
Today’s top AI Highlights:
Command your Terminal coding agents from any browser
Chinese AI startup trains frontier reasoning AI model for $535K
Google releases 3 new Gemini 2.5 models
Pre-built Docker images for 450+ MCP servers
An infinite canvas for vibe-coding production apps
& so much more!
Read time: 3 mins
AI Tutorial
Building good research tools is hard. When you're trying to create something that can actually find useful information and deliver it in a meaningful way, you're usually stuck peicing together different search APIs, prompt engineering for hours, and then figuring out how to get the results into a shareable format. It's a headache, and the results are often inconsistent.
In this tutorial, we'll build an AI Domain Deep Research Agent that does all the heavy lifting for you. This app uses three specialized agents that are built using the Agno, Qwen 3 235B model via Together AI, and use tools via Composio to generate targeted questions, search across multiple platforms, and compile professional reports.
What makes this deep research app different from other tools out there is its unique approach: it automatically breaks down topics into specific yes/no research questions, combines results from both Tavily and Perplexity AI for better coverage, and formats everything into a McKinsey-style report that's automatically saved to Google Docs.
We share hands-on tutorials like this every week, designed to help you stay ahead in the world of AI. If you're serious about leveling up your AI skills and staying ahead of the curve, subscribe now and be the first to access our latest tutorials.
Latest Developments
Your AI agents are doing amazing work on your Mac, but you're stuck monitoring them from that one computer like we're still in the dial-up era. Well, three engineers just decided that was absolutely ridiculous.
One weekend and an unhealthy amount of Claude Code just solved remote terminal access for the vibe coding generation.
They built VibeTunnel, which makes Terminal access as simple as opening a web page.
VibeTunnel proxies your Mac terminal straight into any browser, turning your phone into a command center for monitoring builds, directing AI agents, or debugging that script you left running. It's the missing piece that makes vibe coding truly location-independent.
Key Highlights:
Zero-config remote access - No SSH keys, no complex setups, just drag the app to your Applications folder and start controlling your terminal from any browser in seconds.
Built for AI agents - Built specifically for monitoring Claude Code and other terminal-based AI tools, it lets you monitor long-running builds, or give your agents new tasks without being chained to your desk.
Triple-engine architecture - Ships with Node.js, Swift, and Rust server implementations, letting you choose based on your preferences or deployment needs.
Mobile-ready monitoring - Check on long-running builds, AI agent progress, or system processes from your phone without complex mobile SSH clients.
Instant collaboration mode - Share terminal sessions with colleagues through simple web links, eliminating screen sharing or complex SSH setups.
Smarter Investing Starts with Smarter News
Cut through the hype and get the market insights that matter. The Daily Upside delivers clear, actionable financial analysis trusted by over 1 million investors—free, every morning. Whether you’re buying your first ETF or managing a diversified portfolio, this is the edge your inbox has been missing.
Just when you thought DeepSeek was China's mic drop moment, this Shanghai-based AI company drops M1 and says, "hold my tea."
MiniMax just open-sourced the world's first large-scale hybrid-attention reasoning model, and the benchmarks are making Silicon Valley sweat again.
M1 is a mixture-of-experts model with a proprietary Lightning Attention mechanism, featuring a 1 million token context window that matches Google's Gemini 2.5 Pro and state-of-the-art agentic tool use capabilities. The company trained this 456-billion parameter model using reinforcement learning for just $534,700, proving that you don't need Silicon Valley budgets to build world-class AI.
Key Highlights:
Lightning-fast reasoning - The proprietary hybrid-attention mechanism delivers 8x the context length of DeepSeek R1 while using 70% less compute for deep reasoning tasks, making long-form thinking actually practical.
Model variants and test-time compute - Available in 40K and 80K token variants, with the M1-80K model consistently outperforming M1-40K across most benchmarks, demonstrating clear benefits from scaling test-time compute for complex reasoning tasks.
Benchmark Performance - Both M1 variants deliver some seriously impressive performance. They match or outperform frontier models like Gemini 2.5 Pro, DeepSeek R1, and even Claude 4 Opus on long context reasoning, LiveCode Bench, and agentic tool use. However, the models fall behind in the factuality benchmark.
Availability and Pricing - Weights are available on Hugging Face. The models are also available for free on the MiniMax chat platform. The API pricing is competitive, at $0.4/million input tokens (0-200K) and $1.3/million tokens (200K-1M).
Quick Bites
China's Moonshot AI just dropped Kimi Dev 72B, an opensource coding model that's officially the new software engineering champion - outperforms all opensource models with a 60.4% score on SWE-bench Verified. It even outperforms GPT-4.1 and trails only behind Gemini 2.5 Pro.
What's particularly interesting is their training approach: the model learned by autonomously patching real repositories in Docker environments, only getting rewards when entire test suites passed. Available now on Hugging Face and GitHub.
Google Gemini 2.5 models are now production-ready. The team has released 3 new Gemini models:
Gemini 2.5 Pro (now stable)
Gemini 2.5 Flash (stable with updated pricing), and
New Gemini 2.5 Flash Lite (in preview)
Flash-Lite delivers better performance than the 2.0 models across coding, math, tool-use, and reasoning while maintaining lower latency and supporting the full toolkit including search, code execution, and that coveted 1M token context window. Flash-Lite offers dynamic thinking control that's switched off by default - perfect for high-frequency, cost-efficient tasks.
An open R&D lab Menlo Research has released Jan-Nano 4B model, specifically designed and trained for deep research tasks, and optimized to work seamlessly with MCP servers. The model punches much higher than its weight, outperforming even DeepSeek 671B and GPT 4.5 on the Simple QA benchmark, via agentic tool-use capabilities. You can run it locally through Jan's desktop app.
While the performance comparison might be apples-to-oranges given the different sizes and tooling setups, it's an interesting case study in specialized model design and training.
Tools of the Trade
mcp-containers: Pre-built Docker images for 450+ MCP servers. No manual setup required, just pull the image. The containers are automatically updated daily and run in isolation for secure MCP server deployment with AI agents.
Trieve CLI: A Terminal-based RAG agent that lets you upload documents and ask questions about the knowledge base. It iteratively searches, refines queries, and reasons about what it finds. You can customize the RAG behavior, check upload status, and the responses stream back with source references.
Solar: Build full-stack apps visually with no-code while an agent handles Python code, databases, auth, and more. It’s not just a code generator; it reads logs, fixes bugs, and completes workflows end-to-end.
Awesome LLM Apps: Build awesome LLM apps with RAG, AI agents, MCP, and more to interact with data sources like GitHub, Gmail, PDFs, and YouTube videos, and automate complex work.
Hot Takes
AGI is going to be achieved by a product, not necessarily a “model”. ~
Logan Kilpatrickdating in sf is hard, it’s just like doing sales for your ai startup:
- stuck in a 6-month pilot
- acts of service is the product (no self-serve onboarding)
- ‘thought partner + synergies’ is the wedge
- too much follow-up: closed-lost
- pitching
- waiting for legal to define the partnership
high intent and low conversion ~
Annie Liao
That’s all for today! See you tomorrow with more such AI-filled content.
Don’t forget to share this newsletter on your social channels and tag Unwind AI to support us!
PS: We curate this AI newsletter every day for FREE, your support is what keeps us going. If you find value in what you read, share it with at least one, two (or 20) of your friends 😉
Reply