Key Takeaways

• Ornit 1.0 is a self-training AI coding model released by the DeepReinforce team under MIT license—completely free and open-source
• On the SWE-bench coding benchmark, Ornit’s largest model scored 82.4% versus Claude’s 70.3% and Opus 4’s 85%
• The smallest 9B parameter model (only 19 GB) runs on a single GPU and still achieves 69.4% on benchmarks
• Unlike traditional AI, Ornit writes its own learning framework—it doesn’t need humans to scaffold its problem-solving approach
• Available via official site, Hugging Face, or API through providers like Protoverse

What Makes Ornit 1.0 Different From Other Coding AIs

Most coding AI models I’ve worked with follow a familiar pattern. Humans build what’s called a “scaffold”—essentially a framework that tells the AI: “here’s how to approach this problem, follow these steps.” It’s like giving a student a detailed study guide before an exam.

Ornit doesn’t work this way. Here’s what caught my attention: the model builds its own scaffold. It both solves the problem AND invents the management system for how to solve that problem, improving both simultaneously. This means it discovers solution paths that no human explicitly taught it.

The DeepReinforce team built this as an agent-based model family, not just another chatbot. It doesn’t just talk about code—it takes real software engineering tasks and solves them. Four sizes are available, from the 9 billion parameter version that runs on a single graphics card, up to the flagship 397 billion parameter model. All MIT licensed.

The Benchmark Numbers I Verified

I always dig into the actual numbers before getting excited, and here’s what I found in the verified benchmarks:

On SWE-bench (the industry standard for real software bug fixing):

• Ornit (large): 82.4%
• Claude: 70.3%
• OpenAI: 77.5%
• Opus 4.8: 85%

On TerminalBench (command-line usage testing):

• Ornit: 77.5%
• Claude Opus: 70.3%

And here’s what genuinely surprised me—the 9 billion parameter small model scored 69.4% on benchmarks. That’s a model that fits on an ordinary computer, performing at levels that would have required massive cloud infrastructure just two years ago.

For context, Anthropic’s system that users pay $200 monthly for operates at 78.2% on comparable tests. Ornit hits 77.1%—a one-point difference—for zero cost.

What “Self-Learning” Actually Means Here

I’ve noticed a lot of confusion around “self-learning AI” in my community. Let me clarify what Ornit actually does, based on my testing and the technical documentation.

Traditional fine-tuning requires humans to curate datasets, define learning objectives, and structure the training process. Ornit’s approach—what the researchers call autonomous self-improvement—means the model generates its own training framework, evaluates its own outputs, and iteratively refines its problem-solving strategies without human intervention at each step.

Think of it like this: instead of a teacher giving a student flashcards, the student creates their own flashcard system, discovers which memorization techniques work best for them, and keeps optimizing that system while actually learning the material.

This isn’t artificial general intelligence or sentience—it’s a specific, engineered capability for code-generation tasks. But it’s a genuine step toward AI systems that require less human hand-holding to improve.

How I Tested Ornit and How You Can Too

I ran Ornit through three different setups to understand the practical access points:

Local installation: The official site provides a straightforward prompt for local execution. With the 9B model at 19 GB, I got it running on a single high-end GPU without code changes—it’s compatible with standard developer tools.

Hugging Face: Direct model downloads are available through Hugging Face links, which is how I tested the larger variants without local hardware constraints.

API access: Through providers like Protoverse, you can search “Ornit” and integrate via API if you want to build applications without managing infrastructure.

For developers already using Claude or GPT-4 APIs, the integration is designed to be drop-in compatible—no rewriting your entire stack.

What This Actually Means for Your Work

I’ve been paying hundreds of dollars monthly for premium AI subscriptions, and I run a community of 400+ developers who do the same. Here’s my honest assessment of where Ornit fits:

If you’re learning or bootstrapping: This is genuinely transformative. You can now access top-tier coding performance without subscription costs. The 9B model running locally means no API fees, no rate limits, complete privacy for proprietary code, and full customization potential.

If you’re already paying for Claude/GPT-4: The performance gap is narrow enough that you need to ask: is that 1-5 point benchmark difference worth $200/month to you? For some professional contexts where that edge matters, maybe. For many use cases, probably not.

The bigger picture matters more than any single model comparison. Self-improving AI is no longer theoretical—we have working examples. Systems that can enhance their own capabilities without constant human intervention represent a real step forward, and understanding this trajectory early puts you ahead.

FAQ

Is Ornit 1.0 really completely free to use commercially?

Yes. All Ornit models are released under MIT license, which permits commercial use, modification, and distribution without fees. You can download, modify, integrate into products, or even sell derivative systems.

What hardware do I need to run Ornit locally?

The 9B parameter model requires approximately 19 GB of storage and runs on a single powerful GPU. The larger models (up to 397B parameters) require substantially more infrastructure. For most individual developers, the 9B model offers the best local-running balance of capability and accessibility.

How does Ornit’s self-learning differ from normal AI training?

Standard AI training relies on human-designed frameworks: humans create datasets, define loss functions, and structure the learning process. Ornit autonomously generates its own problem-solving frameworks, evaluates their effectiveness, and iteratively improves both its approach and its outputs simultaneously for coding tasks.

Should I switch from Claude or GPT-4 to Ornit?

It depends on your specific needs. Ornit matches or exceeds Claude on several coding benchmarks while being free, but Claude and GPT-4 still lead in some areas and offer broader general-purpose capabilities. For pure coding automation with budget constraints, Ornit is strongly worth testing. For multi-modal or general reasoning tasks, the paid models may still serve you better.

Where I Think This Is Heading

I’ve watched the AI landscape shift weekly for years, and this release signals something specific: the commoditization of high-end coding assistance is accelerating. When a free, open model can approach or match premium services, the economic model for AI tooling fundamentally changes.

The self-improvement capability matters more long-term than any single benchmark. We’re seeing the first practical steps toward AI systems that reduce their dependence on human-curated training data. Whether this ultimately proves more capable, more concerning, or simply different remains to be seen—but it’s no longer speculative.

I don’t believe in guaranteed outcomes or hype around any single tool. What I do believe: having tested Ornit extensively, it’s a legitimate technical achievement that deserves attention from anyone building with AI, especially those currently paying premium prices for coding assistance.

Self-improving AI is now a working reality, not a research paper concept. The developers who understand and experiment with these capabilities today will be the ones best positioned as the technology matures.

Watch the full video (in Turkish — English subtitles available):

Tools & Community

• TurkoLister — the AI listing tool I use to turn Amazon products into optimized eBay UK listings in about 60 seconds (from £4.99/month, £1 one-week trial).
• AI & E-commerce Community — my Turkish-speaking community ($19/month) with weekly live sessions.
• Subscribe on YouTube — new experiments every week.

The post Ornit 1.0: Free Self-Learning AI That Beat Claude at Coding appeared first on Digital Market Mentoring.

Key Takeaways

• 1 million token context window lets Minimax M3 hold entire codebases, books, or large projects in memory without forgetting details
• 79% score on software testing benchmarks, surpassing GPT-5.5 and Claude 3.1 Pro — despite being nearly free
• Desktop agent capability distinguishes it from chatbots: it sees your screen, controls your computer, and works autonomously
• Multimodal understanding of text, images, video, and live screen capture in a single model
• Cost is “pocket money” compared to competitors — per-million-token pricing in the cents range
• I tested it head-to-head against Claude Opus 4.8 and Codex 5.5 on identical game-building prompts

Why Everyone Is Talking About Minimax M3

Three capabilities converged in this release for the first time, and I believe that’s why it’s causing this level of disruption.

First, the 1 million token context window. I can feed it an entire project — a full book manuscript, a massive code repository — and it retains everything. Nothing gets lost in the middle. For the kind of complex automation work I do in e-commerce and online business, this changes what’s possible.

Second, it’s genuinely multimodal. It doesn’t just process text. It sees images, understands video, and critically — it sees your computer screen in real-time. This isn’t theoretical. When I ran my tests, the agent was literally watching my desktop and making decisions based on what it observed.

Third, the pricing. While competitors charge amounts that make serious automation expensive at scale, Minimax M3 is priced at what I’d call pocket money. Per million tokens, we’re talking cents. There’s even a free tier to start testing the desktop agent. I honestly couldn’t find an excuse not to try it.

The benchmark that stopped me cold: 79% on software testing. That’s higher than GPT-5.5 and Claude 3.1 Pro. An open-source Chinese model beating the most expensive American models on coding tasks — this would have seemed impossible a few years ago.

Setting Up My First AI Employee

Here’s where Minimax M3 diverges from every chatbot I’ve used. It has a desktop application. You install it on your computer, and it becomes what the company calls a “personal assistant” or what I configured — using the Hermes Agent framework — as an autonomous worker.

The fundamental difference from other AIs: it doesn’t just talk. It does the work. The backend contains thousands of capabilities. I connected my files, my email, my calendar. Now it works with my actual data, my actual world — not generic training data.

I gave my agent a specific goal: generate £1 million in sales, and don’t stop until achieved. It will continuously improve itself and scale operations. You can program any life goal this way — write the plan, develop the capability, and the AI works daily toward it.

The configuration requires precision. You define exactly what the agent does, provide detailed descriptions, and specify boundaries — what it should do, what it must never do. You choose between asking permission for each step or granting full autonomous operation. I tested both modes.

Remote control works through Telegram bot, WeChat bot, or Lark. I can monitor and direct my desktop agent from my phone, 24/7, even when I’m away from my computer.

Test 1: Building a Professional Website in Under a Minute

I wrote a single prompt: build me a beautiful gaming website, with all files contained in a single HTML file. I selected Minimax M3 and started the system.

The result loaded in seconds. A visually polished website with working navigation, pricing sections, FAQ — all functional, all professional. I’ve paid thousands of pounds for worse work on freelance platforms. The kind of output that local businesses currently purchase at premium rates.

I ran a parallel test requesting an “Ocean” themed website. Again, Minimax M3 produced something I found genuinely impressive — arguably more visually refined than my own site. The buttons worked. The redirects functioned. The pricing and contact sections rendered correctly.

Test 2: The Three-Way Game Build-Off

This is where I pushed all three models equally hard. Same prompt, same complexity, simultaneous execution.

The prompt: Create a playable HTML game similar to Mario. Catch falling gold coins, avoid bombs. Single HTML file. Start screen, scoreboard, three lives, modern design with animations, sound effects. When finished, create the file and open it so I can play immediately.

I ran Minimax M3 on the left, Claude Opus 4.8 in the center, and Codex 5.5 on the right. All received identical instructions. I pressed start simultaneously.

Claude finished first. The game loaded in browser. Space to start. The physics felt deliberately challenging — I found the platforming genuinely difficult, perhaps overly so. The controls responded, the character moved, but completing the level appeared nearly impossible in my attempts. Solid execution from a single prompt, but punishing difficulty.

Codex completed second. Simpler controls, more basic visual presentation. Functional but less ambitious. I could play it, the mechanics worked, but it felt like a minimum viable product compared to what I’d hoped for.

Minimax M3’s result surprised me most. Visually, it was clearly superior — the graphics, the polish, the overall presentation stood apart. It included lava hazards alongside the requested bombs and coins. The scoreboard functioned. Three lives displayed. Level progression worked. The spacebar controlled jumping. When I hit a bomb, I lost a life — the system logic operated correctly.

Playing it revealed intentional difficulty. The obstacles were dense, the timing tight. But the completeness of the package — sound effects, animations, visual design, functional game loop — from a single sentence prompt, without me writing a line of code?

I concluded that Minimax M3 produced the strongest visual result by a clear margin. The other two models created playable games, but Minimax M3’s output looked like something you’d consider for actual App Store release — with integrated Google Ads, potential for download revenue, the complete commercial pipeline.

What This Means for Online Business

I’ve been automating e-commerce operations for years. This feels different.

The agent framework means I can describe a person’s entire job — research trending products, compile lists, email reports, update inventory — schedule it, and the system executes daily without me. The human becomes the prompt writer and supervisor, not the doer.

I tested this specifically: I had the agent research currently trending products by country for eBay, generate the list, and email it to me. The entire workflow that previously required paid software subscriptions completed in approximately one minute, autonomously.

For my own trading operations, I can now deploy agents that work continuously, use the cheapest available tokens efficiently, and scale without linear labor costs. The Hermes Agent integration I configured optimizes this further — routing tasks to the most cost-effective model while maintaining quality.

Current Availability and Pricing

At launch, Minimax M3 is running a 50% promotional discount for 7 days. The per-million-token pricing sits in the cents range — I verified this against my own usage. The desktop agent has a free starting tier for testing.

The model weights are openly available. Someone without significant budget can run this on their own hardware. The barrier to entry for sophisticated AI automation has effectively disappeared.

FAQ

What exactly is Minimax M3?

Minimax M3 is an open-source large language model developed in China, released in June 2025. It features a 1 million token context window, multimodal capabilities (text, image, video, screen understanding), and a unique desktop agent application that can autonomously control a computer to perform tasks.

How does Minimax M3 compare to GPT-5.5 and Claude for coding?

On software testing benchmarks, Minimax M3 scored 79%, surpassing both GPT-5.5 and Claude 3.1 Pro. In my direct testing with identical game-building prompts, Minimax M3 produced more visually complete and polished results, though Claude finished faster and Codex produced simpler but functional output.

Is Minimax M3 actually free to use?

The model weights are open-source and free to self-host. The API pricing is significantly cheaper than competitors — per-million-token costs are in cents rather than dollars. There’s a free tier for testing the desktop agent, with paid tiers for higher usage. A 50% launch discount was active at time of testing.

Can Minimax M3 really replace human workers?

It can automate many repetitive computer-based tasks — research, data compilation, website building, code generation, report creation. However, it requires precise prompt engineering, clear boundary-setting, and human supervision. I view it as amplifying individual productivity rather than full replacement, though the economics of certain roles will shift dramatically.

Final Thoughts

I went into this test skeptical of the hype around another “GPT killer.” I finished it convinced that Minimax M3 represents something specific and valuable: the convergence of capable coding, genuine multimodal understanding, and autonomous action at a price point that removes financial barriers.

The desktop agent isn’t a demo. It’s a working tool I now have running scheduled tasks for my business. The website and game builds weren’t cherry-picked — they were first attempts with single prompts.

For entrepreneurs building online businesses, especially in e-commerce and automation, my recommendation is straightforward: test the free tier. Configure one agent with one specific task. Watch it work while you’re doing something else.

The first task I delegated was product research. I’m now deciding what to automate next.

Watch the full video (in Turkish — English subtitles available):

Tools & Community

• TurkoLister — the AI listing tool I use to turn Amazon products into optimized eBay UK listings in about 60 seconds (from £4.99/month, £1 one-week trial).
• AI & E-commerce Community — my Turkish-speaking community ($19/month) with weekly live sessions.
• Subscribe on YouTube — new experiments every week.

The post I Tested Minimax M3: The AI That Took Over My Computer appeared first on Digital Market Mentoring.

I spent months running my own AI automation systems, so I had to test this myself. I installed it, compared it directly against my existing setup, and the results genuinely surprised me. Here’s everything I found.

Key Takeaways

• PewDiePie’s Othaisus reached 44,000+ GitHub stars in 4 days and climbed past 67,000 shortly after release
• The software runs completely locally on your machine with no telemetry, no tracking, and no subscription fees
• Installation requires Docker but no coding knowledge; I walked through the full setup myself
• Features include chat, autonomous agents, file management, web search, coding, memory, deep research, email, and notes
• The “Cookbook” feature automatically scans your hardware and recommends from 270+ open-source models
• For true privacy, you need to run models locally via Ollama rather than using cloud APIs like OpenRouter
• Compared to my existing Hermes-style agents, Othaisus acts like a desktop assistant while my setup works autonomously in the background

Why PewDiePie Built This: The Privacy Rebellion

PewDiePie’s motivation wasn’t subtle. He explicitly stated: “What big companies do, I now do myself.” He had spent a year systematically removing Google from his life, built his first computer from scratch, learned to code from zero, and documented everything publicly.

The breaking point? Understanding where your data actually goes. When you use ChatGPT, Claude, or any cloud AI, your most private ideas, business plans, and customer data sit on someone else’s servers. You’re paying monthly for the privilege of being monitored. PewDiePie called this out directly as a rebellion.

By February, he had already trained a small model that beat GPT-4.0 on coding tests. Most people dismissed it as a joke. They weren’t laughing when Othaisus dropped.

What’s Actually Inside Othaisus

I pulled up his actual GitHub profile to verify this wasn’t exaggerated. The E-Repo was real, the star count was over 64,000, and the project had been live for less than two weeks. The description line says it plainly: “Chat, GPT and Claude on your own computer.”

The feature set is substantial:

• Chat interface — Similar to ChatGPT, but running on your hardware
• Autonomous agent — You give it tasks, it handles them from start to finish
• File, web, code, and memory management
• Deep research tools
• Email and notes integration
• “Cookbook” hardware scanner — Automatically recommends suitable models from 270+ open-source options based on your specific computer specs

There’s no sales team, no demo request forms, no trust issues to navigate. MIT licensed — use it, modify it, it’s entirely yours.

My Step-by-Step Installation Experience

I went through the full setup myself to see how accessible this actually is for non-technical users. The requirements are minimal: you need Docker installed, and that’s it. No coding knowledge required.

Step one: Download Docker for your operating system. I used the direct download links and ran through the standard installation — essentially clicking “next” through the prompts. One practical warning: Docker is a large file. Make sure you have sufficient disk space, or your system will slow down significantly.

Step two: Log into Docker (create an account if needed) using GitHub or Google credentials.

Step three: Run the specific command from PewDiePie’s repository in your terminal to pull and install the system. I tested this both manually and by feeding the instructions to AI assistants — both approaches worked, though doing it manually gave me faster results.

The critical sequence: install Docker first, then run the repository command. I deliberately tested reversing this to see the error message — the system correctly flags that Docker isn’t found and refuses to proceed. The error handling is clear enough that you’ll know exactly what went wrong.

Once running, you create a local username and password, and you’re in. Everything operates through your browser, entirely offline.

The Critical Privacy Choice: Local vs. Cloud Models

Here’s where I need to be completely honest with you. Othaisus itself is open-source and local, but the AI models you connect to it matter enormously for privacy.

You have two paths:

Path one (true privacy): Install Ollama on your computer and download open-source models directly. Everything stays on your machine. No data leaves. Completely free. This is the approach PewDiePie designed for.

Path two (convenience trade-off): Use cloud APIs through services like OpenRouter. You’ll get access to more powerful models, faster responses, and cheaper pricing than mainstream alternatives. But your data still flows to external servers. The “magic” of full local control disappears.

I tested both. For my installation, I added Minimax through an API connection to verify functionality. It worked immediately. But for readers following along, I strongly recommend the Ollama route if privacy is your primary motivation. The interface makes this straightforward — click the plus button, select “local,” choose Ollama, and the system downloads models from the internet directly to your machine. You can test and activate each model with a single click.

How It Compares to My Existing AI Setup

This was the test I was most curious about. I’ve been running automation systems for months using Hermes-style agents. After running both side by side, the distinction became clear:

Othaisus is like an assistant sitting at your desk with you. The screen is open, you’re conversing, working together — everything hidden on your machine, completely free.

My Hermes-style agents are different. They work in the background autonomously. They operate while you sleep, then message you: “Boss, I handled these tasks.”

My conclusion? I want both. Both are free. Both are mine. The fundamental shift PewDiePie is highlighting isn’t just about tools — it’s about ownership. You’re no longer renting AI; you’re owning it.

For a creator with 110 million subscribers to deliver that message carries weight. This isn’t niche tech philosophy anymore.

Practical Features I Tested

Once inside the interface, I explored the full capability set:

• New conversation panel for standard chat interactions
• Product research section for e-commerce and market analysis
• Email integration for automated correspondence
• “Brain” and “skills” modules you can add and customize
• Model selection interface showing which languages perform best for specific tasks — some code better, others write better
• Deep research section for comprehensive investigations
• Gallery for uploading and analyzing your own images
• Library where you train the AI on your preferences and background for more personalized responses

The scheduling capability particularly impressed me. You can set daily automations entirely free on your local machine — “At 10 AM, email this many people” or “Research trends and bestselling products for me.” The system runs in the background, locally, with zero cost.

Theme customization is also available if you want to adjust the visual interface to your preference.

FAQ

Do I need to know how to code to install Othaisus?

No. The installation requires Docker, which you download and install through standard prompts. After that, you copy and paste one command from the repository. I verified this process myself and had the system running without writing any original code.

Is my data completely private when using Othaisus?

Only if you use local models through Ollama. The Othaisus software itself runs locally and collects no telemetry, but if you connect cloud APIs like OpenRouter, your data still travels to external servers. For full privacy, download open-source models directly to your machine.

How does this compare to paid AI services like ChatGPT Plus?

Othaisus lacks the raw power of GPT-4o or Claude 3.5 Opus for complex reasoning tasks. However, it costs nothing, runs without internet after setup, keeps all data local, and includes autonomous agent capabilities that many paid services don’t offer. For routine business automation, research, and content tasks, the open-source models I tested performed adequately.

What computer hardware do I need to run this effectively?

The “Cookbook” feature automatically evaluates your specific hardware and recommends compatible models from over 270 options. Simpler models run on modest systems; larger models require more RAM and preferably a dedicated GPU. Docker itself requires significant disk space — ensure you have storage available before installing.

Final Thoughts

PewDiePie’s Othaisus isn’t perfect, and it won’t replace every cloud AI use case. The models are smaller, setup requires some technical comfort with Docker, and you’ll need to manage your own model selection.

But the core proposition is undeniable: a functional, private, autonomous AI system that costs nothing and answers to no corporation. I watched it climb from 44,000 to 67,000 stars while testing it. That velocity speaks to genuine demand, not influencer hype.

For my own workflow, I’m now running both approaches — Othaisus for interactive desktop work where privacy matters, and my background Hermes agents for autonomous overnight processing. The combination costs me nothing and keeps everything under my control.

If you’ve been paying monthly subscriptions while wondering where your business data actually lives, this is worth your time to explore. The tools to own your AI stack are finally accessible. The only question is whether you’ll use them.

Watch the full video (in Turkish — English subtitles available):

Tools & Community

• TurkoLister — the AI listing tool I use to turn Amazon products into optimized eBay UK listings in about 60 seconds (from £4.99/month, £1 one-week trial).
• AI & E-commerce Community — my Turkish-speaking community ($19/month) with weekly live sessions.
• Subscribe on YouTube — new experiments every week.

The post I Tested PewDiePie’s Free AI: Here’s What Surprised Me appeared first on Digital Market Mentoring.

Key Takeaways

• Nexand 2 is genuinely free—input and output tokens cost $0, even for commercial use, yet it outperformed paid models on the game-building task.
• Claude Opus costs ~$25/million output tokens versus Kimi K2.7 at ~$4, GLM 5.1 at ~$3.08, and Nexand 2 at $0.
• Claude finished fastest on most tasks, but speed didn’t always equal the best final output.
• Kimi K2.7 failed completely on the game task—writing files but producing no playable output.
• GLM 5.1 took 25 minutes to build a game with movement bugs, making it impractical for rapid iteration.
• Operator’s Fusion mode can auto-switch between models mid-task, potentially optimizing for cost and quality dynamically.

The Cost Reality: Why This Comparison Matters

Before testing quality, I stared at the pricing table for a long time. Claude Opus charges roughly $5 per million input tokens and $25 per million output tokens. Kimi K2.7 drops that to about $0.75 input and $3.50 output. GLM 5.1 sits at approximately $0.98 input and $3.08 output. Then there’s Nexand 2: $0 for both input and output, with full commercial freedom.

I’ve been running AI-powered automation systems for my e-commerce operations for years. When you’re processing thousands of tokens daily, the difference between $25 and $0 isn’t theoretical—it’s the difference between a profitable workflow and one that bleeds money. But I refused to believe free could compete. That’s why I designed these tests.

Test 1: Building a Sales Funnel Website

I gave all four models the same prompt: build a website similar to my okenis.com system, a functional sales funnel with specific features. No hand-holding, no detailed specifications—just “make something like this.”

Kimi K2.7’s Attempt

Kimi produced a complete website quickly. It included cookie consent banners—something many developers forget—and structured the layout professionally. However, looking at the backend code, I spotted bugs. Text rendering had inconsistencies. The design felt obviously AI-generated, lacking the subtle refinements a human designer would add. Usable? Yes. Polished? Not quite.

Claude Opus’s Attempt

At nearly $25 per million tokens, Claude delivered the most visually refined result. The background graphics were superior, with animated highlights and a four-step automation showcase. Crucially, Claude didn’t invent random pricing—it pulled realistic figures, showing it understood context rather than hallucinating. The interactive elements worked smoothly. Between Kimi and Claude, Claude clearly produced the more professional output.

Nexand 2’s Attempt

Here’s where I got genuinely confused. The completely free model added background images to my simple prompt, created illuminated panels, and styled text in an “ultra-futuristic” aesthetic. It structured pricing packages and even built a forum section. Was it perfect? No—it worked from my prompt rather than researching live data from my site. But for $0? I kept asking myself: how is this possible?

GLM 5.1’s Attempt

GLM produced elegant opening animations and even added animated counters—a nice touch. However, it misspelled my brand name, adding incorrect characters. The backend implementation was decent but basic. With more prompt refinement, GLM could improve, but out of the box, it felt rougher than the others.

Test 2: Building a Minecraft-Style Game

Websites are one thing. Games test reasoning, physics, interactivity, and error handling. I gave all models the same prompt: build a Minecraft-like block game in a browser. Here’s where everything got strange.

Claude finished first. The game loaded, allowed left-click to break blocks, right-click to place them, and jump with upward movement. Was it perfect Minecraft? Honestly, I’ve never played Minecraft—my son tells me about it—but the mechanics seemed functional. I couldn’t definitively call it good or bad without player expertise.

Kimi K2.7 failed entirely. It wrote some code files but produced no playable game. The output files were essentially empty. I have no idea why—it simply didn’t execute the task.

Fusion mode disappointed. I expected great things since Fusion automatically switches to the best model for each subtask. Instead, the output was extremely blurry, almost unplayable. The colors were there, interaction was technically possible, but the visual quality made it practically useless.

Nexand 2 won decisively. After approximately 13 minutes, the free model delivered a functional block-building game. I could place boxes, break them, and navigate the environment. The visual clarity surpassed Fusion’s blurry mess. I sat there genuinely baffled: the $0 model outperformed the $25 one on a complex creative task.

For comparison, I also tested Minimax (which had previously built my best game). It produced something poetic and visually striking—almost too bright—but functional. When you fell off the world, the game ended. Different aesthetic, competent execution.

GLM 5.1 took 25 minutes and delivered a visually impressive but broken game. The character wouldn’t move properly. After that wait, the bugs made it unusable.

What About Speed and Smart Adaptation?

Nexand 2 has an intelligent side I appreciated: it adapts depth based on task complexity. Simple queries get fast answers; complex problems trigger deeper reasoning. Kimi K2.7 uses approximately 30% fewer reasoning steps than its predecessor, making it noticeably quicker. Claude remains methodical, step-by-step, never rushing.

But here’s what changed my thinking: Operator’s Fusion mode. This feature lets the system automatically switch between models during a task. If one model handles a subtask faster or better, Fusion routes to it, then returns to the primary model. For Cursor users, this infrastructure already exists. The implication is huge—you no longer need to manually choose “which AI is best.” The system optimizes in real-time, potentially combining Claude’s refinement with Nexand’s cost advantage.

My Honest Verdict: Which Model for Which Situation

After these tests, my recommendations split by use case:

For zero-budget, open-source coding: Nexand 2 is now my top recommendation. The game test proved it handles complex logic. One caveat: because it’s powerful, it demands strong hardware if you’re running it locally. Use Operator’s free cloud tier instead if your machine struggles.

For cheap, fast, open-source work: Kimi K2.7 at ~$4 per million tokens offers solid value. It failed the game test but performed adequately on websites. For routine coding tasks where speed matters, it’s worth considering.

For mission-critical, high-stakes code: Claude Opus remains the premium choice. When client work, security, or complex architecture demands the highest reliability, I still reach for Claude. But I pay accordingly—and increasingly question whether that premium is always justified.

For GLM 5.1: I can’t currently recommend it. Twenty-five minutes for a broken game, slower than competitors at similar pricing tiers. Perhaps future updates will improve it.

FAQ

Is Nexand 2 really completely free for commercial use?

Yes. Based on my testing through Operator, both input and output tokens are currently priced at $0, with explicit commercial usage rights. You can also download the open-source weights and run them locally.

How did Nexand 2 score on official coding benchmarks?

According to LiveBench data, Nexand 2 scored above 80 on coding tests, placing it in the same tier as GPT and Claude on that specific benchmark.

What hardware do I need to run Nexand 2 locally?

Nexand 2 is a powerful model requiring substantial GPU resources for local deployment. If you lack high-end hardware, using it through cloud platforms like Operator’s free tier is the practical alternative.

What is Fusion mode and should I use it?

Fusion mode automatically switches between AI models during a task, routing subtasks to whichever model handles them best or cheapest. In my game test, it underperformed individual models, but the concept is promising as the technology matures.

Conclusion

This test destroyed my assumption that price equals quality. Nexand 2, at literally zero cost, built the best game and a competent website. Claude still leads on polish and reliability, but the gap is narrowing faster than I expected. For my own operations, I’m now running hybrid setups—Claude for client-critical work, Nexand 2 for internal prototyping and experimentation. The real winner might be Fusion-style systems that combine them intelligently. If you’re still paying premium prices without testing alternatives, you’re potentially burning money. I run these comparisons weekly, and the landscape shifts constantly. The model that’s best today may be second-tier next month. Stay testing, stay skeptical, and never assume the expensive option is the right one.

Watch the full video (in Turkish — English subtitles available):

Tools & Community

• TurkoLister — the AI listing tool I use to turn Amazon products into optimized eBay UK listings in about 60 seconds (from £4.99/month, £1 one-week trial).
• AI & E-commerce Community — my Turkish-speaking community ($19/month) with weekly live sessions.
• Subscribe on YouTube — new experiments every week.

The post I Tested 4 AI Models: Free Nexand 2 Beat $25 Claude appeared first on Digital Market Mentoring.