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.

The model behind this is Alibaba’s Qwen 3.7 Plus—a multimodal agent that doesn’t just talk, but actually does the work. It sees your screen, writes code, tests what it built, and fixes its own mistakes. In this article, I’ll break down exactly what I discovered testing this system, show you the real results, and explain how you can try it yourself for free.

Key Takeaways

• Alibaba’s Qwen 3.7 Plus is a multimodal agent model that sees your screen, uses your mouse and keyboard, writes code, and self-corrects—unlike standard chatbots that only give you text to copy-paste
• In an official demo, the agent worked 11+ hours continuously, wrote 10,000+ lines of code, made 1,000+ agent calls, and completed an entire software development lifecycle from documentation to deployment
• It can replicate applications from just a screenshot—no existing code required—and connect to live data sources like real stock market APIs
• Qwen 3.7 Plus costs approximately 6x less than Qwen 3.7 Max and offers 1 million tokens of context memory for handling complex, long-running tasks
• Unlike previous Alibaba models, this one is not open-source—you can only access it via API or cloud platforms, not download and run locally
• Alibaba is currently offering up to $5,000 in free credits (with $200 immediately available) for new cloud users with no minimum spending requirement

What Makes This Different From ChatGPT or Claude

Until now, AI tools have followed a familiar pattern: you ask, they answer. They might write you some code, explain a concept, or generate an image. But the actual doing? That was always on you. You copied, pasted, applied, troubleshooted.

Qwen 3.7 Plus represents a fundamentally different category. Alibaba calls it a “multimodal agent model”—and that technical label actually matters. This isn’t a chatbot. It’s a system with what I can only describe as digital eyes, hands, and a brain that loops continuously.

Here’s how Alibaba themselves describe the cycle: see, think, write, apply, and verify—all unified inside a single agent. The model looks at your screen, understands which buttons to press, writes the code, executes it, tests the results, and when it finds errors, debugs itself automatically. That closed loop is exactly what separates it from every conversational AI I’ve used before.

On the Terminal Bench benchmark, Qwen 3.7 Plus scores 70.3—competitive with major cloud providers, except those competitors charge you serious money while this system remains dramatically more accessible.

The 11-Hour Demo That Changed My Perspective

The most striking demonstration Alibaba provided—and the one that made me sit up and pay attention—involved what they call their “Hybrid Agent” system built on Qwen 3.7. They gave it a single instruction: build an English vocabulary learning application. Then they walked away.

What happened next is worth examining in detail. The agent worked for over 11 hours continuously and stably. It wrote more than 10,000 lines of code. It made over 1,000 agent calls. And it completed the entire software development lifecycle from start to finish.

But “entire lifecycle” isn’t marketing fluff here. I’m talking about:

• Writing its own design documents and planning what the application should include
• Writing the actual code and building the application
• Creating test scenarios
• Running those tests on-screen by itself
• Testing multiple scenarios simultaneously
• Updating documentation
• Releasing new versions of the application autonomously

Stop and consider what this represents. This is days of work for a small software team, completed in a single night while you sleep. This is precisely what I mean when I talk about “systems that work while you sleep.”

Rebuilding a Finance App From Just a Screenshot

The second demo I found even more impressive because there was no existing code at all—only a screenshot of an interface.

Here’s what the agent did step by step:

1. Entered its own application ecosystem and examined the interface
2. Understood which features were located where
3. Based on what it saw, wrote the application from scratch using Switch UI code
4. Connected to a real stock exchange API (Long Bridge) to pull live market data
5. Compiled and ran the application
6. Tested itself across 10 different functions autonomously

It verified live prices were actually updating, stocks were changing, search was functioning—all tests passed. The agent had reproduced a working financial application with live data, dark theme, and proper layout, from nothing but a screenshot and a command. No designer. No developer. Just an image and instructions.

There’s even a third demo where the model browsed art marketplace websites, clicked through options, and made purchases—going as far as buying items online on your behalf.

The Price Advantage Nobody’s Talking About

Here’s what caught my attention most: the cost. You’d assume this level of capability comes at a premium. The opposite is true.

Qwen 3.7 Plus costs approximately 6x less than Alibaba’s own Qwen 3.7 Max text model. On top of that, it carries 1 million tokens of context memory—massive capacity that lets it track complex, lengthy tasks without losing thread.

Its screen-reading capability—understanding which buttons to press—ranks leader position in industry tests. For context, Alibaba’s cloud platform currently offers up to $5,000 in credits for new users, with $200 available immediately with no minimum spending requirement. Though I should note: identity verification is required to complete setup, which carries its own considerations I leave entirely to your judgment.

How I’m Actually Using This (And How You Can Too)

I’ve been testing this through two main paths. The simplest entry point is chat.qwen.ai—create an account, select your preferred mode (thinking, fast response, or autonomous decision-making), and you can generate images, video, documents, software, or podcasts from the creative panel. The coder section lets you write code and, after connecting GitHub, automatically push to Git repositories.

For the full agent power I described, you’ll need Alibaba Cloud Model Studio, where you can call Qwen 3.7 Plus via API. This is more technical, but it unlocks the complete autonomous capability.

One important limitation: unlike previous Alibaba releases, this model is not open-source. You cannot download and run it on your own machine—only access it through internet-based APIs. This was a controversial decision given Alibaba’s history of open releases.

The model also integrates into existing tools you might already use—Cloud Code, Open Code, Qwen Code—so if you’re already running agent workflows, you can switch the underlying engine to Qwen.

What This Actually Means for Your Business

I keep saying this on my channel, and this model makes the line clearer than ever: AI won’t take your job. People who use AI will take your job. The difference is subtle but enormous.

Consider your situation: small business, app idea, no budget for developers, no coding knowledge. Previously, that idea stayed an idea. Now, you can describe it, and an agent can build it while you handle other priorities.

Here’s a concrete example from my own work. I set up a system to find major affiliate companies online, write blog posts for them, create social media content, and publish it. Over the last 12 months, that’s generated over $2,000 in passive income; lifetime, over $73,000. The concept—combining your ideas with AI execution at speed—is what matters. The specific implementation varies by person and knowledge.

If you run e-commerce, agents can research products, set up stores, track prices. If you create content, they can handle repetitive tasks. The key is your direction and judgment—the AI handles execution.

FAQ

Is Qwen 3.7 Plus really free to use?

Alibaba offers up to $5,000 in cloud credits for new users, with $200 available immediately without minimum spending requirements. The chat interface at chat.qwen.ai has its own free tier limits. For sustained API usage, you’ll eventually pay, but at roughly 6x lower cost than comparable models.

Can I run Qwen 3.7 Plus on my own computer?

No. Unlike previous Qwen releases, this model is not open-source. Alibaba made it available only through API and cloud platforms. You cannot download weights or run it locally.

How does this compare to Claude or GPT-4 for coding?

The critical difference is autonomy. Claude and GPT-4 generate code you must then implement, test, and debug yourself. Qwen 3.7 Plus executes the full loop: it sees your screen, writes code, runs it, detects errors, and fixes them—potentially for hours without human intervention. Benchmark scores are competitive, but the workflow difference is the real distinction.

Is this reliable enough for production use right now?

Honestly, probably not consistently. Alibaba’s demos ran in controlled environments. Real-world complexity—unexpected UI changes, API failures, edge cases—will challenge any agent. I view this as extraordinarily powerful for prototyping, learning, and accelerating development, but not yet a complete replacement for human oversight in critical systems.

Conclusion

Alibaba has released something genuinely significant. A model that sees your screen, writes code, tests itself, and works for 11 hours straight to complete applications—at a fraction of competitor pricing. The trajectory is unmistakable: AI is transitioning from talking to doing.

This isn’t magic. Demos are controlled, real life brings friction, and you’ll still need your own knowledge to direct these tools effectively. But the capability gap between idea and execution is narrowing fast. The question isn’t whether this technology will impact your work—it’s whether you’ll be among those learning to direct it.

What project or idea would you build with an agent that can work through the night while you sleep? That’s the question worth sitting with.

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 Let AI Code for 11 Hours Straight—Here’s What Happened appeared first on Digital Market Mentoring.

So what are we supposed to use right now? I got curious and decided to find out. I headed to OpenRouter and tested 11 of the world’s leading large language models with the exact same prompt. The same game, the same website request, the same everything. What happened next genuinely surprised me—and the best performer was completely free.

Key Takeaways

• Claude Opus 4.5 remains unavailable with no official reopening date; Anthropic is negotiating with US authorities
• GPT-5.6 has not launched despite widespread speculation; GPT-5.5 is still OpenAI’s current flagship
• I tested 11 models on OpenRouter: Gemini Pro, OpenAI (GPT-5.5), Claude Opus 4.8, Kimi 2.7, Grok 4.3, Ollama, DeepSeek V4, GLM 5.1, Qwen 3.7, MiniMax M3, and the free open-source Nexa N2
• MiniMax M3 produced the best playable game from a single prompt—visually impressive and fully functional
• Nexa N2, a free open-source model, built a professional website in seconds that outperformed paid alternatives
• Grok and Kimi failed outright—Grok’s game was unplayable, Kimi crashed repeatedly and required multiple reprompts
• Completion speed varied dramatically: Grok and Gemini finished in ~5 minutes, MiniMax took 11 minutes, others fell in between

Why the Top Models Are Missing Right Now

Before diving into my test results, let me clarify why I even had to do this comparison. Claude Opus 4.5 was released and then immediately pulled—I’ve confirmed in my own system that it’s still inactive and unavailable for selection. Anthropic says they’re working to bring it back “as soon as possible,” but there are only three realistic paths forward: the US government reverses its decision, Anthropic launches a geographically restricted version for US users only, or a lengthy legal battle unfolds. My honest assessment? It’s going to stay unavailable for a while.

As for GPT-5.6, the rumored specs sound impressive—over 1 million token context memory, significantly improved autonomous agent capabilities, and 2-5x faster code execution. But I’ve seen the “it’s out now” videos too, and they’re wrong. The official top model from OpenAI is still GPT-5.5. When GPT-5.6 actually drops, I’ll notify my community immediately.

How I Set Up the 11-Model Showdown

I wanted a test that reflected real-world use: vibe coding—building something functional from a single, natural-language prompt without extensive back-and-forth refinement. I chose a classic space shooter game similar to ones I played as a kid at Atelier arcades. Simple concept, but it tests visual generation, game logic, controls, and overall polish.

I sent the identical prompt to all 11 models through OpenRouter and let them work independently. I also asked each model to build a website for “Okyanus.com” to test web development capabilities. No hand-holding, no follow-up fixes—just one shot to see what each could deliver.

The lineup included five American companies (OpenAI, Anthropic, Google, Grok, Ollama) and six Chinese companies (Kimi, DeepSeek, GLM, Qwen, MiniMax, and Nexa). Yes, I burned through credits running this. Several models hit rate limits mid-test and I had to purchase additional credits to complete the comparison.

The Game Results: From Broken to Brilliant

The Fast Finishers: Gemini and Grok

Gemini Pro and Grok 4.3 completed their games in roughly 5 minutes—the fastest of the batch. Gemini’s result was functional but basic: simple interface, working sound, movement controls operational. Playable? Yes. Impressive? Not particularly.

Grok, on the other hand, was a disaster. The game killed the player before it even properly loaded. Enemies fired from every direction simultaneously, making forward progress impossible. I genuinely wanted to test it properly, but it was unplayable by any reasonable standard.

The Visual Standout That Couldn’t Run: GPT-5.5

ChatGPT’s system (GPT-5.5) produced what I consider the best visual design of any model. The game’s aesthetics were genuinely stunning. The problem? Severe performance issues—lag, freezing, unresponsive controls. Beautiful to look at, impossible to play. There’s a lesson here about the gap between appearance and functionality.

The Control Problem: Opus 4.8 and GLM 5.1

Claude Opus 4.8 created something visually superior to most competitors, but with a critical flaw: the left-right arrow keys didn’t work properly. Only forward and backward movement functioned. GLM 5.1 had a similar issue initially, though I discovered mouse controls worked for lateral movement—still not the intended keyboard experience.

The Reliable Middle: DeepSeek V4 and Qwen 3.7

DeepSeek V4 delivered what I’d call the best balance of playability and visuals among the mid-tier performers. Controls worked smoothly, graphics were decent, though enemy spawning seemed limited—I only encountered 4-5 enemies total. Qwen 3.7 was playable but visually primitive, with audio that felt more like noise than designed sound effects. Both worked on the first try, which matters more than you might think.

The Complete Failure: Kimi 2.7

Kimi was the most frustrating experience by far. The game didn’t open on the first attempt. I sent a second prompt—it still failed. On the third try, the game launched but crashed during play. This illustrates a critical real-world problem: models that require constant reprompting don’t just waste time, they often degrade the output with each iteration. Despite Kimi’s strong reputation, it was the only model that couldn’t produce a working result after multiple attempts.

The Clear Winner: MiniMax M3

At 11 minutes, MiniMax M3 was among the slower finishers. But what it delivered made the wait irrelevant. The best game of the entire test, by a significant margin. Smooth controls, polished visuals, proper game mechanics, genuinely playable. I’ve been following MiniMax for a while—the M3 series has become one of my favorite model families—and this performance validated that enthusiasm. From a single prompt, it built something I’d actually show to someone.

The Website Test: Where Free Beat Expensive

While the games were generating, I had each model build a website for “Okyanus.com.” Here the results flipped in an interesting direction.

Nexa N2—completely free, open-source, no subscription—built a professional, functional website in seconds. I’m talking genuinely impressive output that would pass for a paid service. Same prompt to GLM 5.1 produced solid results too, and Kimi actually created something more polished and detailed than its broken game, with custom graphics and professional layout.

But Nexa N2’s performance stuck with me. This is a model you can run without spending a penny, yet it outperformed systems people pay $25+ monthly for. When I say “open source, completely free” in the AI space, I know there’s skepticism. This test made me a believer.

What the Leaderboards Get Wrong

Before running my own tests, I checked the standard benchmarks. They show Claude Opus 4.5 leading in writing, coding, and most categories. Opus 4.7 leads in visual reasoning and document analysis. ChatGPT Image 2 wins at image generation. Gemini 1.0 Flash dominates video creation.

But here’s what benchmarks miss: the single-prompt experience. Most impressive AI demos you see online? They’re the result of extensive prompt engineering, iterative refinement, sometimes dozens of back-and-forth exchanges. My test reflected what happens when you don’t have time for that—when you need something that just works from your first request.

Also worth noting: the visual gap between what Claude Opus 4.5 produced in other users’ refined projects (I showed examples of arcade-perfect Space Invaders clones and Minecraft-style games) and what the available models achieved in my single-prompt test was substantial. When Opus 4.5 returns—if it returns with fewer restrictions—it will likely reset expectations. I’m genuinely curious whether it’ll come back limited or at full strength.

FAQ

Is Claude Opus 4.5 coming back?

There is no official reopening date. Anthropic has stated they are negotiating to restore access, but the most likely scenario is an extended period of unavailability. A geographically restricted US-only version or prolonged legal process are possible outcomes.

Has GPT-5.6 been released?

No. Despite rumors and misleading videos claiming otherwise, OpenAI’s current flagship model remains GPT-5.5. GPT-5.6 has not launched and no official release date has been announced.

What is the best free AI model for coding right now?

Based on my testing, Nexa N2 delivered exceptional results for web development completely free. For game development from single prompts, MiniMax M3 outperformed all competitors including paid options, though it requires API access.

Why do benchmark rankings differ from real-world results?

Benchmarks typically measure performance under optimized conditions with extensive prompting. Real-world use often involves single prompts without refinement. My test showed significant gaps between benchmark leaders and models that actually deliver functional outputs on the first try—completion speed, control reliability, and crash resistance matter in practice but rarely appear in benchmark scores.

My Honest Recommendation

If you’re building right now and need something that works without constant babysitting, MiniMax M3 for game development and Nexa N2 for web projects are my top picks based on this test. Neither requires the patience that Kimi demanded or accepts the broken output that Grok delivered.

The landscape shifts weekly. Models that failed today may improve tomorrow. But the core lesson from this experiment? Price and reputation don’t determine practical utility. The free, open-source option beat expensive competitors. The lesser-known Chinese model outcoded famous American brands. Test for your specific use case—assumptions about “the best” will cost you time and money.

What’s your go-to model for coding projects? I’m genuinely curious whether your experience matches my results.

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 11 Top LLMs: The Free Winner Shocked Me appeared first on Digital Market Mentoring.