Hugging Face shipped ml-intern on June 18, 2026, a free open-source ML engineer that can clean data, run training loops, evaluate models, and write deployment code. The agent runs on a 7 billion parameter instruction model with a 32,768 token context window. The weights are released under Apache 2.0. Hugging Face published the announcement on its homepage and the model weights are available through the Hub. The full bf16 checkpoint takes 14.2 GB of disk, while a 4-bit quantized version needs 4.7 GB. That size puts real ML engineering assistance on hardware that many developers already own.
Hugging Face built ml-intern with a small team and outside contributors, then opened the entire pipeline. The code and evaluation harness are linked from the Hugging Face homepage and GitHub release notes. Unlike closed coding agents, ml-intern is not a thin wrapper around a remote API. You can run it locally, fine tune the 7B base, and inspect every prompt. This matters because free tier limits from commercial tools have tightened in 2026, as covered in our free AI pricing changes tracker.
The release matters because it attacks the cost problem directly. Commercial ML copilots often charge per token, per agent step, or per seat, and those fees climbed in mid 2026. ml-intern under Apache 2.0 removes the monthly bill and lets an enterprise host the agent inside its own VPC. Early benchmark numbers show the agent at 78.4 percent on ML-Bench, 68.9 percent on HumanEval, and 41.3 percent on SWE-bench Verified. Those scores are not top of the closed-model leaderboard, but they are close enough for many everyday ML tasks. Read our best free AI models in 2026 for more background.
Why now is obvious if you track the AI coding tools pricing changes from June 2026. GitHub Copilot moved to usage based billing, OpenAI and Anthropic changed free tier resets, and teams got a rude awakening. An open-source agent resets the conversation. It can run offline, does not phone home, and costs nothing after the hardware you already have. That makes ml-intern a pressure release valve for students, researchers, and smaller shops that cannot absorb another per token invoice.
How Do the Top Options Compare?
| Option | Best For | License | Params | Context | Local Run |
|---|---|---|---|---|---|
| ml-intern 7B | Free local ML engineering | Apache 2.0 | 7B | 32k | Yes |
| ml-intern hosted Space | Zero setup testing | Apache 2.0 | 7B | 32k | No, runs on HF |
| Closed coding agent | Managed convenience | Proprietary | Varies | Varies | No |
| Manual OSS stack | Full control | Mixed OSS | Varies | Varies | Yes |
Scores and sizes reflect the June 2026 release as announced on the Hugging Face homepage. Closed agent pricing may change after publication.
1. Hugging Face ml-intern 7B , Best for free, local ML engineering
Hugging Face ml-intern 7B is the core model. It uses a dense 7 billion parameter transformer with a 32,768 token context window. Hugging Face released the weights under Apache 2.0, so commercial use and derivative models are permitted. The full checkpoint is 14.2 GB in bf16. A 4-bit GGUF quant is 4.7 GB. On an RTX 3060 12GB, the agent generates roughly 32 tokens per second with vLLM. On Apple Silicon with M3 Max, it runs via MLX at 18 tokens per second. The model card is on the Hugging Face Hub and points to the open training recipe. Hugging Face has not paywalled the better quant or held back the eval harness. The agent knows common ML libraries, including scikit-learn, PyTorch, Hugging Face Transformers, and XGBoost. It can read a messy CSV, propose a train test split, generate a baseline pipeline, and then write an evaluation script. On ML-Bench it scored 78.4 percent. On HumanEval it scored 68.9 percent. That puts it within a few points of some paid coding assistants on code generation tasks, but it is not a frontier reasoning model. The main gap appears in multi step debugging, where the 7B size shows.
Key strengths:
- ✅ Apache 2.0 license allows commercial use, fine tuning, and private forks
- ✅ 7B model runs in 4.7 GB quantized form on a single RTX 3060 or Apple Silicon
- ✅ 32k context handles long data schemas, multi file repos, and evaluation logs
- ✅ Benchmark scores track paid copilots on ML-Bench and HumanEval
- ✅ No API keys, rate limits, or usage based billing
- ❌ SWE-bench Verified score still trails frontier closed models by a wide margin
- ❌ Requires local hardware or a Hugging Face Space with cold start delays
- ❌ No vendor support line for break fix issues
Who it’s for: Developers who want an open agent they can inspect, self host, and modify without a monthly bill.
2. Hugging Face ml-intern hosted Space , Best for zero setup testing
The hosted option removes the setup barrier. Hugging Face offers a public Space that runs the 4-bit quant on a free CPU basic tier. Response times are slower, often 5 to 10 tokens per second, but the interface works for quick data tasks. The Space uses the same Apache 2.0 weights, so you can download the model and leave at any time. Hugging Face has not added a paywall to the base Space, but private Spaces require a paid plan. This matters because free tier policies have tightened across the industry, as tracked in our AI API free tier limits update. The hosted version is not ideal for production. Free Spaces sleep after inactivity and may lose session state. For real projects, you should download the GGUF and run it locally or use a dedicated inference endpoint. The benefit is a zero install test bed. You can paste a CSV, ask for a baseline model, and see the agent’s output in a browser. That low friction mirrors what closed tools offer without the immediate credit burn. Watch our AI free tier landscape shifts page for changes to hosting allowances.
Key strengths:
- ✅ No local GPU required
- ✅ Free tier Space runs the 4-bit quant
- ✅ Shares the same weights and prompt as local version
- ✅ Community Spaces show example deployments
- ❌ Cold starts and short session limits on free hardware
- ❌ Data leaves your machine unless you use a private Space
- ❌ CPU only free tier is slower than a local GPU
Who it’s for: Users who want to try ml-intern without installing Python or buying a GPU.
3. Closed-source ML coding agents , Best for managed convenience and frontier scores
Closed coding agents from major vendors still lead on hard multi file refactoring. In June 2026, several providers shifted to usage based billing, and developers pushed back. The developer outcry over GitHub Copilot hidden costs shows why trust is fragile. These tools are convenient, but the meter runs constantly. For a small team doing ML work, a single agentic debugging session can cost several dollars to over a hundred dollars. The bill is hard to predict before the sprint ends. Closed tools also introduce data governance issues. Your training code, data schemas, and evaluation logs move through a third party. Contracts may allow model trainers to retain prompts. For regulated shops, that is a nonstarter. Open source ml-intern can run offline, which closes that gap. The tradeoff is clear: closed agents are smarter but rent seeking. Open agents are inspectable but less polished.
Key strengths:
- ✅ Frontier reasoning models score higher on SWE-bench and debugging
- ✅ Managed cloud removes local hardware constraints
- ✅ Tighter integration with existing IDEs and enterprise identity
- ❌ Usage based billing creates unpredictable monthly costs
- ❌ Proprietary code and weights cannot be inspected or self hosted
- ❌ Free tier limits have gotten tighter through June 2026
Who it’s for: Teams that need the highest benchmark scores and can pay per token or seat.
4. Manual open-source ML stack , Best for full control and reproducibility
The manual stack is the old way. You combine pandas, scikit-learn, PyTorch, and Weights and Biases or MLflow. There is no agent to generate the first pass. That gives you total control, but it takes longer. ml-intern sits in the middle. It automates the boring parts while leaving the code visible and modifiable. The manual route still wins for unusual data or custom loss functions that a 7B model handles poorly. Cost is not the only reason to stay manual. Agents can produce plausible but wrong evaluation code. A manual script forces you to read every line. ml-intern reduces some of that risk because you can inspect the prompt and the output locally. But it does not remove the need for code review. The honest downside is that an open 7B agent will miss edge cases a senior engineer would catch. That makes it a junior assistant, not a replacement. The name ml-intern is accurate.
Key strengths:
- ✅ Complete control over every library, version, and training script
- ✅ No agent layer to hide decisions
- ✅ Free tools remain widely available for core ML work
- ❌ Slower to set up than a purpose built agent
- ❌ No unified interface for data cleaning, training, and deployment
- ❌ Debugging long pipelines still falls on the human engineer
Who it’s for: Engineers who prefer direct code and versioned scripts over an agent.
Frequently Asked Questions
What is Hugging Face ml-intern?
Hugging Face ml-intern is a free, open-source autonomous ML engineer released June 18, 2026. It uses a 7 billion parameter model with a 32,768 token context window under Apache 2.0. It handles data preparation, training loops, evaluation, and deployment code.
Is ml-intern really free for commercial use?
Yes. Apache 2.0 allows commercial use, modification, and private distribution. You can fine tune the weights and deploy them without paying Hugging Face. You still pay for any cloud hardware or private Spaces you use.
What hardware do I need to run it locally?
The 4-bit quantized version needs about 4.7 GB of disk and runs on a 12 GB GPU. CPU inference works but is slow. Apple Silicon with MLX also works for smaller tasks.
How does it compare to closed coding agents?
It scores 78.4 percent on ML-Bench and 68.9 percent on HumanEval, close to some paid assistants. It trails frontier closed models on SWE-bench Verified at 41.3 percent. It wins on cost, privacy, and custom fine tuning.
Can I fine tune ml-intern on my own data?
Yes. The model weights and training recipe are open. You can use LoRA or full fine tuning on domain specific data. Quantized tools like PEFT and TRL support the architecture.
Where can I get it?
The model weights are available from the Hugging Face Hub, and the code is linked from the Hugging Face homepage and GitHub. Start with the official Hugging Face Spaces demo for a zero install test.
What Should You Remember?
- Apache 2.0: Commercial use, private forks, and fine tuning are allowed without a vendor contract.
- 7B dense: The agent fits on a single 12 GB consumer GPU, no data center required.
- 32k context: Long data schemas and multi file repos fit in one prompt.
- 78.4 ML-Bench: The model matches paid copilots on common ML tasks but trails frontier tools on hard debugging.
- Local by default: No API keys, no rate limits, no per token meter.
- Hosted Space caveat: Free Spaces sleep and are slower, so production users should self host.
- Cost reset: This release lands as commercial AI pricing changes squeeze developers, giving teams a no invoice escape hatch.
Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.
