{"slug": "thx-01-just-processed-70-million-tokens-our-open-source-decision-model-is-just", "title": "THX-01 Just Processed 70 Million Tokens. Our Open-Source Decision Model Is Just Getting Started.", "summary": "HAL-X AI has open-sourced THX-01, a 322-million-parameter non-autoregressive decision model that returns structured answers with calibrated probabilities in a single forward pass, and reports over 70 million tokens processed through its API since launch. On an internal benchmark of 2,843 multilingual support tickets, THX-01 scored 98.4% accuracy with 0.003 expected calibration error and roughly 10 ms measured latency, versus 97.4%, 0.007 and 331 ms for TypeSafe Jev 1.13, though the company notes the latency figures came from different serving environments. The model is released under Apache 2.0 with weights on Hugging Face, a pip-installable package, and a free API tier of 200 decisions per minute.", "body_md": "**70 million tokens processed. 322 million parameters. One forward pass. Free API. Fully open-source.**\n\nA few days ago, we open-sourced **THX-01**, our 322M-parameter non-autoregressive decision model designed to make AI decisions faster, more efficient, and accessible to everyone.\n\nToday, we've reached another milestone: **over 70 million tokens processed through our THX-01 API since launch.**\n\nAnd we're just getting started.\n\nMost AI applications don't actually need to generate long responses.\n\nSometimes, all you need is a reliable decision:\n\nInstead of generating tokens one by one, THX-01 processes a document or message alongside typed questions and returns structured answers with calibrated probabilities in a single forward pass.\n\nOn our GPU benchmark, inference takes approximately **10 milliseconds**.\n\nWe evaluated THX-01 against TypeSafe Jev 1.13 on our internal benchmark of 2,843 multilingual support tickets.\n\n| Metric | THX-01 | Jev 1.13 | \n|---|---|---|\n| Accuracy | **98.4%** | 97.4% | \n| Calibration Error (ECE) | **0.003** | 0.007 | \n| Measured Latency | **~10 ms** | 331 ms | \n| Open-Source Weights | Yes | No | \n\nTHX-01 outperformed Jev on this classification benchmark, with approximately 33x lower measured latency.\n\nThe latency figures come from different serving environments, so they should not be interpreted as a hardware-controlled comparison.\n\nOur results show that small, specialized decision models can deliver competitive performance without requiring billions of additional parameters.\n\nTHX-01 is trained using **Reinforcement Learning for Calibrated Decisions (RLCD)**.\n\nRather than optimizing only for correct predictions, RLCD encourages probability estimates that reflect the model's actual uncertainty.\n\nThis is particularly useful for AI agents and production applications that need to decide whether to accept a prediction, request additional information, or escalate to another model.\n\nTHX-01 supports:\n\nThe model supports 18 post-training languages, including Azerbaijani, English, Russian, and Turkish.\n\nAnd at just 322M parameters, it can even run locally on smartphones with a compatible inference runtime.\n\nWe believe developers should be able to experiment with new AI architectures without worrying about infrastructure costs.\n\nThat's why we've opened the **THX-01 API with a free tier of 200 decisions per minute**.\n\nYou can test it with a simple request:\n\n```\ncurl https://api.hal-x.ai/v1/systemone \\\n  -H \"Content-Type: application/json\" \\\n  -d '{\n    \"state\": \"My card was charged twice for one order\",\n    \"questions\": {\n      \"team\": {\n        \"type\": \"choice\",\n        \"question\": \"Which team should handle this?\",\n        \"criteria\": {\n          \"billing\": \"Billing issues\",\n          \"technical\": \"Technical issues\",\n          \"sales\": \"Sales inquiries\"\n        }\n      }\n    }\n  }'\n```\n\nCheck the API documentation for the current authentication requirements and usage limits.\n\n**API Documentation:** [https://api.hal-x.ai/docs/thx-01](https://api.hal-x.ai/docs/thx-01)\n\nWe released THX-01 under the **Apache 2.0 license**.\n\nDevelopers can access the model weights, source code, and training resources on Hugging Face.\n\nYou can also install the Python package directly:\n\n```\npip install thx01\n```\n\n**Model & Source Code:** [https://huggingface.co/doofz/THX-01](https://huggingface.co/doofz/THX-01)\n\n**Interactive Demo:** [https://huggingface.co/spaces/doofz/THX-01-demo](https://huggingface.co/spaces/doofz/THX-01-demo)\n\nAt HAL-X AI, we're working toward a future where AI systems don't just generate answers, but make reliable, measurable, and well-calibrated decisions.\n\nWe believe the next generation of AI infrastructure will need specialized models that are smaller, faster, and capable of expressing uncertainty.\n\n**70 million tokens processed is an exciting milestone for our team, but this is only the beginning.**\n\nWe're continuing to improve THX-01, expand its capabilities, and contribute to the global open-source AI ecosystem.\n\nOur mission is simple:\n\n**We need more AI models that know how to make decisions, how confident they should be, and when they don't know.**\n\nAnd we're building them for everyone.\n\n**Try THX-01 today:**\n\n`pip install thx01`", "url": "https://wpnews.pro/news/thx-01-just-processed-70-million-tokens-our-open-source-decision-model-is-just", "canonical_source": "https://dev.to/montydrief/thx-01-just-processed-70-million-tokens-our-open-source-decision-model-is-just-getting-started-2o8l", "published_at": "2026-10-12 09:36:54+00:00", "updated_at": "2026-10-12 09:40:43.686504+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "ai-agents", "ai-tools", "ai-research"], "entities": ["HAL-X AI", "THX-01", "TypeSafe Jev 1.13", "Hugging Face", "Apache 2.0", "RLCD"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/thx-01-just-processed-70-million-tokens-our-open-source-decision-model-is-just", "markdown": "https://wpnews.pro/news/thx-01-just-processed-70-million-tokens-our-open-source-decision-model-is-just.md", "text": "https://wpnews.pro/news/thx-01-just-processed-70-million-tokens-our-open-source-decision-model-is-just.txt", "jsonld": "https://wpnews.pro/news/thx-01-just-processed-70-million-tokens-our-open-source-decision-model-is-just.jsonld"}}