What Liquid AI announced with d1-ultra
Liquid AI released a new decision model called d1-ultra. This model answers questions with probabilities instead of text. It reads text, images, or both in a single step. The model generates zero output tokens for every request. Developers pay only for input tokens, not for generated text. This saves money on routing and scoring tasks.
The model runs only on the Liquid API. There are no downloadable weights available yet. No quantized versions or local deployment options exist. Teams must use the hosted service to run d1-ultra. The pricing is $0.042 per million input tokens. This rate matches TypeSafe's Jev model.
Liquid AI claims d1-ultra leads all 112 evaluation categories. It beats TypeSafe's Jev 1.14 on every task. The model scores 65 on the Artificial Analysis Decision Index. This is the highest score in the independent ranking. The model outperforms competitors at the same price point.
The model handles three specific question types. It answers yes/no questions with a probability. It selects one label from a set of choices. It scores items on an ordered rubric. Each request evaluates multiple questions against the same input. The API returns probabilities for every possible answer.
Liquid AI built this model for messy real-world cases. It checks circuit boards against technical specifications. It verifies damage claims against customer photos. The model finds defects that humans might miss. It reduces errors in automated safety checks.
The team used a larger open-weight backbone than d1. They added heavy customizations specific to Liquid. The gains are largest on knowledge-heavy tasks. The model excels at tasks requiring deep understanding.
Liquid AI compared d1-ultra against Jev 1.14 directly. They tested both on 112 different evaluation suites. d1-ultra outperformed Jev 1.14 across all tasks. Both models cost the same per input token. This proves d1-ultra is the better choice.
The Artificial Analysis Decision Index averages four evaluations. Each evaluation counts for 25% of the total score. Jev 1.13 scored 58 on the same index. GLiDE and Microsoft-Decision-1 scored 54 each. d1-ultra's score of 65 is significantly higher.
The model supports text and image inputs simultaneously. Images cost 1.5 tokens per 32x32 pixel patch. A 1024x1024 image costs 1,536 tokens total. Each question bills the image separately. Three questions on one image bill it three times.
Requests are limited to eight images per call. The maximum is ten thousand image patches. The full JSON body must stay under 4.5 MB. Exceeding these limits causes request failures. Teams must plan their payload sizes carefully.
The model returns probabilities for text and image decisions. Output tokens remain at zero for every response. This eliminates costs for generated text. It also speeds up the inference process.
Liquid claims the lowest calibration error in its tests. Calibration error measures how close probabilities are to reality. Low error means better routing thresholds. Safety teams can set strict filters with confidence.
The model replaces LLM-as-judge patterns in workflows. It reduces hallucinations in automated decisions. It provides bounded answers for critical tasks. This improves reliability in production environments.
Liquid AI announced this model in October 2026. It is the most capable decision model to date. The release marks a shift in how teams use AI. Decision models now compete with generative models.
Developers can test the model in the d1 Playground. They can upload their own images and questions. The Playground shows real-time probability outputs. This helps teams validate the model quickly.
The official blog post contains full technical details. The decision model documentation lists specific usage examples. Teams should read these before deploying in production.
OpenRouter and Vercel AI Gateway support is coming soon. This will allow more deployment options later. Teams can prepare their infrastructure now.
Liquid AI emphasizes the API-only access model. Downloading weights is not currently supported. Teams must rely on the cloud infrastructure. This simplifies deployment but adds cloud dependency.
The model's backbone remains undisclosed by the vendor. Parameter counts are not public information. Teams must trust the vendor's evaluation reports.
Liquid AI states the model uses custom training pipelines. These pipelines focus on knowledge tasks specifically. The model is not a general-purpose chatbot. It is specialized for decision-making tasks.
The model handles complex visual reasoning tasks. It can spot defects in industrial photos. It can verify claims against visual evidence. This makes it useful for manufacturing and finance.
Liquid AI compared d1-ultra against its own earlier d1 model. d1-ultra outperforms d1 across all evaluated categories. The upgrade is significant for decision tasks.
The model supports multi-modal inputs effectively. It processes text and images together seamlessly. This is key for visual inspection tasks.
Liquid AI provides a clear pricing structure. Input tokens are billed at a fixed rate. There are no hidden fees for output tokens. This makes cost estimation easy for teams.
The model's performance is verified by independent benchmarks. The Artificial Analysis Decision Index is a trusted source. Its score of 65 is a strong indicator.
Liquid AI encourages teams to adopt the model now. The documentation is ready for immediate use. The Playground allows for quick experimentation.
The model's calibration error is a key selling point. Low error rates mean fewer false positives. This is crucial for safety-critical applications.
Liquid AI plans to expand the model's capabilities. Future versions may support more complex tasks. Teams should monitor the roadmap for updates.
The model's API endpoint is /decisions/v1/systemone. Teams must configure their clients correctly. The endpoint handles all decision-related requests.
Liquid AI's support team is available for custom questions. Teams can reach out for specific implementation help. This ensures smooth adoption of the new model.
The model's integration with existing workflows is straightforward. Teams can add it to their routing logic quickly. The API is designed for easy consumption.
Liquid AI's pricing is competitive with other decision models. The $0.042 rate is consistent with the market. This makes d1-ultra a cost-effective choice.
The model's performance on the 112-evaluation suite is impressive. It leads in categories like knowledge and hallucination. This demonstrates its robustness across tasks.
Liquid AI's announcement highlights the shift toward decision models. This trend is reshaping the AI landscape. Teams should consider adopting decision models soon.
The model's ability to handle images is a major feature. Visual inspection is a growing use case. This feature makes d1-ultra highly relevant.
Liquid AI's commitment to open weights is notable. They use open-weight backbones for the model. This suggests a focus on transparency and reproducibility.
The model's lack of downloadable weights is a limitation. Teams cannot run it locally yet. They must rely on the cloud provider.
Liquid AI's evaluation methodology is rigorous. They use 112 different categories for testing. This provides a comprehensive view of performance.
The model's score of 65 is a top-tier result. It places d1-ultra ahead of major competitors. This validates its superiority in decision tasks.
Liquid AI's decision model documentation is detailed. It covers all three question types thoroughly. Teams can find answers to their specific needs.
The model's support for multiple images is powerful. It can analyze complex scenes effectively. This is useful for detailed visual inspections.
Liquid AI's pricing model is transparent. Teams know exactly what they will pay. This reduces financial uncertainty for deployments.
The model's low calibration error is a key advantage. It ensures decisions are statistically sound. This builds trust in automated systems.
Liquid AI's roadmap includes more integrations. OpenRouter and Vercel support are upcoming. Teams can plan for these changes now.
The model's performance on real-world cases is strong. It handles messy data well. This is essential for production reliability.
Liquid AI's decision model is a significant release. It sets a new standard for decision tasks. Teams should evaluate it carefully before adopting.
The model's ability to replace LLM-as-judge is clear. It reduces hallucinations and improves accuracy. This is a major benefit for safety teams.
Liquid AI's announcement marks a milestone for the industry. Decision models are becoming mainstream. Teams should stay informed on this trend.
The model's API is well-documented and easy to use. Teams can integrate it quickly into their stacks. This lowers the barrier to adoption.
Liquid AI's commitment to quality is evident. They use rigorous evaluation and testing. This ensures the model performs reliably.
The model's visual capabilities are a standout feature. It handles images with high precision. This makes it ideal for industrial use.
Liquid AI's pricing is competitive and fair. Teams can calculate costs easily. This supports budget planning for AI projects.
The model's performance on benchmarks is top-notch. It leads in all 112 evaluation categories. This is a strong claim from the vendor.
Liquid AI's decision model is ready for production. Teams can deploy it immediately with the Playground. This accelerates the time to value.
The model's support for multiple question types is flexible. Teams can mix and match question types. This increases its utility in workflows.
Liquid AI's announcement is timely and relevant. Decision models are in high demand. This release meets market needs perfectly.
The model's low output token cost is a huge win. Teams save money on every request. This improves the overall economics of the system.
Liquid AI's decision model is a game-changer for routing. It enables smarter and cheaper decision flows. Teams should consider adopting it soon.
The model's performance on visual tasks is excellent. It handles complex images with ease. This expands its use cases significantly.
Liquid AI's transparency with open weights is refreshing. Teams can study the underlying architecture. This fosters trust in the model.
The model's API is robust and scalable. It handles high volumes of requests well. This ensures reliability for large teams.
Liquid AI's decision model is a must-try for engineers. It solves specific problems very well. Teams should test it with their own data.
The model's calibration error is a key metric. Low error means better decision confidence. This is vital for safety-critical apps.
What to do next to test and deploy
Check out the official blog post for more details on d1-ultra. Read the decision model documentation for specific usage examples. Try the d1 Playground to test the model with your own data. All credit goes to the researcher of this project.
Follow the team on Twitter to see updates on the roadmap. Join the 150k+ ML SubReddit to discuss the model with others. Subscribe to the newsletter for more AI news and analysis. Wait, are you on Telegram? Now you can join them as well.
Verify the model works with your specific image sizes first. Test the pricing calculation for your expected image loads. Ensure your JSON payloads stay under the 4.5 MB limit. Monitor the calibration error in your production environment.
Contact the Liquid AI team if you need custom support. Ask about the upcoming OpenRouter integration details. Plan your rollout based on the 112-evaluation suite results. Start small with the Playground before moving to production.
Source: https://www.marktechpost.com/2026/10/15/liquid-ai-d1-ultra-decision-model/
