Accessibility Adjustments

Use these optional tools to adjust reading and display preferences. These tools cannot resolve every accessibility barrier. Please contact the website owner if you need assistance.

  • Text adjustments
  • Content scaling 100%
  • Font size 100%
  • Line height 100%
  • Letter spacing 100%
  • Colour adjustments
  • Orientation adjustments

Liquid AI d1.2 adds audio and typed answers, with API limits

Listen to this article

Liquid AI released d1.2 and d1.2-ultra on November 16, adding audio and new structured answer types to its hosted decision models. They are available through Liquid’s API. The launch announcement describes OpenRouter and Vercel support as coming soon, not available at launch.

The useful change is broader than another classification score. Developers can extract fields, locate objects and interpret recordings without asking a general chatbot to write an answer and then parsing its prose. That still leaves applications responsible for checking uncertainty and deciding what happens next.

What the new answers can do

The decision model overview describes five additions. These cover reading text, numbers or dates, finding passages, locating objects, estimating a quantity within a range and selecting multiple applicable categories. A support workflow could identify both delivery and billing issues instead of forcing one label.

Extraction and estimation are different operations. Reading a number preserves the source span and does not convert its unit. Estimation returns a median and interval instead. Missing extracted values carry a found signal and can return null. Object boxes require exactly one image. Treat these distinctions as part of the application’s data contract, not interchangeable formats.

These are hosted models, not new downloadable d1.2 weights. Liquid separately lists local d1-2B and d1-omni models. Its overview still recommends a language model for conversation, summaries, generated code and open-ended reasoning.

Pricing counts shared inputs once

The current price list puts d1.2 at $0.02 and d1.2-ultra at $0.042 per million input tokens, with no output charge. Ultra’s input rate is 2.1 times the standard model’s rate. That price difference should be weighed against errors on the tasks a team actually needs.

The hosted API guide bills shared state, images and audio once per request. Each question adds its own instructions, options and answer slots. Raising the requested maximum for passages or boxes therefore increases input usage.

Audio costs 10 input tokens per second, rounded up per clip. The current audio requirements accept WAV with PCM or floating-point samples, MP3, FLAC and OGG with Opus or Vorbis. A request can contain up to eight clips totaling 300 seconds, while the entire encoded JSON body must stay under 4.5 MB, including audio, images, state and questions.

That size ceiling can limit uncompressed 16-bit mono PCM WAV at 16 kHz to roughly 100 seconds. It is not a universal duration limit. Compressed formats can fit longer recordings within the same request budget, subject to the 300-second maximum. Check the encoded request size, not just the source file size.

Correction, November 17, 2026. An earlier version incorrectly stated that 16-bit PCM WAV was required. The supported formats and request limits are corrected above.

Migration has limits worth checking

The guide permits up to 128 questions. Each question, including shared media, must fit 65,536 tokens and 250,000 bytes of text. The older d1, d1.1 and d1-ultra IDs are deprecated but still support their original question types. They do not gain audio or the new typed tasks.

That is an important distinction from the earlier d1 Ultra release. Reusing an old model identifier is not the same as migrating. Before switching production traffic, test missing fields, ambiguous dates and recordings with background noise, then compare the returned structures with what downstream code expects.

Benchmark claims are not a deployment test

Liquid reports that d1.2-ultra scores 75 on Artificial Analysis’s Decision Index v2.1. It also states that no public benchmark yet covers the five new task types, so its results for those additions come from internal evaluations using held-out public datasets. Those are different kinds of evidence and should not be collapsed into a universal accuracy claim.

For a pilot, measure false positives and missed cases separately, preserve an uncertain-result path and inspect the actual token usage. A low per-token price does not compensate for silently forwarding an incorrect extracted value. ByteForward has not independently benchmarked these models.

Featured image is an AI-generated editorial illustration by ByteForward. Liquid logo belongs to Liquid AI.

Jordan Reid
Jordan Reid

Jordan Reid is focused on AI tools, agents, developer products, and the way technology changes everyday work. Jordan approaches a launch from the user’s side of the screen. What can it actually help someone finish? The voice is practical, conversational, and skeptical of products that turn a simple job into five new settings. Coverage follows coding assistants, creative software, browser agents, and the workflows around them, with attention to pricing, permissions, setup, and the human work that remains.

Leave a Reply

Your email address will not be published. Required fields are marked *

Gravatar profile