AI FRONTIER
Signals worth understanding before they become obvious.
OpenRouter new model: LiquidAI LFM3-1.7B
LiquidAI's LFM3-1.7B is now listed on OpenRouter: a compact 1.7B-parameter text model with a 262,144-token context, post-trained for agents, tool calling, and structured outputs.
- 1.7B parameters with a 262,144-token context window and support for up to 8,192 completion tokens.
- Priced at $0.08 per 1M input tokens and $0.20 per 1M output tokens, with cache reads at $0.02 per 1M tokens.
- Accepts tools and tool_choice for function calling, and supports structured outputs via a JSON schema in response_format.
- The OpenRouter page lists a release date of November 6, 2026 and states that prompts and outputs may be retained and used to train Liquid models.
For agent and structured-output workloads that need long context at a low per-token price, LFM3-1.7B is a small hosted option worth testing, provided the data-retention terms and the page's reported throughput fit the use case.
Palette: teaching a small model to build decks in your design language
IBM Research's Palette splits slide generation into a designer role and a coder role served by a single LoRA adapter on a 20B model, and it ranks first on two of three PresentBench domains.
- Palette fine-tuned a 20B model; training data follows the IBM Design Language, with the designer writing a structured JSON brief and the coder writing pptxgenjs JavaScript.
- Because each slide's context is fixed before any code is written, the coder calls run in parallel, making the coding stage about 6–7× faster.
- Training used LoRA rather than full fine-tuning; coder examples are about 5× longer than designer examples, so designer examples were oversampled 5×.
- The synthetic data covers around 1000 decks and roughly 10,000 slides, each rendered and checked for overlaps and overflow, with repaired slides' reasoning rewritten as counterfactuals.
For teams that need their own design language, a verifiable render checker plus reasoning-bearing synthetic data lets a 20B self-hosted model approach frontier output.
nvidia releases Nemotron-3.5-Transcribe-30B-A4B, an offline 40-language speech model
NVIDIA published Nemotron-3.5-Transcribe-30B-A4B on Hugging Face, an open-weight offline speech and audio model covering transcription in 40 languages, recordings up to one hour, and a 65,536-token context, with a mixture-of-experts architecture activating about 4.3B parameters including the audio encoder.
- 34B total parameters with about 4.3B active, including an approximately 800M-parameter audio encoder and about 3.5B active parameters in the language backbone
- 65,536-token context and recordings up to one hour; weights released in BF16 with FP8 and NVFP4 quantized variants
- Reported mean WER of 4.71 on 9 public sets of the Open ASR Leaderboard, versus 4.94 for the next best open-weight model, Qwen3-ASR-1.7B
- License is OpenMDW-1.1 and deployment requires 1× H100 80 GB; release date is October 28, 2026
For speech application builders, this open-weight checkpoint offers commercial-use-licensed offline transcription across 40 languages, but capacity planning must account for 1× H100 80 GB per deployment.
Granite Switch 4.1.1 — Update models like you update software
Granite Switch 4.1.1 is a patch release fully backwards compatible with 4.1, requiring no migration. It updates the Granite Speech ASR model to multilingual speech-to-text across 10 languages with native punctuation and capitalization, and updates the Decision adapter in granitelib-core; apart from training the new ada...
- The patch release is fully backwards compatible with 4.1 with no migration required.
- The updated Granite Speech ASR model covers speech-to-text across 10 languages with native punctuation and capitalization.
- The updated Decision adapter in granitelib-core features improved overall performance on the Jev Decision Index test set.
- Apart from training the new ASR model and Decision adapter, nothing was recomputed and the base model is untouched.
For builders, modular adapters let model updates ship like software patches: compose the latest adapters and deploy without retraining the base model.
Do AI models lie? Researchers set 1,532 traps for nine top large models
Researchers set 1,532 test traps for nine top large models to examine whether the models lie.
- The test targets nine top large models.
- A total of 1,532 traps were set.
For AI builders, this kind of deception-focused test design can serve as a reference direction for evaluating model trustworthiness.
AI agents attempted to research open questions and none reached expert standard
A paper abstract reports that AI agents attempted to research open questions, and none reached the expert standard.
- The paper abstract states that AI agents attempted to research open questions.
- The paper abstract states that none reached the expert standard.
For builders of research agents, the abstract indicates current performance on open questions falls short of the expert standard.