Liquid Ai
PulseAugur coverage of Liquid Ai — every cluster mentioning Liquid Ai across labs, papers, and developer communities, ranked by signal.
- 2026-07-29 product_launch Liquid AI released LFM2.5-Encoders, a model for fast long-text analysis on CPUs. source
- 2026-07-29 product_launch Liquid AI released two new open-weight bidirectional encoder models, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M. source
- 2026-06-19 product_launch Liquid AI released two new retrieval models for multilingual search. source
- 2026-06-19 product_launch Liquid AI released two new retrieval models, LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M, designed for efficient multilingual search on edge devices. source
- 2026-05-29 product_launch Liquid AI released its LFM2.5-8B-A1B model, featuring an expanded context window and increased training data. source
- 2026-05-26 product_launch Liquid AI launched two new Japanese-capable AI models, LFM2.5-1.2B-JP-202606 and LFM2.5-Audio-1.5B-JP. source
4 day(s) with sentiment data
Liquid AI's LFM2.5-2.6B model to see widespread adoption in on-device agent applications within 6 months
The release of LFM2.5-2.6B, with its strong emphasis on agentic capabilities, compact size, and ability to run on-device with a large context window, positions it well for use in applications requiring local AI processing. The partnership with MacPaw for macOS integration further suggests a strategic push towards broader adoption in consumer devices. We expect to see this model integrated into various agentic applications targeting mobile and desktop within the next six months.
Liquid AI's LFM2.5-2.6B model demonstrates strong performance in tool-use and instruction-following benchmarks
Multiple releases highlight that Liquid AI's LFM2.5-2.6B model, despite its small parameter count (2.69B), shows competitive performance against much larger models in specific benchmarks, particularly instruction-following and tool-use. This suggests that Liquid AI is effectively optimizing for agentic workflows and efficient execution on resource-constrained devices.
Liquid AI to release a fine-tuned version of LFM2.5-2.6B optimized for coding tasks within 3 months
While current reports indicate LFM2.5-2.6B is not recommended for complex coding, its compact size and agentic focus make it a strong candidate for specialized fine-tuning. Given the general demand for capable coding assistants, Liquid AI may prioritize developing a version of this model specifically for code generation or assistance, leveraging its efficient architecture. We hypothesize such a release within the next three months.
-
Liquid AI releases compact agent model; Mistral launches safety classifier
Liquid AI has released LFM2.5-2.6B, a compact text-only model optimized for agent harnesses and tool interaction, featuring a large context window and multilingual support. While not recommended for complex coding, its …
-
MacPaw and Liquid AI partner for on-device AI on Macs
MacPaw and Liquid AI have formed a partnership to develop on-device AI capabilities for macOS. This collaboration will integrate Liquid AI's foundation models with MacPaw's Elix and Mnemos technologies to create a local…
-
Liquid AI releases on-device agentic model LFM2.5-2.6B with 128K context
Liquid AI has released LFM2.5-2.6B, an open-weights, on-device agentic model designed for mobile and edge devices. This model boasts 2.69 billion parameters, a 128,000-token context window, and can perform multi-step ta…
-
Liquid AI releases 2.6B model capable of 128K context on phones
Liquid AI has released LFM2.5-2.6B, a compact language model designed for local AI applications. Despite its small size of 2.69 billion parameters, the model boasts a 128K context window and supports tool calling, makin…
-
Liquid AI releases LFM2.5-2.6B model with agentic focus
Liquid AI has released LFM2.5-2.6B, a new small-scale language model. The release emphasizes enhanced agentic capabilities, aiming to excel at high-volume tasks such as document summarization. The developers are particu…
-
Liquid AI releases fast, CPU-friendly bidirectional encoders with 8K context
Liquid AI has released two new open-weight bidirectional encoder models, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M. These models are built on the LFM2 hybrid backbone and support an 8,192-token context window, designe…
-
Liquid AI open-sources Antidoom to fix AI model doom loops · 2 sources tracked
Liquid AI has introduced Antidoom, an open-source method designed to mitigate "doom loops" in AI reasoning models. This issue causes models to repeatedly output the same text, consuming the context window, and is partic…
-
Liquid AI releases smallest on-device model, LFM2.5-230M
Liquid AI has launched LFM2.5-230M, their most compact model yet. This open-weight model, with 230 million parameters, is designed for on-device operation on edge hardware. It achieves speeds of 213 tokens/s on a Samsun…
-
Liquid AI ships tiny LFM2.5-230M for on-device agent tasks
Liquid AI has released LFM2.5-230M, its smallest model to date, designed for on-device inference on edge hardware like phones and robots. This 230-million-parameter model excels at data extraction and tool use, outperfo…
-
OpenAI internal AI usage surges; new open models and agent capabilities emerge
OpenAI has reported a significant increase in internal AI model usage across various departments, with median output tokens for Codex growing by as much as 56x in Research since November 2025. This surge in token consum…
-
Liquid AI releases compact multilingual search models for edge devices
Liquid AI has launched two new retrieval models, LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M, each with 350 million parameters. These models are designed for efficient multilingual search across 11 languages and can o…
-
AI-DNA startup Radical Numerics raises $50M seed funding
Radical Numerics, a startup founded by researchers who developed AI models for DNA generation, has secured $50 million in seed funding. The company aims to revolutionize biology by teaching AI to understand and generate…
-
New tool simplifies local AI model selection with curated filters
A new tool aims to simplify the selection of local AI models by curating options from various labs, including Google and Mistral. The platform allows users to filter models based on specific criteria to find suitable on…
-
Liquid AI releases free Japanese-language AI model for edge inference
Liquid AI has released a new Japanese-language AI model designed for edge inference, making it freely available. This model is capable of processing voice and language, enabling on-device AI capabilities. The release ai…
-
Tutorial shows LFM2 fine-tuning with QLoRA and DPO
This tutorial demonstrates how to fine-tune the LFM2 model using QLoRA and Direct Preference Optimization (DPO) on Google Colab. It covers loading the base LFM2 model with 4-bit quantization, preparing a dataset for sup…
-
Anthropic ships Opus 4.8; Tencent model tops Claude; Groq seeks $650M
Anthropic has released Opus 4.8, though details about its capabilities are scarce. Separately, a model developed by Tencent has surpassed Anthropic's Claude in performance on the OpenRouter platform. Meanwhile, Groq is …
-
Liquid AI ships 8B MoE model with 128K context for on-device tool use
Liquid AI has released LFM2.5-8B-A1B, an 8-billion parameter Mixture-of-Experts model designed for efficient on-device tool calling. This new version significantly expands the context window to 128,000 tokens and scales…
-
Liquid AI launches Japanese text and audio LLMs
Liquid AI has released two new Japanese-language models: LFM2.5-1.2B-JP-202606, a text-based chat model that shows improved performance in knowledge, instruction following, and coding, and LFM2.5-Audio-1.5B-JP, a multim…
-
Liquid AI ships LFM2.5-8B-A1B on-device MoE model
Liquid AI has released LFM2.5-8B-A1B, a new on-device Mixture-of-Experts (MoE) model designed for complex tasks and tool chaining. This model features 8.3 billion total parameters but activates only 1.5 billion per toke…
-
Liquid AI releases LFM2-24B-A2B, an efficient 24B parameter MoE model
Liquid AI has released an early checkpoint of its LFM2-24B-A2B model, a sparse Mixture of Experts (MoE) architecture with 24 billion total parameters and 2 billion active parameters per token. This model demonstrates th…