Applications

Can a compact late interplay retriever index as soon as and ship correct cross lingual search with quick inference? Liquid AI launched LFM2-ColBERT-350M, a compact late interplay retriever for multilingual and cross-lingual search. Documents could be listed in a single language, queries could be written in lots of languages, and the system retrieves with excessive…

Can we render lengthy texts as photos and use a VLM to attain 3–4× token compression, preserving accuracy whereas scaling a 128K context towards 1M-token workloads? A staff of researchers from Zhipu AI launch Glyph, an AI framework for scaling the context size through visual-text compression. It renders lengthy textual sequences into photos and processes…

Large language fashions (LLMs) supply a number of parameters that allow you to fine-tune their habits and management how they generate responses. If a mannequin isn’t producing the specified output, the problem usually lies in how these parameters are configured. In this tutorial, we’ll discover among the mostly used ones — max_completion_tokens, temperature, top_p, presence_penalty,…

Liquid AI launched LFM2-VL-3B, a 3B parameter imaginative and prescient language mannequin for picture textual content to textual content duties. It extends the LFM2-VL household past the 450M and 1.6B variants. The mannequin targets greater accuracy whereas preserving the velocity profile of the LFM2 structure. It is obtainable on LEAP and Hugging Face beneath the…

In this tutorial, we discover LitServe, a light-weight and highly effective serving framework that permits us to deploy machine studying fashions as APIs with minimal effort. We construct and check a number of endpoints that exhibit real-world functionalities resembling textual content era, batching, streaming, multi-task processing, and caching, all operating regionally with out relying on…

Do you really want a large VLM when dense Qwen3-VL 4B/8B (Instruct/Thinking) with FP8 runs in low VRAM but retains 256K→1M context and the total functionality floor? Alibaba’s Qwen workforce has expanded its multimodal lineup with dense Qwen3-VL models at 4B and 8B scales, every delivery in two job profiles—Instruct and Thinking—plus FP8-quantized checkpoints for…

How a lot functionality can a sparse 8.3B-parameter MoE with a ~1.5B lively path ship in your telephone with out blowing latency or reminiscence? Liquid AI has launched LFM2-8B-A1B, a small-scale Mixture-of-Experts (MoE) mannequin constructed for on-device execution underneath tight reminiscence, latency, and vitality budgets. Unlike most MoE work optimized for cloud batch serving, LFM2-8B-A1B…

We will construct a Regression Language Model (RLM), a mannequin that predicts steady numerical values immediately from textual content sequences on this coding implementation. Instead of classifying or producing textual content, we concentrate on coaching a transformer-based structure that learns quantitative relationships hidden inside pure language descriptions. We begin by producing artificial text-to-number information, tokenizing…

Thinking Machines has launched Tinker, a Python API that lets researchers and engineers write coaching loops regionally whereas the platform executes them on managed distributed GPU clusters. The pitch is slim and technical: preserve full management of knowledge, goals, and optimization steps; hand off scheduling, fault tolerance, and multi-node orchestration. The service is in non-public…