Tag
Liquid AI releases LFM2.5-230M, a small 230M parameter model optimized for fast inference on CPUs, NPUs, and GPUs, targeting agentic tasks on devices like phones and robots.
Liquid AI released LFM2.5-230M, a compact 230M-parameter hybrid model optimized for on-device deployment with fast edge inference speeds (213 tok/s on Galaxy S25 Ultra) and built for agentic tasks via reinforcement learning.
SupraLabs released supra-title-FFT-preview, a full fine-tuned 0.4B parameter model for chat title generation, trained on 115K samples — nearly 10x larger than their previous dataset.
Nanbeige 4.1, a 3B model, outperforms Qwen3-30b-A3b and Qwen 3.5 4b in coding tasks with focus on algorithmic efficiency, achieving long horizon tasks with 600+ tool calls.
VibeThinker, a 3B parameter model fine-tuned on Qwen 2.5, achieves performance comparable to Claude Opus 4.5 and much larger models like DeepSeek v3 through innovative post-training that includes multi-path thinking and staged training on math, coding, and science.
Microsoft is testing Phi Silica support on Nvidia GPUs, allowing developers to run the small language model locally on Windows devices with RTX 30-series or newer GPUs, though it lacks NPU-only features like prompt compression.
Weibo's VibeThinker-3B, a 3B parameter model, claims to match or exceed the reasoning performance of much larger models like DeepSeek V3.2 and Gemini 3 Pro on math and coding benchmarks, sparking debate over benchmark reliability and the necessity of scaling.
The VibeThinker-3B model achieves state-of-the-art math and coding reasoning performance, scoring 94.3 on AIME'26 and 96.1% on unseen LeetCode problems, demonstrating that small models can reach frontier-level reasoning in verifiable domains.
Cleo is a finetuned version of Qwen3.5-2B-Base designed for text-to-SQL tasks, using a unified harness for training and inference that supports live execution evidence and safety checks. All code, model, and datasets are open-source.
Microsoft launches Fara-7B, an efficient Computer Use Agent with only 7B parameters, surpassing larger models on web tasks, supporting pure local deployment, and achieving low-cost desktop automation.
Microsoft has released Fara-7B, a small 7B-parameter language model focused on pure local desktop automation. It can directly take over your mouse and keyboard to execute repetitive workflows, with low cost and no need for internet connectivity.
Supra Labs released Supra Title, a 350M parameter model specialized for generating chat conversation titles. Built on LFM2.5, it runs on any hardware in GGUF format and requires no system prompt.
VibeThinker-3B is a 3B-parameter model that achieves frontier-level reasoning performance on math, coding, and STEM benchmarks by optimizing the Spectrum-to-Signal Principle (SSP) post-training pipeline, reaching performance comparable to much larger models.
This paper introduces IAPO, a reinforcement learning algorithm that improves tool-calling capabilities in multimodal small language models by aligning input attribution with a stronger teacher. Experiments on Qwen2.5-VL-3B show an average 3% improvement in visual question answering accuracy across six test sets.
The author shares learnings from training a 160M parameter LLM from scratch, experimenting with architectures like multi-token prediction and hierarchical reasoning models. They emphasize the importance of fast iteration, simplifying ideas, and understanding why architectures work.
Microsoft announced two new on-device AI models at Build 2026: Aion 1.0 Instruct, an open-weights small language model, and Aion 1.0 Plan, a 14B parameter reasoning and tool-calling model for local agentic workflows.
Trained a 75M parameter LLM called KeyLM from scratch on 18B tokens, achieving competitive instruction-following scores against larger models while using fewer parameters and less data.
OpenBMB releases MiniCPM5-1B, a leading 1B open weights LLM that achieves the highest Artificial Analysis Intelligence Index score (17.9) in its size class, surpassing larger models like Qwen3.5 2B while using fewer parameters.
MiniCPM5-1B is a new 1B parameter AI model from OpenBMB featuring hybrid reasoning with Think/No-Think modes, 128K context, and Apache 2.0 license, running on various hardware.
MiniCPM5-1B is fully open-sourced with weights, training data, and deployment code; it achieves top scores among sub-2B models and runs on edge devices.