Best Small Language Models in the UK - Page 3

Nemotron 3 Nano is the smallest model in the NVIDIA Nemotron 3 family, built for agentic AI applications with strong reasoning, conversational ability, and cost-efficient inference. It is a hybrid Mamba-Transformer Mixture-of-Experts model with 3.2 billion active parameters, 3.6 billion including embeddings, and 31.6 billion total parameters. NVIDIA describes it as more accurate than the previous Nemotron 2 Nano while activating less than half of the parameters per forward pass, improving efficiency without sacrificing performance. The model is positioned as more accurate than GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507 on popular benchmarks across different categories. On an 8K input and 16K output setting using a single H200, it delivers inference throughput 3.3 times higher than Qwen3-30B-A3B and 2.2 times higher than GPT-OSS-20B. Nemotron 3 Nano supports context lengths up to 1 million tokens and is reported to outperform GPT-OSS-20B and Qwen3-30B-A3B-Instruct-2507.

Aion-1.0-Instruct is a pre-release small language model introduced in Microsoft Edge as a developer preview for early testing and feedback. It is designed to power Edge’s on-device Prompt and Writing Assistance APIs, giving web developers a faster, smaller, and more efficient model for AI-powered browser experiences. Microsoft previously used Phi-4-mini for these APIs, but its hardware requirements limited availability across devices. Aion-1.0-Instruct expands support to significantly more devices, including machines with less capable GPUs and, through CPU inference, devices without a GPU, while still delivering strong quality for a wide range of web use cases. The model is available in Edge Canary and Dev channels, allowing developers to evaluate it in real-world web scenarios, test API interoperability, and provide feedback before final optimizations. Aion-1.0-Instruct is meant to help developers build AI features directly into websites and browser extensions.

SubQ 1.1 Small is a long-context AI model from Subquadratic designed to reason over complete enterprise artifacts such as codebases, document collections, contracts, and financial filings. It uses Subquadratic Sparse Attention, or SSA, to reduce the high compute costs normally associated with processing very large context windows. The model delivers near-perfect long-context retrieval across 1M, 2M, 6M, and 12M token tests while using far less attention compute than dense attention. SubQ 1.1 Small also maintains strong general reasoning, coding, knowledge, and agentic task performance across multiple benchmarks. Its capabilities make it useful for financial analysis, legal review, contract work, software engineering, due diligence, and other workflows where information is spread across large artifacts. SubQ is built for organizations that want to move beyond fragmented retrieval pipelines and enable direct reasoning over massive bodies of information.

Bonsai 27B is the new multimodal flagship of the Bonsai family and the first 27B-class model to run on a phone. Based on Qwen3.6 27B, it brings a new capability tier to local devices: multi-step reasoning, structured tool calls, vision tasks, and computer-use agentic loops that stay coherent across many steps. Bonsai 27B comes in two variants. Ternary Bonsai 27B uses ternary weights with FP16 group-wise scaling, giving 1.71 effective bits per weight and a 5.9 GB footprint for the quality-oriented laptop-class version. 1-bit Bonsai 27B uses binary weights with the same group-wise scaling, giving 1.125 effective bits per weight and a 3.9 GB footprint that fits within the memory budget of an iPhone 17 Pro. Both variants run end-to-end across the language network, embeddings, attention, MLPs, and LM head with no higher-precision escape hatches. They are multimodal, with a compact 4-bit vision tower, so on-device workflows can understand screenshots, documents, and camera input.

CodeGemma is a collection of powerful, lightweight models that can perform a variety of coding tasks like fill-in-the-middle code completion, code generation, natural language understanding, mathematical reasoning, and instruction following. CodeGemma has 3 model variants, a 7B pre-trained variant that specializes in code completion and generation from code prefixes and/or suffixes, a 7B instruction-tuned variant for natural language-to-code chat and instruction following; and a state-of-the-art 2B pre-trained variant that provides up to 2x faster code completion. Complete lines, and functions, and even generate entire blocks of code, whether you're working locally or using Google Cloud resources. Trained on 500 billion tokens of primarily English language data from web documents, mathematics, and code, CodeGemma models generate code that's not only more syntactically correct but also semantically meaningful, reducing errors and debugging time.

OpenAI o3-mini is a lightweight version of the advanced o3 AI model, offering powerful reasoning capabilities in a more efficient and accessible package. Designed to break down complex instructions into smaller, manageable steps, o3-mini excels in coding tasks, competitive programming, and problem-solving in mathematics and science. This compact model provides the same high-level precision and logic as its larger counterpart but with reduced computational requirements, making it ideal for use in resource-constrained environments. With built-in deliberative alignment, o3-mini ensures safe, ethical, and context-aware decision-making, making it a versatile tool for developers, researchers, and businesses seeking a balance between performance and efficiency.

The o4-mini model is a compact and efficient version of the o3 model, released following the launch of GPT-4.1. It offers enhanced reasoning capabilities, with improved performance in tasks that require complex reasoning and problem-solving. The o4-mini is designed to meet the growing demand for advanced AI solutions, serving as a more efficient alternative while maintaining the capabilities of its predecessor. This model is part of OpenAI's strategy to refine and advance their AI technologies ahead of the anticipated GPT-5 launch.

Llama (Large Language Model Meta AI) is a state-of-the-art foundational large language model designed to help researchers advance their work in this subfield of AI. Smaller, more performant models such as Llama enable others in the research community who don’t have access to large amounts of infrastructure to study these models, further democratizing access in this important, fast-changing field. Training smaller foundation models like Llama is desirable in the large language model space because it requires far less computing power and resources to test new approaches, validate others’ work, and explore new use cases. Foundation models train on a large set of unlabeled data, which makes them ideal for fine-tuning for a variety of tasks. We are making Llama available at several sizes (7B, 13B, 33B, and 65B parameters) and also sharing a Llama model card that details how we built the model in keeping with our approach to Responsible AI practices.

OpenELM is an open-source language model family developed by Apple. It uses a layer-wise scaling strategy to efficiently allocate parameters within each layer of the transformer model, leading to enhanced accuracy compared to existing open language models of similar size. OpenELM is trained on publicly available datasets and achieves state-of-the-art performance for its size.

LTM-2-mini is a 100M token context model: LTM-2-mini. 100M tokens equals ~10 million lines of code or ~750 novels. For each decoded token, LTM-2-mini’s sequence-dimension algorithm is roughly 1000x cheaper than the attention mechanism in Llama 3.1 405B1 for a 100M token context window. The contrast in memory requirements is even larger – running Llama 3.1 405B with a 100M token context requires 638 H100s per user just to store a single 100M token KV cache.2 In contrast, LTM requires a small fraction of a single H100’s HBM per user for the same context.

The o3-mini-high model from OpenAI advances AI reasoning by refining deep problem-solving in coding, mathematics, and complex tasks. It features adaptive thinking time with adjustable reasoning modes (low, medium, high) to optimize performance based on task complexity. Outperforming the o1 series by 200 Elo points on Codeforces, it delivers high efficiency at a lower cost while maintaining speed and accuracy. As part of the o3 family, it pushes AI problem-solving boundaries while remaining accessible, offering a free tier and expanded limits for Plus subscribers.