On-device AI

AIGoogle's AI Edge Team open-sourced ML Drift under Apache 2.0, a GPU compute engine for on-device AI inference across OpenGL ES, OpenCL, Metal, and WebGPU. It serves as the core GPU acceleration engine within LiteRT and succeeds the legacy TFLite GPU delegate, which will no longer receive new features. The post cites benchmarks showing up to 40% lower frame latency in YouTube Shorts and up to 30% faster on-device performance in Adobe Lightroom and Photoshop.

Why it matters: The post explains how ML Drift unifies GPU shaders across platforms and replaces the TFLite GPU delegate, which matters for developers deploying on-device models.

AINVIDIA and Microsoft announced RTX Spark laptops and compact desktops that run the full NVIDIA AI stack locally, with laptop preorders open today and sales from October 16. Microsoft also announced general availability of Microsoft Execution Containers (MXC), an OS-level infrastructure for agents to run securely in the background, while NVIDIA previewed DGX Station for Windows with 748GB of coherent memory and up to 20 petaFLOPS of FP4 compute.

Why it matters: The announcement pairs Windows agent infrastructure with local hardware, showing how agents may move onto personal computers and enterprise desktops rather than only cloud services.

AILiquid AI released two open-weight d1 decision models, d1-3B and d1-omni-600M, on Hugging Face. d1-3B scores 48.57 on the Decision Index v0.2.1 public split and answers a single question in 8 ms on an NVIDIA GeForce RTX 4090 and 50 ms on a Jetson Orin Nano. d1-omni-600M is an experimental checkpoint that handles text with images or audio and scores 15.95 on the same index.

Why it matters: The release pairs open-weight decision models with measured latency across Apple, NVIDIA, and Jetson hardware, showing how edge deployment changes what is practical.

AIGoogle DeepMind has released EmbeddingGemma 2, an open 740 million parameter model that maps text, images, audio, and video into one embedding space. It is built on the Gemma 4 architecture under an Apache 2.0 license and supports an 8K token context window. The company reports a code benchmark gain from 68.76 to 78.68 on MTEB Code and says the model can run on-device with about 567MB of active RAM for the full multimodal version on a Google Pixel 11 Pro.

Why it matters: The release shows how a 740M-parameter embedding model can cover text, code, images, audio, and video on local hardware, with memory and storage figures to compare against other on-device options.

AIGoogle DeepMind has released EmbeddingGemma 2, a 740-million-parameter embedding model that maps text, images, audio, and video into a shared space and runs on local hardware under an Apache 2.0 license. Matryoshka Representation Learning lets developers truncate output vectors from 768 dimensions to 512, 256, or 128, and the model supports an 8K-token context window. The model weights are available on Hugging Face and Kaggle, with Gemini Enterprise Agent Platform availability coming soon.

Why it matters: The release shows how a 740M-parameter multimodal embedder runs locally with a 768-to-128 dimension truncation option, useful for judging on-device retrieval designs.

AIGoogle DeepMind launched EmbeddingGemma 2, an open-weight 740M parameter model that maps text, images, video frames, and audio into one vector space. The model can run on-device, with about 567MB active RAM for the full multimodal model on a Google Pixel 11 Pro, and is available through Google AI Edge Gallery, Google AI Edge Foresight on Mac, and MediaPipe Tasks, with ML Kit support coming in the weeks ahead.

Why it matters: The post names concrete on-device apps, memory footprints, and latency figures, showing how a multimodal embedding model can power local search without cloud calls.

AIArtificial Analysis has open-sourced AA-AgentPerf-Local, a tool that replays recorded agent trajectories to measure inference speed on laptops and workstations. Initial results cover NVIDIA DGX Spark, NVIDIA GeForce RTX 5090, AMD Ryzen AI Halo, and MacBook Pro M5 Pro, with the RTX 5090 fastest for models that fit its 32 GB. The source states the tool and leaderboard will expand to more hardware, frameworks, and models.

Why it matters: The source gives per-system completion times and memory bandwidth figures, letting readers compare local hardware for running agentic workloads.

AIGoogle announced that the Antigravity SDK supports local agent workflows, with initial support for Gemma 4 26B A4B through Google AI Edge's LiteRT. The post includes Python setup steps and says a recommended machine has more than 24GB VRAM or unified memory. It also describes a hybrid pattern in which a cloud Gemini 3.8 Flash planner hands work to local Gemma 4 26B models, with 97.2% of tokens in one recorded run staying local.

Why it matters: The source shows how to run an agent with a local Gemma 4 26B model using LiteRT, plus a hybrid cloud-planner pattern that keeps most tokens on-device.

AIUnsloth says the 7B Qwen-Image-2.1 text-to-image and editing model can run locally on 12GB VRAM using its GGUF builds. It also states that the model performs on par with Nano Banana 2.0, and that Dynamic FP8 can run on 6GB of VRAM via offloading for higher quality. The image lists int8 at 7.26 GB with mean LPIPS 0.064 and fp8 at 7.12 GB with mean LPIPS 0.112, and says int8 is the default.

Why it matters: The post gives concrete local-run settings, VRAM figures, and GGUF and FP8 options, which helps readers judge whether the model fits their hardware.

AIOpenBMB has released MiniCPM5-2B, a dense 2B Transformer built for on-device and resource-constrained deployment, with an average score of 53.9 in its comparison set. The release also opens the UltraData datasets behind it, including UltraX, UltraData-Code, UltraData-SFT-Agent-2609 and UltraData-RL-2609, and includes GGUF, MLX, GPTQ and DSpark variants for common runtimes.

Why it matters: The release pairs a 2B model with open training datasets and reports per-benchmark comparisons against named same-size and larger models, letting readers check the claims directly.

AIUnsloth AI published a guide for running GLM-5.3 locally using quantized GGUF weights. The 2-bit version is reduced from 1.51TB to 239GB and retains about 81% accuracy, and it can run on a 256GB Mac or RAM/VRAM setups.

Why it matters: The guide shows which quantization levels fit local memory budgets and how much accuracy each costs, useful for planning a local deployment.

AIOpenBMB released MiniCPM5-2B-SFT, an SFT-only BF16 checkpoint taken before RL and OPD, within its MiniCPM5-2B series. The model is a 2B dense Transformer built for on-device and local deployment, with 131,072-token context and the same training recipe as the final release.

Why it matters: The source gives concrete benchmark averages against same-size and larger models, plus released training data and multiple deployment formats, useful for judging a compact on-device model.

AILiquid AI released DSpark speculative decoding draft models for LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B on Hugging Face. The draft models reach up to 3.18x throughput improvement on an H100 GPU and up to 2.87x on-device, and the outputs match baseline greedy decoding by construction. Support is available in llama.cpp and SGLang, with the speedup varying by model and dataset.

Why it matters: The release reports measured speedups on both H100 and MacBook hardware, with per-dataset results and acceptance rates that show where speculative decoding helps most.

AILiquid AI released 4-bit Q4_0 GGUF checkpoints for LFM2.5-230M, LFM2.5-350M, LFM2.5-1.2B-Instruct, and LFM2.5-2.6B, trained with Quantization-Aware Distillation. The company says the checkpoints recover most accuracy lost to quantization, reaching roughly 97% of their BF16 averages while keeping Q4_0 memory footprint and throughput. Benchmarks compare them against post-training quantized Q4_0 GGUFs and against Q5_K_M, Q4_K_M, and Unsloth's UD-Q4_K_XL.

Why it matters: The post shows how quantization-aware distillation recovers accuracy lost in Q4_0 checkpoints, with throughput measured across four hardware backends for deployment tradeoffs.

AILiquid AI released LFM2.5-VL-3B, an open-weight 3B vision-language model that it says rivals models twice its size while running faster on CPU and GPU. Benchmarks show large gains over LFM2-VL-3B, including ScreenSpot-v2 averaging 80.7, RefCOCO precision@1 rising from 57.1 to 87.9, and ToolSandbox rising from 26.4 to 59.5. The model is available on Hugging Face and decodes 228 tokens/s on an Apple M5 Max.

Why it matters: The post pairs benchmark gains with on-device and GPU throughput figures, showing how a 3B vision model trades size against speed and accuracy.

AILiquid AI released LFM2.5-2.6B, a 2.6B-parameter agentic model that runs on-device on phones and CPUs, along with a base variant on Hugging Face. The company reports it leads on every instruction-following benchmark and nearly every tool-use benchmark it tested, and decodes 220 tokens/s on an M5 Max. The source says larger models may still suit complex agentic or coding-heavy tasks.

Why it matters: The source reports benchmark results against several same-tier models and notes where larger models still lead, which helps judge fit for edge agent workloads.

AIMistral AI released Voxtral Mini 4B Realtime 2602, a multilingual realtime speech-transcription model with 13 supported languages under the Apache 2.0 license. The model has a configurable transcription delay from 240ms to 2.4s, and it matches leading offline open-source models at a 480ms delay. The source says it is optimized for on-device deployment and is currently supported only in vLLM.

Why it matters: The source specifies the 480ms delay operating point, 4B size, Apache 2.0 license, and vLLM serving path, which matter for teams weighing realtime transcription deployment.