English text is being prepared.
LatestUnsloth Releases 1-bit GGUF, Mistral Le Chonk Can Run Locally on Mac Studio
Why read thisUnsloth reports that the 1-bit GGUF quantization retains 91% top-1 accuracy, allowing readers to judge the feasible boundaries of running trillion-parameter models on local devices.
Liquid AI has released the LFM3 series of open-weight models, all branched from LFM3-1.7B-Base, covering text, vision, audio, Omni, and the d1 decision model. LFM3-1.7B-Base adopts a three-hybrid-layer architecture, is pretrained on 22T tokens, and in the mid-training stage over the final 3.5T tokens incorporates more agent trajectories and tool-calling data, with the context window extended to 256K tokens. The company says LFM3-1.7B is about 25% faster than LFM2-1.2B at an 8K context, and scores 38.42 versus 3.02 on WebArena-Verified; LFM3-VL-2B's average score across 62 vision benchmarks rises from 57.6 to 71.4. All models are available for download on Hugging Face, along with a single-GPU fine-tuning notebook.
Why read thisThe original article presents the architectural changes shared by the six models on a common backbone and on-device benchmark data, which can be used to assess the trade-offs of the small-model specialization approach on long-context and agentic tasks.
Google and XREAL have launched the XREAL AURA, the first wired XR glasses powered by Android XR. Orders open today in the United States, Canada, and South Korea, with sales in Japan starting November 2, at a starting price of $1,099. The glasses weigh less than 95 grams, use transparent lenses and a 70-degree field of view to project apps into the user's field of view, and offload heavy computing to a portable computing puck connected by cable; the puck is powered by the debut Qualcomm Snapdragon Reality Elite platform. Users can open multiple apps at once and pair them with a Bluetooth keyboard and mouse, or connect a laptop screen as a window via USB-C wired mirroring, and Gemini can provide assistance based on the content in the field of view and on the screen.
Why read thisThe original text explains the combination of transparent lenses, a 70-degree field of view, and a split-computing puck, allowing readers to judge the boundaries of everyday use for wired XR glasses.
Why read thisThe original text provides the accuracy retention ratio after quantization compression and the memory threshold, so readers can judge the feasibility of local deployment based on these figures.
Yanxin Matrix, under RockAI Technology, together with universities including HKUST and CMU, has released and open-sourced EdgeRock-4B. At a 256K context, its KV Cache is fixed at 72MiB, a 99.1% reduction compared with the source model Qwen3.5-4B. A single 80GB GPU can run 512 concurrent long-inference streams, with an output throughput of 12,124 tokens/s, 6.42 times that of the source model; the average pass@1 on AIME 2024–2026 is 82.9%. The model is derived from Qwen3.5-4B through staged conversion, requiring no pretraining from scratch; the weights and inference code have been open-sourced and are available for commercial use, supporting text and image inference.
Why read thisThe original text provides measured comparisons of fixed memory and throughput, based on which readers can assess changes in memory cost for long-inference scenarios.
Block's internal AI research lab released the full-duplex speech-to-speech model Dessa on October 23, stating that it was trained entirely from scratch on human conversational speech, can listen and speak simultaneously, and natively perceives emotion, sarcasm, and laughter. The model's inference speed is faster than real time, and it can run locally on a MacBook Pro M4 Max or higher configuration; the team completed training with a single NVIDIA GB200 NVL72. Block has published demo recordings and a public waitlist, and has opened early access to a small number of AI safety researchers.
Why read thisDessa was trained from scratch using only speech data, demonstrating how full-duplex conversation handles emotion and interruption timing, and also providing local running requirements.
Why read thisThe original text gives the model's positioning of running locally on 6GB of memory and its visual capabilities, so readers can judge the usable boundaries of small models in on-device assistant scenarios.
Recursive, founded by Tian Yuandong, has open-sourced its first model Flint, focused on local Agent coding, with 40 billion total parameters, 2.5 billion activated per token, and support for a 512k token context window. The company says that paired with its self-developed inference engine rec, Flint can reach at least 125 tokens/s on a MacBook Pro and over 150 tokens/s on RTX Spark, roughly 1.4–2.6 times higher throughput than llama.cpp. Flint scores 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2, and 54.9% on Terminal-Bench 2.1; rec currently supports M4 Pro, M5 Pro, and M5 Max, requiring at least 48GB of unified memory and about 30GB of free disk space.
Why read thisTian Yuandong's startup Recursive open-sources its first model, letting readers understand the trade-offs of local Agent coding models in speed and coding benchmarks.
Why read thisUnsloth provides the GGUF quantized version of Flint along with a local running guide, so readers can use it to actually deploy this 39.8B-parameter MoE model on a 24GB-memory device.
Google announced that Gemini is now available globally on Android Auto and in cars with Google built in, and has begun rolling out Personal Intelligence, which, with authorization, can connect to Workspace (including Gmail and Google Calendar), Google Photos, Google Maps, YouTube, and Search to answer questions such as flight gate numbers, restaurants along the route, and locations in photos. Personal Intelligence begins rolling out on Android Auto today, with cars with Google built in to follow over the coming months. It is currently available to eligible personal Google accounts for users aged 18 and over, excludes Workspace business, enterprise, and education users, and is not available in the European Economic Area, Switzerland, the United Kingdom, or Nigeria.
Why read thisThe original text explains how Personal Intelligence pulls information across Gmail, Photos, and Maps, so readers can judge the actual usage and privacy boundaries of the in-car assistant.
Why read thisPerplexity Computer runs both models and agents locally on the Mac, so readers can judge for themselves how sensitive data can stay on the device.
PyTorch 2.15 adds FP8 support to the MPS backend and makes MoE model decoding 5–6.5 times faster than 2.14: on an M3 Ultra with torch.compile, Qwen3.6-35B-A3B goes from 9.4 tokens/s to 61.3 tokens/s, and LFM2.5-8B-A1B goes from 23.6 to 123.6 tokens/s. For Wan2.2's tiled VAE video decoding, peak driver memory drops from about 270 GiB to 8 GiB, and the image corruption issue with Qwen-Image-2.1's VAE encoding on MPS is also fixed with a padding change, raising the encode-decode round-trip PSNR from 10.06 dB to 29.95 dB. This version requires macOS 15 or later and Python 3.11 or later.
Why read thisThe original article provides specific figures for the MoE decoding speedup and the reduction in video VAE memory, which can be used to gauge the actual changes when running models locally on a Mac.
Why read thisMeta has published the technical report for a 30B open-weight model, covering architecture, distillation recipe, and quantization, for readers interested in on-device agent deployment.
Why read thisUnsloth provided the accuracy retention ratio and size change after 1-bit quantization, so readers can judge the local deployment threshold accordingly.