Meta released open weights for its Muse Spark AI model, shifting it from a proprietary system to Muse Spark Open designed for coding and long-running agent work. The model has 816 billion parameters, with 42 billion active per token, and it supports a context window of up to 1 million tokens. The release adds a license with limited commercial terms and changes usage by enabling free downloads/adaptation for developers (with restrictions on hosted resale, and on military/critical-infrastructure uses), while keeping Meta’s top systems closed.
Meta launched Muse Spark Open, an open-weight AI model designed to run on Macs, and paired it with Muse Code for coding and other agentic workflows. The model uses 816 billion parameters and a 1 million token context window. Running it locally requires at least 1 terabyte of RAM, and Meta is making the model available now through multiple API platforms for desktop-focused coding and tasks.
NVIDIA released Nemotron 3.5 Super VL, an open multimodal agent vision-language model that can read text, images, documents, and MP4 video clips, then reason, call tools, and write code. It has 120B total parameters with 12B activated per token and supports up to 1M tokens of context. The release consolidates text and vision checkpoints into one model and adds separate edge variants (Nemotron 3.5 Bolt 74B-A9B and Bolt E4B) plus open weights, datasets, training recipes, and deployment options.
Nvidia launched customizable Nemotron 3.5 Super LV and agent-focused Nemotron 3.5 Bolt LLMs for always-on AI agents, with an additional smaller Bolt E4B planned later this month. Super VL is an open 120-billion-parameter model trained on about 26 trillion tokens. Organizations can download the open weights to run and post-train models behind their own firewalls, while Bolt models enable local and edge deployment within 24 GB or as little as 4 GB.
Nvidia announced new Nemotron 3.5 models, including a 120B vision-language model and two smaller Bolt models meant for local use. Nemotron 3.5 Bolt 74B-A9B is quantized to fit in 24GB of memory. The launches add multimodal enterprise agent capability plus downloadable local models on Hugging Face under an OpenMDW license for commercial use and modification.
NVIDIA released Nemotron 3.5 Super VL plus two new edge variants, Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB, to expand its open Nemotron family for agent deployments. Super VL is a 120-billion-parameter model that NVIDIA says is 5x faster than the next best open model in its class for high-volume agentic workflows. Organizations can now pick an open model that matches their deployment tier, run it on their own infrastructure with post-training control, and use NVIDIA’s production packaging (NIM) to serve more users with lower latency claims.
An early Anthropic backer predicts Anthropic’s IPO will price below its reported $2 trillion target but then jump after trading begins. They expect the IPO shares to be priced around $1.8 trillion, with a trading value of $2.3 to $2.4 trillion versus SpaceX’s about $2.1 trillion opening-day market cap. The story frames Anthropic’s perceived strength in foundation models and enterprise focus as driving a valuation that the investor says will justify earlier profitability than SpaceX.
Mistral released Timestral, its first family of time-series foundation models, and reports top results on major public forecasting benchmarks with probabilistic, multivariate forecasts. The largest model has 209M parameters and ranks first on GIFT-Eval, TIME, and fev-bench. The release introduces four model sizes (5M to 209M) designed for accuracy/efficiency tradeoffs, CPU inference, and zero-shot forecasting, with optional fine-tuning through Mistral Forge and code made available on GitHub.
Apple received final regulatory approval to launch Apple Intelligence in mainland China. The green light came from China’s Ministry of Industry and Information Technology in time for the next week launch of the iPhone Duo. The rollout in China uses Alibaba’s Tongyi (Qwen language models) and Tongyi Wanxiang for image/video, with answers subject to China’s content rules.
OpenAI introduced GPT-6.1 Sol and its Dots autonomous agent during DevDay, while Google announced Gemini 4 Argon for long-context, security-focused work. GPT-6.1 Sol costs $2 per million input tokens via API (and $10 per million output tokens) and is positioned at one-fifth the per-task cost of GPT-6 Astra. The changes shift agent workflows toward higher tool-call volume and longer computations, with tighter availability tiers (e.g., Dots rollout to specific ChatGPT plans; Gemini 4 Argon limited to Fairwind program users) and ongoing delays/scraps tied to security concerns.
NVIDIA launched Nemotron 3.5 Super VL, an open multimodal model for specialized agents. It is a 120B mixture-of-experts model with 12B active parameters. NVIDIA says it can deliver up to 5x higher agentic throughput on a single GPU than open models of similar size, with checkpoints and training materials provided on Hugging Face.
TypeSafe AI released Jev 1.14, an update to its System One typed decision model that is live under the jev-preview alias with new experimental capabilities. The default context is 64k tokens per request and 32k per question, with an opt-in header that raises this to 262k tokens. Developers can now enable experimental image input and a native Box bounding-box question type, but TypeSafe says none of the three new danger-zone features are ready for production.
Tool-augmented LLM agents benchmark two tool-discovery architectures—Retrieve-then-Plan and Plan-then-Retrieve—on ToolBench to see which finds the right APIs before execution. PtR beats RtP on multi-tool queries across retrieval budgets k=1 to 20 and registry sizes from 1k to 47k APIs. The results imply complexity-adaptive routing (potentially using RtP’s top-1 retrieval score) to choose between the two pipelines.
Apple Machine Learning Research·2 days ago·
26
● 2 sources
MLX with RDMA-over-Thunderbolt support on Mac Studio was used to run and benchmark local and distributed inference for large language models, including Muse Spark Open, using the latest MLX release and benchmark tooling. The setup reportedly generated over 36 tokens per second on a cluster of four Mac Studios. As a result, very large MoE models can be tested on private hardware either on a single Mac with enough unified memory or by scaling out across multiple Macs via pipeline or tensor parallelization.
Black Forest Labs released FLUX.3 Image Klein and made it available in Diffusers from the start. The distilled checkpoint generates a 1024 × 1024 image in about 2 seconds on an H100. The pipeline now supports a single model for text-to-image, multi-reference image-to-image, and bounding-box layout control, along with distilled, base, and quantized variants plus LoRA training and editing workflows.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
