Meta has released the weights for Muse Spark Open, an 816-billion-parameter model that scores 39 on the Artificial Analysis Intelligence Index, leading US open-weight models. However, among the nine highest-scoring open models, it ranks last on four agent benchmarks. Microsoft has unveiled MAI-Base-2, with 2.8 trillion total parameters. Scaling-law estimates suggest it is 4.2–12.3 times as compute-efficient as DeepSeek-V4-Pro-Base at matching the latter's held-out performance; the model is still undergoing post-training. TypeSafe has raised $870 million in Series A funding at a $7.5 billion valuation. Its Jev model does not generate text, but instead makes calibrated decisions within software.
My take
My biggest takeaway today isn’t a leaderboard win—it’s the gap between high scores and execution. Muse Spark Open leads US open-weight models on the intelligence index, yet finishes last on four agent benchmarks among the nine highest-scoring open models; that makes task-specific performance matter more to me than the overall ranking.
Meta and Black Forest Labs release model weights, while Microsoft discloses foundation-model training efficiency estimates.
Meta has released the weights for Muse Spark Open, with 816 billion total parameters, 42 billion active parameters, and a one-million-token context window. It scores 39 on the Artificial Analysis Intelligence Index, leading US open-weight models, but ranks last on four agent benchmarks among the nine highest-scoring open models. The license requires attribution, and model-as-a-service companies with annual revenue above $20 million must sign a separate agreement.
Read original →Microsoft has completed pretraining for MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active parameters. Based on scaling-law estimates, Microsoft says its training recipe is 4.2–12.3 times as compute-efficient as DeepSeek-V4-Pro-Base at matching the latter's held-out performance. The coding evaluation in this comparison uses MAI's own codebase. The model is still undergoing post-training.
Read original →Black Forest Labs has released FLUX 3 Image Klein, a 5.6-billion-parameter open-weight image model. It supports native 4K output, bounding-box control, and pixel-precise editing, and can run on a single consumer GPU. The company says it combines the image quality of large frontier models with customizability and speed.
Read original →🔍 Analysis: Muse Spark Open's overall score contrasts with its agent benchmark rankings. MAI-Base-2's claimed 4.2–12.3-fold efficiency comes from scaling-law estimates, and its coding evaluation uses Microsoft's own codebase. Neither directly guarantees application performance. FLUX 3 Image Klein brings another selection criterion to the fore: whether a model can run on a single consumer GPU. For developers, these three models call for separate scrutiny of task performance, the basis of training comparisons, and running costs. A single leaderboard position or efficiency multiple cannot substitute for model selection.
Behind its $870 million Series A, Jev focuses on decision-making capability per unit of cost.
TypeSafe has raised $870 million in Series A funding at a $7.5 billion valuation. Its decision model, Jev, does not generate text; instead, it makes fast, low-cost, calibrated decisions within software. CEO Diogo Almeida introduced RLCD (reinforcement learning for calibrated decisions), saying he focuses less on growth metrics such as user counts and more on intelligence per unit of cost—and on whether developers can build applications they previously abandoned because AI was too expensive.
Read original →Sakana orchestrates parallel research workflows, while Perplexity provides isolated sandboxes for task execution.
Perplexity has open-sourced its SPACE agent sandbox platform under the Apache 2.0 license. Each task runs in a separate Firecracker microVM, with credentials and network egress controlled outside the VM. Across 108 escape attempts in which models had root privileges inside the VM, none reached the host. Perplexity also uses SPACE internally for reinforcement learning and evaluation tasks.
Read original →Sakana has introduced Agentic Scientist (AI Scientist v3), which autonomously improved Doc-to-LoRA's multi-needle retrieval score from 52% to 98%. The system uses a meta-controller to coordinate parallel coding agents, combining preregistered scoring criteria, shared memory, and automated ablation experiments to organize research workflows. It also supports human–AI collaboration: a paper produced with contributions from a development version under human guidance was accepted at ALIFE 2026.
Read original →🔍 Analysis: Sakana uses a meta-controller, shared memory, and automated ablation experiments to organize parallel coding. SPACE puts task execution inside isolated microVMs, with credentials and network egress controlled outside the VM. The two address separate challenges in autonomous execution: how to organize work and how to constrain execution privileges. For enterprises, better research results cannot substitute for runtime isolation. Likewise, the results of SPACE's 108 escape attempts provide test evidence for its isolation mechanisms, not a substitute for evaluating task performance.
Goodfire and Neel Nanda debate neuron interpretability, while METR gains evaluation access approaching that of Anthropic's internal team.
Goodfire's Nick Cammarata argues in a post that most neurons in trillion-parameter-scale models are interpretable. He says some classic examples of polysemanticity yield clear explanations after further analysis by agents. By contrast, the 2023 OpenAI automated interpretability work he contributed to found only about 2% to be interpretable. Neel Nanda says this does not change his core view: he still expects representational superposition to exist and weak activations to carry memorized facts.
Read original →Anthropic has given six METR staff members access to internal tools and systems approaching the level available to its own risk evaluation team. This embedded evaluation goes beyond model access, allowing evaluators to understand the training process and how safety measures work. METR can publish independent risk analyses, and Anthropic cannot veto publication. It can, however, request redactions of predefined categories of information, such as customer data and model architecture; METR can publicly comment on whether those redactions are justified.
Read original →GPT-Live-1.1 adds voices in six languages, with the user rollout underway and API support to follow.
OpenAI has released a GPT-Live-1.1 voice update for ChatGPT, Codex, and dots. It adds voices in French, German, Hindi, Mexican Spanish, Korean, and Japanese, while improving accents and reducing filler phrases such as “checking…”. The update has begun rolling out to all users, with API support to follow.
Read original →