Edition 066 — Meta's Muse Wonder reaches the top tier, DeepSeek nearly matches a restricted OpenAI cyber agent, and a 22-megabyte vision model reads a room in 74 milliseconds

In this edition

Frontier & Text Models (4)
Recursive Self-Improvement — Research (1)
Video & World Models (2)
Audio, Voice & Music (2)
Robots & Embodied AI (1)
Small & On-Device Models (2)
Papers & Research (5)

Meta's Muse Wonder Full lands in the top three on six of ten benchmarks, a $3 open-weight model comes within eight points of OpenAI's restricted cyber agent, and a 51-million-parameter vision model describes a scene on a phone in 74 milliseconds. Seventeen items today: two frontier releases and a training-data fight, four audio and video systems that run in real time, robot agents that improve themselves in the physical world, three models small enough to ship on consumer hardware, and three benchmarks that make last year's scores look generous.

Frontier & Text Models

Meta Muse Wonder Full benchmark results

Meta's Muse Wonder Full enters the AI top tier with a 56 on the Artificial Analysis Intelligence Index v4.3, up from Muse Spark 1.3's 48. Terminal-Bench 4.0 performance jumped from 33 to 57 percent, and the model places in the top three on six of ten benchmarks. Both Muse Wonder and Muse Wonder Full support a one-million-token context window and accept text, images, video and audio. Hallucination rate fell from 33 to 23 percent; only Gemini 4 Argon at 15 percent is lower among models scoring at least 45 index points. On the Coding Agents Index, Muse Wonder Full hits 68, tying Claude Code with Sonnet 5.5. New features include Active Memory and Active Focus; Muse Wonder is available via Meta's API and Muse Code, while Muse Wonder Full is limited to Muse Code users.

OpenAI logo

Starting November 29, OpenAI will train its models on Daybreak Blue session data that was previously off-limits. An email to Trusted Access program members confirmed the change, which applies even when users had "improve the model for everyone" turned off. Following backlash, OpenAI sent a second email promising a separate opt-out; interactions flagged during safety reviews can still be used for safety purposes. Researchers cite the Buckmaster-Alpöge case, where mathematicians uploaded Navier-Stokes drafts to Codex and OpenAI soon presented its own proof while denying it used their inputs for training.

Vals AI CyberOffense benchmark

On the Vals CyberOffense benchmark, OpenAI's restricted Daybreak agent GPT-6 Sol reached 80.8 percent of objectives while open-weight DeepSeek V4.1 Flash 1017 hit 72.8 percent at roughly $3 per run instead of $22. Each agent works solo on an isolated Kali Linux machine with only a public web address and no human help, advancing from the web application through Windows servers to admin privileges. In the harder Lighthouse test, Daybreak took admin control of the first company, stole passwords and broke into a second, completing 27 of 31 objectives in under three hours; DeepSeek managed 21 at a tenth of the cost. DeepSeek scored zero in its first 25 million tokens and made most progress only after 100 million.

A frontier LLM called Astra directly edited the weights of Qwen3.5-0.8B with no gradient descent and no training, in an experiment published as "Sparks of Weight-Space Programming." Working in a Python session with a budget of 50 forward passes on GSM8K and MATH training problems, Astra traced the base model's thinking loops to one direction written by the MLPs of layers 15 and 19, then subtracted it from 64 output columns of down_proj in each layer: 131,072 weights, 0.016 percent of the model. MATH-500 rose from 7.9 to 48.1, GSM8K from 13.8 to 67.1, IFEval from 21.0 to 62.3, and the share of answers hitting the token limit fell from 93.0 to 33.1 percent. An independent re-run of the checkpoint scored 51.0 percent on MATH-500 with 31.6 percent hitting the token limit.

Recursive Self-Improvement — Research

Coding agents like GPT-6-Astra now drive robot behavior through embodied recursive self-improvement, reaching a 91.1 percent average success rate across eight real-world tasks. PyRUA-REAL lets agents retain interaction experience as persistent memory and reusable executable skills that inform later exploration, code refinement and deployment. On 42 RoboDojo benchmark tasks the same framework averages 84.4 percent. The study also analyzes efficiency when the resulting programs run without an online agent making decisions.

Read more → https://arxiv.org/abs/2611.06768

Video & World Models

NetProjects published an MLX port of NVIDIA's Cosmos3-Edge world model for Apple Silicon, keeping bf16 weights unchanged with the timestep embedder in fp32. Builds are 7.6 GB in bf16 and 5.0 GB in 8-bit; with the 8-bit build, NVIDIA's full image-to-video example at 832x480, 121 frames, 24 fps and 20 steps takes 113 seconds end to end with an 18.8 GB peak. Latent cosine similarity against NVIDIA's fp32 pipeline from the same initial noise is 0.9960 for image-to-video and 0.9959 for text-to-video in 8-bit, and 0.9975 for both in bf16. Benchmarks ran on a MacBook Pro M5 Max with 128 GB on 7 November 2026. Cosmos3 is NVIDIA's omnimodal world-model family taking text, images, video and action trajectories as input.

EvoWorld is a video generation framework that pairs a world model with a renderer so a scene can keep evolving causally, even out of sight. The paper frames the gap as Out-of-Sight Causal Evolution: a model must render an event, reason about its consequences for world state, and reveal a causally plausible outcome as the world continues to evolve. EvoWorld is a brain-renderer collaboration where a vision-language model acts as the world brain, observing generated segments, maintaining an evolving account of the world and reasoning about state transitions, while a segment-autoregressive video renderer realizes them. The authors introduce EvoWorld-Data and EvoWorld-Bench for event realization, causal reactions and revisit consistency, and report substantial improvement over recent models while preserving visual quality and consistency.

Read more → https://arxiv.org/abs/2611.06748

Audio, Voice & Music

Fish Audio Drama 3 is a 35-billion-parameter text-to-speech model with 3 billion active per token, built for multi-character scenes and long-form content. The MoE text backbone uses hybrid full and linear attention, and a Codebook-MTP audio decoder decodes 10 codebooks causally over the whole sequence; MoE experts run in FP8 while the rest stays bf16. Speaker tags allow up to 16 characters in one text, and long text is chunked automatically and streamable per chunk. Measured throughput hits 166.7 frames per second at concurrency 1 with 88.9 ms median time to first audio, 927.6 at concurrency 16, and 1,687.3 at concurrency 64. It ran on SGLang-Omni and vLLM-Omni from launch day.

Read more → https://huggingface.co/fishaudio/drama-3

LiveFoley is a streaming audio generation model that turns video into synchronized sound effects with look-ahead below the human perceptual threshold, matching offline models in temporal alignment while leading streaming systems in quality. A Trailing Sliding Window converts non-causal segment-level encoders into streaming modules without retraining, and a configurable hop size reduces frontend look-ahead below the human perceptual threshold. A Unified Varhop Framework adapts to different latency and compute budgets, while History LocDiT adds past context at no latency cost. The authors report LiveFoley preserves audio-visual alignment that streaming methods typically degrade when shrinking receptive fields to cut look-ahead latency.

Read more → https://arxiv.org/abs/2611.05988

Robots & Embodied AI

BEAKER is the first benchmark for embodied robots in self-driving chemical laboratories, scoring actionability, knowledge, experimentation and reasoning. BEAKER comprises 1,000 expert-curated question-answer samples across 11 image and video task types, all manually created and reviewed by chemistry and embodied AI experts with no AI-assisted generation. Evaluations of 28 general-purpose, science-oriented, chemistry-specialized and embodied MLLMs show models can follow coarse experimental procedures in videos but struggle to ground actions into operable regions and feasible trajectories, recognize fine-grained apparatus, states and safety norms, and order steps in mid-length procedures.

Read more → https://arxiv.org/abs/2611.06911

Small & On-Device Models

Mirage is a 51.3-million-parameter on-device vision model that fits in about 22 MB using 2-bit quantization, bringing multimodal vision to phones, wearables and robots. It runs end to end in 74 ms on a Samsung Galaxy S26 NPU and 107 ms on its CPU, describing scenes, detecting objects, reading text, generating segmentation masks, identifying spatial relations and making decisions from user-supplied labels in a single pass. It describes images in English, German, French, Spanish, Italian, Dutch and Polish, and runs in the same binaries as Cactus models Needle and Whistle.

Read more → https://cactuscompute.com/blog/mirage

Whistle is a 16.9 MB edge model that runs on-device speech recognition, word timestamps and speech embeddings in seven languages. It supports English, German, French, Spanish, Italian, Dutch and Polish, and runs through the same cactus-compute runtime as Cactus's other on-device models. The repository ships WebAssembly-compatible builds, targeting phones and other consumer hardware for fully local speech processing.

Papers & Research

VoxCanvas-41M is a large-scale bilingual dataset of roughly 41 million speech editing pairs covering content and paralinguistic editing. Content editing includes single-span, multi-span and location-reference tasks, while paralinguistic editing covers insertion and deletion of 39 classes of nonverbal events such as laughter, sighs and coughs. The authors train VoxCanvas-Edit, an instruction-based speech editing model, and study how data scale, task interactions and training strategies affect editing performance. The work addresses the lack of large, high-quality datasets and the poorly understood learning dynamics of speech editing models.

Read more → https://arxiv.org/abs/2611.08251

NegWAM is a safety post-training framework for autonomous driving world-action models that learns explicit safety boundaries from near-expert failure trajectories instead of copying expert driving alone. It builds negative trajectories that sit geometrically close to the expert path but carry safety risks, uses them as repulsive anchors to steer predictions away from dangerous behaviour, and encodes them as scene-conditioned rider tokens to learn scene-trajectory compatibility. Training jointly optimises the flow-matching and near-expert negative objectives, and the auxiliary rider branch is discarded at inference, adding no overhead. On NAVSIM v2 it reaches 40.4 EPDMS on navhard, a 4.1 percent relative improvement over the strongest scorer-free baseline, and 93.3 EPDMS on navtest, with zero-shot closed-loop evaluation on HUGSIM.

Read more → https://arxiv.org/abs/2611.08111

Tencent's DSpine is a five-layer speculative-decoding draft model for Qwen3.8-27B that lifts acceptance length and end-to-end throughput across reasoning and code benchmarks. With block size 16 and an 8,192-token sliding window, it generates draft tokens in parallel while passing predicted predecessor features to adjacent positions to carry causal dependencies through depth. Against the official DFlash2 draft at greedy decoding, acceptance length rises on GSM8K from 5.69 to 7.20, on MATH-500 from 5.49 to 6.49, on HumanEval from 4.87 to 5.71 and on MBPP from 5.04 to 6.14. End-to-end throughput on WorkBuddy Code goes from 253.35 tokens per second autoregressive to 686.23 with DSpine, a 2.71x speedup, reaching 3.18x at high reasoning effort.

Co-Design Bench is a reference-free benchmark of 273 tasks curated from 37,463 deduplicated queries across 26,314 real user coding sessions. Its two tracks, Multimodal Editing and Repository Editing, cover localized edits with optional image, audio, video or document inputs and coordinated repository revisions, with 2,331 hidden expert-authored acceptance criteria. A judge-then-refute protocol grades criteria using source and browser evidence and challenges positive verdicts with a separate LLM verifier. Under a shared agentic harness, the strongest of ten model endpoints passes an average of 69.9 percent of criteria per task but fully satisfies only 26.4 percent of tasks; all ten do worse on repository edits spanning three or four change types.

Read more → https://arxiv.org/abs/2610.35138

TB-MathModel turns complete contest-style mathematical modeling problems into agentic, verifiable, terminal tasks over 18 real CUMCM and MCM problems from 2023 to 2025. Each task ships the verbatim statement, official attachments and figures in a containerized terminal, and the agent returns one JSON artifact scored on the numeric core of the final question. Two-layer feasibility verification combines a programmatic gate over six decidable invariant forms with 624 Lean 4 obligations, and a hard gate zeroes infeasible answers while direction-aware one-sided scoring never penalises beating the anchor. Across seven multimodal entries the strongest reaches a multi-reference O-Eval of 64.58 percent, while a $2.54 open model holds 59.81 percent at one hundredth of the cost.

Read more → https://arxiv.org/abs/2611.08237

After Intelligence is published every morning. Seventeen items, every claim traceable to its source.