Muse Wonder ships without scores, Anthropic tests self-copying

Meta launched a frontier model without publishing a score. Anthropic published a pile of scores and then told you how easy the test was.

Muse Wonder, from Meta

Meta announced Muse Wonder on October 12 as a model for personal and professional agents. The pitch rests on two features. Active Memory tracks context across conversations, and Active Focus keeps the model working toward a long goal. You can reach it through the Meta Model API and Muse Code, with a Muse Wonder Full tier for subscribers. There are no open weights and no parameter count. Meta claims state-of-the-art results and cites Harvey's Legal Agent Benchmark, FocusBench and SWE-Marathon, but none of those come with a number. Until someone reproduces one, I read this as a product launch.

Anthropic on self-replication

The Frontier Red Team asked whether small models can copy themselves across cheap IoT devices with no datacenter. A 9B model ran inference on 8 of 9 hardware types. GLM-5.2 reached all 8 nodes in 4 of 5 trials. Gemma4-26B, spread over 15 smart TV boxes at about $55 each, managed 0.39 tokens per second. Setup is slow: 28 to 51 hours for Qwen3.5-9B and 71 to 150 hours for Gemma4-26B-A4B.

The number I keep rereading is the one on inherited notes. Success rose from 45% to between 61% and 84% across six generations when each model got notes from the one before. Anthropic projects small-model replication capability by the end of 2026. It also says the test used unauthenticated ADB ports, that a cycle takes days, and that its capability metric is only a proxy.

Open weights and images

Qwen-Image-2.1-Turbo is an 8-step checkpoint of the 7B Qwen-Image-2.1, for text-to-image and editing, and it loads directly in Diffusers. OpenBMB's MiniCPM-V 4.7 35B-A3B is the first mixture-of-experts model in its series. The team claims 13+ tokens per second in 3 GB on an iPhone 17 Pro Max, and a higher overall score than Qwen3.6 35B-A3B in less time. Both are vendor claims. vLLM v0.31.0 landed with 717 commits from 307 contributors, plus a preload command that keeps weights resident across engine restarts.

Inference and agents

NVIDIA and Princeton's Dao AI Lab published ReplaySSM, which stores state-space layer state as a checkpoint plus a buffer of recent inputs. Decoding kernels ran 1.46x to 1.80x faster at batch size 256 on B300 GPUs, and speculative throughput rose 1.85x to 2.10x on 120B-class models. Three of four models stayed within about 2 points of FP32 on GPQA-Diamond.

Codex 0.162.0 adds managed Git worktree tools and task pinning.

JetBrains studied 4,782 real IDE agent sessions and found only 33% had three or more user messages. In a 700-task SWE-Bench Pro pilot, solving in several steps roughly doubled cost with no stable change in resolve rate.

On the robotics side, Reflex catches thrown boxes from egocentric RGB-D. On a real Unitree G1 it succeeds 65% of the time at 2.5 m and 40% at 5 m.

What I'd watch next

Whether Meta publishes any Muse Wonder score on FocusBench or SWE-Marathon. Whether the Anthropic team reruns the test against devices that have real authentication. And whether MiniCPM's phone speed holds up on hardware that isn't the team's own.