AI Frontier Daily · October 13, 2026

Daily global AI trends digest, caught up in five minutes.


Headlines

1. Meta Ships Its First Frontier Models: Muse Wonder and Muse Wonder Full, With ‘Active Memory’ That Beats the Field

Meta’s Superintelligence Lab released a pair of new state-of-the-art models on Monday — Muse Wonder and Muse Wonder Full — representing the company’s first genuine shot at the frontier. Where Muse Spark aimed to be a lightweight, low-latency model, the Wonder models are larger and markedly more capable, with stronger benchmark scores and better results on complex tasks like coding. According to Meta’s model card, Wonder and Wonder Full achieve benchmark scores comparable to Fable 5.1 and GPT-6 Astra; the company calls them “the next step on our scaling ladder and a major leap toward personal superintelligence.” Pricing runs $4 per million input tokens and $20 per million output tokens through the Meta Model API — roughly equivalent to GPT-6 Sol or Opus 5.5 — and notably, Wonder Full will only be available to Muse Code subscribers, not through the API. Both models will also power the Muse agent. Staying true to Meta’s long-horizon personal-assistant ambitions, the release puts special emphasis on “active memory”: the ability to retain and synthesize user information across days, weeks, and months. On Meta’s new MemoryBench benchmark, Wonder scores higher than Fable, Opus, or Astra. This is Meta’s first frontier-class release since Scale AI founder Alexandr Wang restructured its AI efforts in August 2025, and both models were trained at Meta’s Prometheus data center in central Ohio, which came online earlier this year.

Source: TechCrunch

2. GPT-6 Astra Teams Up With an Amateur Mathematician to Prove a 57-Year-Old ‘Circle Packing’ Problem

Math enthusiast nor picked up GPT-6 Astra and formally proved the remaining N ≤ 20 cases of a classic “circle-in-circle packing” problem — specifically the cases 15, 16, 17, 18, and 20. The puzzle looks like an innocent olympiad problem, but the hard part is rigor: in 1969, researcher Udo Pirl found a beautiful arrangement and carried the proof to 10 disks, yet nobody had rigorously shown optimality for nearly six decades. Nor’s approach: instead of searching for the optimal arrangement directly, use geometric constraints to shrink infinitely many possible layouts into finitely many cases, then eliminate them one by one with the computer — pushing outer disks toward the boundary, ruling out arrangements via angular constraints (if required angular separations sum past 360°, the layout dies), enumerating distance intervals and winding orders, and finally using geometric rigidity to show that “very close to the classical arrangement” must mean “strictly identical,” capped by proving at least one disk always touches the outer boundary. In his write-up, nor noted the real work lay in scoping the problem, guiding the model, wrestling a buggy web UI, manually verifying results, and re-checking everything with Astra and Opus 5.5. Commenters marveled that AI is making math feel more explorable — and the field’s atmosphere more electric — than ever.

Source: QbitAI

3. ByteDance Unveils Its World Model: AI Moves From ‘Image-Making’ to ‘World-Building’

At Douyin Group’s “2026 Wonderful Fantastic Night” tech showcase, film-tech creator Tim (影视飓风) demonstrated something straight out of science fiction: riding a paper airplane through a free-roaming cyber world, interacting in real time, editing objects, and stepping seamlessly between different spaces. The engine behind it all is ByteDance’s long-in-development world model — not yet officially released, but the demo already showed 3D scene generation, free exploration, real-time interaction, object editing, and cross-scene continuous modeling. The industry has been racing to it: Fei-Fei Li’s World Labs (acquired by AMD for $8.2B in September), Google DeepMind’s Genie, and Nvidia’s Cosmos are all betting that AI must learn spatial structure and how the world responds to actions. In June, Fei-Fei Li’s team published “A Functional Taxonomy of World Models,” carving the space into three functional classes. Chinese vendors are taking a differentiated path built on their image/video-generation muscle: Seedance 2.5 stretched single generations to 30 seconds, and Seedream 5.0 Pro supports interactive region-based editing. World models matter beyond content, too — robotics and autonomous driving have long been throttled by the cost of real-world data and the impossibility of rehearsing dangerous scenarios, and world models can synthesize those long-tail situations on demand, moving part of the training loop from the physical world into the virtual one.

Source: QbitAI

4. Huawei’s Memento 3 Clears All 25 ARC-AGI-3 Games With a Perfect RHAE — 56% Fewer Actions Than Humans

Huawei’s Noah’s Ark Lab, working with UCL and other teams, released Memento 3: an agent system built on Claude Opus 5 that never updates model weights and is never fine-tuned for gaming, yet scored a perfect 100 RHAE (Relative Human Action Efficiency) across all 25 games and 183 levels of ARC Prize’s ARC-AGI-3 benchmark — using just 7,518 actions, about 56% fewer than the human baseline, and even beating independent researcher Sergey Rodionov’s baseline1 (99 points, 8,347 actions) by 829 actions. For reference, Claude Opus 5 alone averages only ~40.7 RHAE under the official ARC Prize framework. Memento 3’s secret is Model-Based Recursive Self-Improvement: it builds inspectable, executable, rollback-able external world models (code world models) so a frozen general model can iterate and evolve in unfamiliar territory. The details are striking: in symbol-decryption game tr87 the AI initially couldn’t tell whether symbol orientation mattered, then discovered differently oriented symbols can encode the same letter; in game wa30, one world model took 899 actions while dual world models dropped it to 597. The team even threw it at Atari Pong — without a world model the AI scored -19 (about the same as random button-mashing at -20.7), but with one it re-derived the collision and bounce physics, correcting its model on the fly until it could volley reliably. The authors caution that the experiments focus on rule-clear game environments, and whether the approach survives the messy real world remains to be seen.

Source: QbitAI


Model Releases

5. Zhipu’s New Model Leaks at 490 Tokens/s — Likely GLM-5.3-Highspeed

A new Zhipu model has reportedly opened up in Zcode 3.15.1 for random subscribers, per tipster @leo114119, and its headline number is speed: 490 tokens/s. At that sustained rate, generating 10,000 tokens would take about 20 seconds of pure generation time. Context is everything here — last month Zhipu’s GLM-5.3-FlashX topped out at 200 tokens/s, and Artificial Analysis measures GLM-5.3-Flash at roughly 60 tokens/s on the Z.ai API. The leaked directory entry lists a 1M-token context window, a maximum generation cap of 131,072 (~128k) tokens, and three thinking-effort tiers (low/high/max). Reddit sleuths are already speculating about a link to the anonymous “Omen Alpha” model on OpenCode, with other guesses including GLM-5.3-Omni and GLM-5.4-Flash. Zhipu set the global API speed record with GLM-5.1-highspeed (400 tokens/s) back in May; if GLM-5.3-Highspeed really hits 490 TPS, it would be a major win for long-horizon agent tasks.

Source: QbitAI


Industry

6. $48B Has Poured Into Physical AI in 2026 — More Than All of Last Year

Per the Financial Times citing PitchBook, robotics and physical AI companies globally have raised nearly $48 billion (~RMB 321.3 billion) this year as of October 8 — roughly 1.7x the full-year 2025 total, versus the $10–15 billion annual range of 2022–2024. Q2 alone brought in $18.6 billion, up 25.3% quarter over quarter. Mega-rounds are stacking up: Anduril closed a $5B raise in May at a $61B valuation (led by Thrive Capital and a16z); Skild AI took $1.4B in January led by SoftBank with Nvidia and Bezos’s vehicles participating, at a valuation above $14B; NEURA Robotics announced an up-to-$1.4B Series C in June (Qualcomm, Amazon, Nvidia); Physical Intelligence raised $1B in Q2; and Generalist added $400M in June, with Fei-Fei Li appearing among new individual investors. Data acquisition has become its own arms race: Figure’s Index program pays ordinary people worldwide to record everyday tasks — pass 264,000 downloads across 108 countries, over 16 million video clips collected, $15M already paid out, and more than $1B planned for data and compute over the next 12 months. China has at least 90 humanoid-robot data collection and training centers operating, under construction, or planned (64 already live). And Dyna Robotics’ Dyna-2, pretrained on over 1 million hours of first-person human video, lifted average performance on 14 real manipulation tasks from 20% to 53% — evidence that LLM-era scaling laws are starting to replicate in the robot world.

Source: QbitAI, Financial Times

7. Lenovo’s TianxiCode Tops SWE-bench-Live With a 71% Solve Rate

On SWE-bench-Live (Lite sub-leaderboard) — the dynamic benchmark widely considered the closest proxy for real-world software engineering — Lenovo’s Tianxi AI code-agent framework TianxiCode, paired with DeepSeek-v4.1-Flash, took first place globally with a 71% problem-solving rate, and passed the official “Verified” audit, which requires submitting the full agent trajectory for review to rule out any leakage of ground truth. TianxiCode’s edge is three systems-level engineering capabilities: cross-file multi-hop retrieval with precise context pruning, letting the agent trace root causes through large codebases like a senior engineer; autonomous planning with multi-turn tool calling, so it opens terminals, runs tests, and reads logs on its own; and closed-loop patching with test-driven self-correction, executing code in an isolated environment and revising until patches pass. Lenovo frames the pairing as an engineering team: DeepSeek-v4.1-Flash is the brain, TianxiCode is the hands, eyes, toolbox, and workflow discipline. The framework is slated to land inside Lenovo’s AI hardware products.

Source: QbitAI


Compiled from multiple media reports.