AI Brief, 28 October 2026: the world model that shipped without a number

Google DeepMind published Genie 4 at about 16:00 UTC on 27 October, and it took 794 points on Hacker News inside a day, submitted four minutes after the post went up. It is the loudest launch of the window by a wide margin. Measured against its own predecessor it is also the least documented: no paper, no technical report, no model card, no arXiv entry, no benchmark table, no parameter count. Genie 3, announced in August 2025, published 720p at 24 frames per second with visual memory reaching about one minute back. For Genie 4 none of those figures exists in public, and the one duration statement Google does make, "a few minutes of continuous interaction", is the phrase Genie 3's materials already used. On Google's own documents, no improvement in session length is claimable.

What is new is an axis rather than a number. Genie 4 is described as DeepMind's first real-time world model to generate audio natively — ambient sound, music, and dialogue that adapts while the scene is being steered — with Gemini 3.8 Flash acting as a guiding agent that decomposes a goal into steps and checks incoming frames against them. Every capability claim is self-reported, unaudited and demonstrated on curated recordings. Nothing is independently measured, because nothing is released: access is a waitlist form, with a research preview for trusted testers and a promise of Google AI Ultra subscribers and early-access Gemini API partners later. Ars Technica notes that Genie 3 never expanded beyond a limited web experience and puts AI Ultra at $100 a month; TechCrunch, from the briefing, reports a two-minute consistency limit that appears in no Google document. The two disagree over whether the rollout is weeks away or has no timeline.

Ninety minutes later, two researchers at Helmholtz Munich published the question the announcement does not answer. "Understanding the Limits of Latent World Models" defines what it would mean for a world model to be right, proposes a metric for checking it, and reports that across three state-of-the-art latent world models, high task performance did not imply the model had learned the environment's dynamics — in one case the task score was all but identical while the metric collapsed ninefold. The paper's authors are careful about something the pairing above invites: they explicitly scope generative video models out, and Genie is never mentioned. That is the real state of play: the loudest world model of the year and a serious attempt to measure whether world models model anything arrived ninety minutes apart, and they do not reach each other.

• Genie 4, 27 October: native audio and interactive dialogue, research preview and waitlist only, no weights, no paper, no benchmarks, fewer published specifications than Genie 3.
• Faithfulness is now defined and measurable for action-conditioned latent world models; of the three families tested, the pixel-based ones kept almost all their reward with a fraction of the fidelity.
• Hy-Nomas, Tencent's 8B diffusion world model, released weights, code and a benchmark on 27 October with no score on it. The 24th flagged exactly this.
• Ling-3.1-flash-1019: 560B total, 25B active, MIT, 63 linear-attention layers to 9 full-attention, FP8-only at 573.8 GB, and no press coverage at all. Its card advertises a 1M-token context window; the shipped config sets 262,144 with no scaling configured.

The loudest world model of the year published no specifications #

Native audio is a real capability direction: a generated world that also generates its own sound, including dialogue that responds while the scene is being steered, is not something the previous generation did. DeepMind also claims gains in physics, naming aerodynamics and collisions, and calls human simulation "the most dramatic leap". The system around the model is described mostly in terms of other models: Gemini 3.8 Flash plans and verifies, Gemini Omni anchors appearance, and a separate experiment pairs Gemini 4 Argon with Gemini Robotics to simulate a humanoid playing checkers. The Decoder adds the caveat the others omit: no real robot.

Google's own materials list the failure modes in prose — object permanence "can still shift or drift over time", multi-agent interaction unsolved, legible text reliable only when it appears in the input description — and show no failure cases at all. The sharpest reading of the demos came from the thread rather than the coverage. One commenter itemised consistency breaks in the showcase clips: a second castle appearing in the dragon scene, the people on a train platform changing as they walk away, a moped rider acquiring a helmet between shots. Another argued this is structural rather than a tuning issue, since holding an occluded object stable implies maintaining something like a 3D scene model. A third questioned controllability, reading the comparison clip as video generation with game widgets on top, and noting that the checkers demo shows a piece lifting off the board and then not, plus a move that is not legal checkers — which lands on the robotics-training framing, the announcement's main argument for mattering at all. None of that is a measurement either, only careful viewing of curated video, which is all anyone outside the trusted-tester set has.

A metric that asks whether a world model models anything #

Rodrigo Gonzalez Laiz and Steffen Schneider, at Helmholtz Munich, posted their paper at 17:25 UTC on 27 October. Their argument is that the field evaluates world models almost entirely on downstream reward, which cannot distinguish a model that has learned the environment's dynamics from one that found a shortcut scoring well.

Their definition has two parts. Let $\text{bm}{s}_t$ be the unobserved environment state, $\text{bm}{a}_t$ the action, and $\mathbf{f}$ the environment's transition map, so $\text{bm}{s}_{t+1}=\mathbf{f}(\text{bm}{s}_t,\text{bm}{a}_t)$ . Observations come from a mixing function $\mathbf{g}$ , the encoder is $\mathbf{h}$ , and the learned latent dynamics are $\hat{\mathbf{f}}$ ; composing the first two gives the state-to-latent map $\mathbf{r}:=\mathbf{h}\circ \mathbf{g}$ . The model is faithful when $\mathbf{r}$ is injective — no two environment states collapse to one latent — and when the two ways of taking a step agree:

$$\hat{\mathbf{f}}(\mathbf{r}(\text{bm}s),\text{bm}a)=\mathbf{r}(\mathbf{f}(\text{bm}s,\text{bm}a))$$

Stepping the real environment and then mapping in should land where mapping in and then stepping the model lands. Their metric, latent rollout consistency, is the $R^2$ between the latent the encoder infers from the realised observations and the one the model predicted, over a multi-step open-loop rollout: 100% is exact, 0% matches a mean predictor, and negative values are kept.

They evaluated DreamerV3, TD-MPC2 and LeWM, a joint-embedding predictive architecture, four seeds each from both state vectors and pixels, on three DeepMind Control Suite tasks. The dissociation is cleanest on CartPole swing-up.

The DreamerV3 pair is the cleanest result in the paper: reward 872 from pixels against 871 from states, statistically indistinguishable, while 30-step consistency falls from 81.9% to 8.9%. TD-MPC2 keeps 95% of its reward from pixels and goes below zero.

Two caveats belong with it, and the authors raise both. The metric compares the predicted latent against the same model's encoder output, so it measures internal self-consistency rather than agreement with the true state, which is why they pair it with probes that decode physical state. And they offer a less damning reading they decline to adjudicate: the predictive information may sit implicitly in the policy and value networks, compensating for an inaccurate dynamics model. Their summary is that these are "controlled counterexamples", "rather than as evidence that all latent world models exhibit the same failure modes".

Tencent's Hy-Nomas paper, posted at 06:48 UTC the same day with 38 authors, is the contrast case on evidence rather than method: an 8B diffusion backbone for interactive world generation, aligning structured captions with action-aware causal attention so descriptive context survives across segments while action instructions stay in their intended windows. Weights, code and benchmarks are released, which puts it ahead of Genie 4 on everything except noise. The evaluation is one sentence: human raters preferred it. No table, and no score on Hy-Nomas-Bench, the 280-case benchmark Tencent published with the weights on 23 October and which this brief flagged on the 24th. A lab that ships a benchmark and declines to score itself on it has turned a gap between conversion and evaluation into one between publication and evaluation.

560B parameters under MIT, and a context window the weights do not have #

While that ran, inclusionAI — the Ant Group organisation behind the LLaDA models covered here on 6 September — uploaded Ling-3.1-flash-1019 at 10:45 UTC on 27 October under an MIT licence: 560B parameters, 25B active. It has 65 likes and no press coverage at all.

The architecture is the interesting part, and the card states it plainly: a hybrid-linear MoE of 72 layers, of which 63 are KDA linear-attention layers and 9 are gated multi-head latent attention, a 7:1 ratio, with the first four layers dense and the rest routing 8 of 512 experts plus one shared expert. That ratio is the whole engineering story, because only the full-attention layers keep a cache that grows with the sequence; the linear layers carry a fixed-size recurrent state whatever the length. At the weights' own FP8 precision, one byte per value, that is the difference between caching everything and caching almost nothing:

# conventional: all 72 layers cache keys and values
2 x 64 heads x 128 dims       = 16,384 bytes per token per layer
16,384 x 72 x 262,144         =    309 GB at full context

# Ling: 9 attention layers cache a rank-512 latent plus a 64-dim rotary key
512 + 64                      =    576 bytes per token per layer
576 x 9 x 262,144             =   1.36 GB at full context

A factor of roughly 228, at the same sequence length and head geometry. The practical meaning is not whether one long context fits but how many fit at once: on the 8-way H200 deployment the card recommends, the weights take 573.8 GB and what is left serves hundreds of concurrent 256k-token sequences rather than one.

Three caveats, and the first is a flat contradiction. The card advertises a 1M-token context window. The shipped config.json sets max_position_embeddings to 262,144 with rope_scaling null, and the card carries no scaling recipe:

from huggingface_hub import hf_hub_download
import json
cfg = json.load(open(hf_hub_download("inclusionAI/Ling-3.1-flash-1019", "config.json")))
cfg["max_position_embeddings"], cfg["rope_scaling"]   # (262144, None)
cfg["quantization_config"]["fmt"]                     # 'e4m3' — FP8, the only format shipped

inclusionAI's hosted endpoint does advertise 1,000,000 tokens, so the 1M figure describes the service, not the download. Second, the weights are FP8 only: 54 shards, 573.8 GB, no BF16 reference, which leaves readers where MiniMax left them on 23 October — every published number describes the quantised model with nothing to compare against. Third, the headline score, 44 on the Artificial Analysis Intelligence Index against 41 for its predecessor, is attributed by the card to version 4.3.2, and Artificial Analysis now publishes v5.0, built from nine evaluations, so the quoted figure describes a superseded scale.

Also notable #

• Maya Research published maya2-native under Apache 2.0, weights on 26 October and the model card, launch video and benchmark table on the 27th: a 2.5B streaming text-to-speech model on a Qwen3-1.7B base, 11 Indic languages plus English, 20 voices, cloning from 3 to 10 seconds of audio. It claims second for Hindi on VoiceArena's real-time board at 1047 Elo against Cartesia's Sonic-3.6 at 1094 — a third-party ranking, rarer than it should be. The card also discloses, against its own interest, that VoiceArena dates its entry to July 2026 and names no checkpoint revision, so the ranking may not describe this week's weights.
• Liquid AI shipped the LFM3 family on 27 October: a 1.73B text model with vision, audio and omni variants, plus two d1 decision models that generate no tokens at all, returning typed answers with calibrated probabilities in a single forward pass, timed by Liquid at 5.7 ms per question on an H100. Every figure is Liquid's own, and the licence, named the "LFM Open License", caps commercial use at $10 million of group-wide annual revenue.
• Anthropic published partner results from six drug-discovery collaborations, at 170 points on Hacker News. The checkable figures belong to the partners: Enveda reports that across 1,061 pages of manufacturing documents Claude found all 11 issues human reviewers had caught and flagged 7 they had missed, cutting review time 89%. All self-reported by interested parties, and mostly preclinical.
• An aggregator began reselling safety-stripped open weights, advertising five providers for an abliterated DeepSeek V4.1 Flash 1017 and three for GLM-5.3 from $0.22 per million input tokens. It does not check out: OpenRouter's public list carries the unmodified model across 36 endpoints and none of those five providers, and names neither the source repository nor the abliterator.

What to watch #

• Whether Genie 4 produces a single specification. Resolution, frame rate, latency and parameter count are absent and the duration wording matches Genie 3's, so the claimed advance rests on curated video. The two-minute figure in circulation came from a briefing, not a document.
• Whether latent rollout consistency is extended to a video world model. Its authors scope generative video out and tested three latent families on three control tasks, so it does not currently reach Genie 4 or Hy-Nomas, whose released weights make it the obvious first target.
• Whether anyone scores Hy-Nomas-Bench, flagged on the 24th and now four days open. The benchmark, the weights and the paper all exist; a number does not, Tencent's own included.
• Still open: Ling-3.1-flash-1019's 1M claim stands against a config capped at 262,144, with no unquantised reference, the same gap MiniMax left on the 23rd. Kimi K3.1 Pro's weights remain absent against a promised 4 November date, flagged on the 24th, and Nanbeige4.3-33B still has no independent benchmark number, flagged on the 26th. And no decision-model vendor has published a fail-open rate, asked here daily since 10 October and now nineteen days: Liquid's two new d1 models publish a calibrated probability per answer and no rate at which they answer anyway.