This week in models, the cost math flipped. Z.AI, Alibaba, and Tencent all shipped big Mixture-of-Experts systems that activate only a fraction of their parameters per token. The theme is sparsity: more total capacity, tiny per-token cost, long context, open weights. Here is what landed and what it means.
The three launches that reset the cost math
GLM-5.3-Flash runs 320B parameters with only 18B active per token, combining hybrid sparse and linear attention with vision and a 1M-token context window. It ships under an MIT license and is free to trial. The architecture signals where frontier efficiency is heading: massive total capacity, tiny per-token cost. Read more →
Qwen 3.8 Flash Next activates 6B of its 125B total parameters and serves a native 262,144-token context that extends to 1M via the official API. Built on the qwen4_exp architecture, it targets long-context workloads without the inference penalty of dense models at similar total scale. Read more →
Tencent Hy4-preview is a 770B MoE model with 49B active parameters, released under Apache-2.0 with context beyond 1M tokens. It is explicitly aimed at long-horizon reasoning tasks. The open license and scale make it one of the largest openly available systems this week. Read more →
The visual break
A pause for the visual side of AI — video generation, 3D reconstruction, and spatial models are moving as fast as the language stack. Below: this week's image and 3D releases.
BRIA Fibo 1.5 is an open, JSON-native image generation model trained on long structured captions, expanding short prompts into structured JSON for fine-grained control. Its training data is licensed for commercial safety. The JSON-native design makes prompt structure a first-class input rather than a post-hoc interpretation. Read more →
Block3D generates high-precision text-to-3D assets with sharply reduced latency, producing production-ready meshes for games and design. The work comes from Zhejiang University, Monash, and Adelaide. The latency reduction matters because iterative 3D workflows live or die on per-iteration speed. Read more →
OVOW 4D turns a single monocular video into an editable physical 4D world with object instances, meshes, and lights at roughly 3.35 seconds per frame. That is one to two orders of magnitude faster than baselines. Editability of the reconstructed scene — not just reconstruction — is the differentiator. Read more →
FixAnything is a single model that repairs artifacts across 3DGS, NeRF, meshes, and sparse point clouds by repurposing a pretrained video generative model for implicit multi-view priors. It comes from CMU and appears at ECCV. One repair model across representations is simpler than a dedicated fixer per format. Read more →
Sponsored
After Intelligence is a weekly briefing for builders who need to track AI releases without the noise. No hype, no sponsored "thought leadership," just the models, the numbers, and the links. This edition is brought to you by After Intelligence — the independent weekly digest for people who ship AI systems.
Audio, transcription and more
Gemini 3.5 Transcribe delivers low-latency speech-to-text with speaker diarization, word-level timestamps, and utterance-based language detection. It is built as a transcription product, not just a research endpoint. The diarization and timestamp granularity matter for anything downstream that needs to know who said what and when. Read more →
Gemini Omni 1.1 Flash adds production video updates including scene extension that reads up to 10 seconds of prior footage and chains to roughly 40 seconds, plus start/end-frame control and 4K output. Language-driven edits let you steer generation without re-prompting from scratch. The scene-extension chaining is the practical unlock for longer-form video. Read more →
Google's Planetary Prediction Engine turns natural-language queries into executed geospatial models, covering spatial regression, downscaling, and nowcasting. It runs on the Earth AI initiative. The appeal is removing the hand-coded pipeline between a question and a runnable spatial model. Read more →
MiniMax H3 is an open omni video model that generates synchronized video and stereo audio from a single prompt, shipping with two task-specific checkpoints for text encoding and Visual/Audio VAEs. It uses sparse attention in training with full attention at local inference. One prompt for both modalities removes the audio-sync post-processing step. Read more →
FastH3 is a 4-step DMD2 distillation of MiniMax H3 that runs text-to-video-and-audio up to 14× faster while keeping 90% sparse attention and H3 quality. It targets Blackwell GPUs. The distillation schedule compresses generation steps without retraining the underlying model. Read more →
Robotics, code and AI memory
Skild S1 performs in-context learning on ten-minute unseen manipulation tasks from a single video prompt, with no finetuning. The pancake-flip is not in its million-hour training corpus, yet it flips after one demo. That is the test of in-context skill transfer, not memorization. Read more →
Zero-WAM averages 46.95% zero-shot across seven unseen RoboTwin 2.0 tasks, beating LingBot-VA by +29.5 points. It is an in-context world-action model that predicts future frames and actions jointly, with no parameter updates at inference. The world model and the policy are the same network. Read more →
Meta CWM is a 32B open-weights code LLM that reasons over execution traces rather than only over source. The trace-conditioned formulation improves debugging and issue resolution on real GitHub tasks, where the failure signal lives in runtime behavior, not in the diff. Read more →
DiffusionOPSD uses on-policy self-distillation to close the supervision gap in reward-optimized diffusion. Image-level rewards are routed into intermediate denoising targets, and the step-distilled schedule cuts training GPU-hours by 63%. The gain is compute, not just sample quality. Read more →
RSI-Exam is a benchmark for agent self-improvement over long horizons and generalization to unseen data. It ships 88 active tasks across six domains and probes recursive self-improvement directly, not as a side effect of benchmark overfitting. Read more →
VoiceMem is a streaming dual-brain memory for real-time spoken assistants, splitting an information stream from an emotion-and-persona stream. The two memory streams run in parallel, and a memory-aware SLM training pipeline keeps retrieval grounded in what was actually said and felt. Read more →
That is the issue
If you want a slower read, the full index with model cards and benchmark tables lives on the After Intelligence site; the newsletter is the fast path.
Subscribe to get next week's briefs in your inbox before they hit the web.