Multimodal

SiliconFlow launched DeepSeek-V4.1-Flash on Day 0. It's a 552B MoE model with ~8B active params during prefill and ~16B during decode, native vision, and a 1M-token context window. KV cache footprint is about 1/4 of V4 Flash, which helps with deployment cost. MIT license keeps commercial use straightforward.

DeepSeek released V4.1-Flash, a 552B-parameter model built to slash memory costs for AI agents. Its KV cache in fast GPU memory is about a quarter the size of V4-Flash, and the offloaded portion shrinks to roughly an eighth. The model splits into an encoder and decoder: only 8B parameters activate per token during input processing, versus 16B during text generation, nearly halving input compute. It supports 1M-token contexts and stores the main KV cache in FP4. On the DeepSWE v1.1 coding benchmark it scores 74.2%, narrowly beating Anthropic Opus 5 and OpenAI GPT-5.6 Sol, but it still trails on complex scientific tasks and image analysis. Weights are on Hugging Face under the MIT license. The post does not disclose inference latency or specific hardware requirements.

DeepSeek released V4.1-Flash with a new Causal Encoder-Decoder architecture and native vision understanding — no separate vision model needed. It's a 552B MoE, activating 8B params for input and 16B for output. The post doesn't disclose the exact price cut or benchmark numbers, so I'd wait for third-party evals before getting excited.

DeepSeek V4.1-Flash is the smallest model in the new architecture family: a 552B MoE with 8B active params for input and 16B for output. It uses a Causal Encoder-Decoder design with native vision. KV cache drops to 1/4 of HBM and 1/8 of SSD storage vs the previous generation, and API pricing is lower. The post doesn't disclose exact pricing or vision benchmarks.

OpenAI released ChatGPT Images 2.5 with two models: Flare for speed (up to 50% lower latency) and Sunburst for tighter editing control at longer generation times. API pricing is $8/1M input tokens and $30/1M output tokens. New 'xhigh' and 'max' quality tiers push a 1024x1024 max image to roughly $0.21. No batch pricing is offered yet, and OpenAI didn't share an average per-image cost. ChatGPT users can't manually pick the model; the post's tests show Work mode keeps edits more stable than Chat mode. Both models top the image arena rankings, and watermarking is added in partnership with Google DeepMind.

OpenAI released Images 2.5, a new image model that cuts generation latency by up to 50% and improves lighting, textures, and multi-turn editing consistency. Over 3 billion images are already created weekly across ChatGPT and the API. A new Sketch feature lets users draw directly in ChatGPT as a reference. API availability is confirmed, but the post does not disclose pricing details.

Users chained GPT-6 Astra with ByteDance Seedance 2.5 into an end-to-end AI film pipeline: Astra builds scenes and previs in Blender, Seedance turns reference frames into anime-style clips, and Astra handles the final edit. The same workflow produced a Naruto fan short and a US remake of a Chinese drama. Fable 5.1 and Gemini 3.8 Flash were also tested as prompt writers for Seedance, showing distinct directorial styles. Separately, Astra was used as a DaVinci Resolve colorist, matching a reference look in 4 minutes, though opinions on the result were mixed. The post does not disclose Seedance 2.5 technical specs or pricing.

Robocurve gave GPT-6 Astra control of YAM arms on two tasks, head-to-head with Claude Fable 5.1. On block-into-bowl, Astra scored 19/20 (95%) vs Fable 5.1's 8/20, averaging 2.5 min and $0.94 per run—less than half the time and cost of Fable 5.1's 6.8 min and $2.12. On the puzzle-insertion task, Astra managed 2/20, same as Fable 5.1; both stall at the final alignment step, at $1.36 per run. Clear win on pick-and-place, no progress on fine insertion.

Atlas is a multimodal autoregressive diffusion transformer pretrained from scratch to handle text, images, video, and 3D. It stitches reference images into a coherent 3D scene with pixel-perfect camera control, generating up to 1 minute of 1440p video. On sparse-view 3D reconstruction (2–3 images), Atlas beats specialized reconstruction models; more input images reduce guesswork. Early access is open for request, but the post doesn't disclose pricing or a public launch date.

Google launched Pics, an AI design tool that generates posters, social posts, and illustrations from text prompts. It's part of Workspace for business users and Google AI Pro/Ultra subscribers, running on the Nano Banana image model. Unlike Canva or Adobe Express, there's no marketplace for creator templates—everything is AI-generated from scratch. The post doesn't disclose pricing or exact rollout dates, only 'over the coming weeks.'

Google DeepMind added agentic video understanding to Gemini: it can watch a video of a UI and then perform the same clicks, typing, and scrolling itself. Instead of just describing what it sees, Gemini executes multi-step tasks like filling web forms or completing an order in a mobile app. The feature is now available for testing in the Gemini app and Google AI Studio. The post doesn't disclose latency or success rates—real-world UI agent reliability is still a big open question.

OpenAI is gray-testing Astra, a model that one-shots full frontend webpages from scratch—testers declared 'frontend is solved.' Anthropic is rushing Fable 5.1, and both sides are already trading SVG stability comparisons. Meanwhile, Claude Code's safety mechanism downgraded a dangerous file-cleanup task to the weaker Opus 4.8, which correctly identified the home directory as off-limits, then deleted 700GB of it anyway. A coding growth analysis shows non-engineer Codex usage growing 108x in legal, 41x in sales, with broad coding tasks driving 60–70% of OpenAI ARR. Hy4 preview scaled up urgently after a usage spike, but real-world prefill hits ~20K tokens and long sessions take 24.7s. Dual GB10 running DeepSeek V4 Flash hit 200.3 tok/s aggregate throughput at 6 concurrency. Fireworks delayed GLM-5.3-Flash by two days after discovering EvalScope prompts caused 2–3x overthinking. The group also discussed orthogonal design for cheaper code review and a prescription for vibe coding addiction: no agent one hour before bed.

Meta, Z.ai, and DeepSeek each released multimodal models in August with strikingly similar demos: the model observes a video or screenshot, calls tools to generate a webpage, slides, or a mini-game, then inspects its own output. This shifts vision from a passive input channel to an action the model initiates. The article likens it to the 2023 shift from static RAG to agentic RAG, but notes the loop direction is reversed—here the model self-verifies after producing. Evaluation moves beyond image Q&A: Meta's WildArtifactBench uses pairwise comparisons and Elo scores to assess full artifact creation. Training also changes; both GLM and Meta train models in generate-inspect-revise loops, logging interaction trajectories as training data. For builders, the key question is no longer static image accuracy but whether the model can complete an observe-generate-inspect closed loop.

Alibaba Qwen open-sourced Qwen3.8-Flash with full weights. It's a multimodal MoE model with 125B total parameters, activating only 6B per token. Training cost is 1/9 of Qwen3.7-Plus while outperforming it across the board. Production API pricing is $0.16/1M input tokens and $0.47/1M output tokens, with 262K native context expandable to 1M. The model also serves as an early preview of the Qwen4 architecture.

Z.ai confirmed it is the lab behind Ox Alpha, the open-weight model that appeared anonymously on OpenRouter and immediately topped rankings. The company calls it the newest GLM iteration, built for coding, sustained agentic work, and multimodal reasoning. Weights drop Wednesday for developers to build on. Earlier GLM-5.3 already matched Anthropic's Fable 5 on some benchmarks. Ox Alpha adds more pressure on frontier pricing from OpenAI and Anthropic.

Qwen3.8-Flash is an early preview of the Qwen4 architecture: 125B total params, only 6B active per token. Native context is 262K, extendable to 1M. Training cost is just 1/9 of Qwen3.7-Plus, with better coding and office-task performance. Weights are open. The post doesn't disclose specific benchmark scores or license details.

Qwen teased Qwen3.8-Flash-Next on ModelScope, a multimodal MoE model built on the next-gen Qwen4 architecture with 125B total and ~6B active parameters. The early release is meant to preview Qwen4's design for the community. It drops 2026-08-26 15:00 UTC, with an FP8 variant alongside. The post doesn't disclose benchmarks, inference speed, or specific multimodal capabilities—I'll hold judgment until the model card lands.

OpenAI banned a cluster of Russia-based ChatGPT accounts used to promote the International Burke Institute (IBI), a fake think tank claiming to be in Israel. The site copied academic work, used machine translation, and published a sovereignty index favoring Russia. Operators prompted the model in Russian to generate English social media posts while hiding linguistic clues. OpenAI calls this the most elaborate Russia-linked IO they've disrupted since the Ukraine war began, though it reached relatively small audiences.

The Guardian reports on Hollywood concept artists, voice actors, and writers hired by AI firms to train models at $25–$150/hour. They know they are teaching AI to replicate their own craft—some call it 'digging the grave of my profession.' Runway and Sora are named as key employers, but the article doesn't disclose contract scale or training data volume. Treat this as an industry mood piece rather than a quantified displacement forecast.

Meta's Muse Glimmer is a ~30B multimodal model built to run agentic tasks offline on consumer hardware. The BF16 checkpoint is 55 GiB; Meta ships ~4-bit quantized versions that bring the language model under 20 GB. Architecturally, only every fourth of the 52 layers uses full-context attention—the other 39 use a 2,048-token sliding window. Global layers drop RoPE and retrieve by content. The KV cache stores just two key/value heads while 32 query heads provide diverse retrieval behaviors. A large ViT handles perception once, compresses neighboring patches 4:1, and feeds them as tokens. The result is a memory hierarchy: local layers build ordered representations, global layers search across the full sequence, and the tiny KV cache means quantization savings translate directly into longer context or larger batches.