xAI ships Grok Imagine Image 2.0 — region-level editing becomes first-class
xAI’s Imagine Image 2.0 treats region-level editing as a core feature, expanding Grok’s image stack into wand, background, and multi-reference edits.
Models
AI models shape what today’s assistants can reason through, create and automate. ByteForward covers large language models, multimodal systems and open weight releases with a focus on what their capabilities mean in practice. Explore reporting on reasoning performance, coding ability, context windows and the tradeoffs between speed and cost. Model benchmark comparisons are most useful when the test setup, version and limits stay attached to the score. API pricing, licensing terms and access restrictions also matter when choosing a model for a real workload. Our AI model release coverage follows new launches and major updates from announcement to availability. For the methods behind performance claims, explore our AI research reporting. Start with the stories below to understand what changed, what the evidence supports and which questions remain open.
Alibaba opens Wan 3.0 video generation to public beta testers
While the rest of the week argued about coding agents, Alibaba opened a public beta for Wan 3.0. The product page is selling longer clips, more reference assets, and a free-quota on-ramp that turns into a per-second bill.
OpenAI updates GPT 5.6 Sol and expands access to Luna
OpenAI just retuned the ChatGPT front door without touching the agent stack behind it. In today’s product post, Plus and Pro get an updated GPT-5.6 Sol built for tighter, more factual chat answers plus a reasoning-effort slider. Free and Go…
Meta’s Muse Spark breached a real company — via a misconfigured Irregular sandbox
Meta’s Muse Spark 1.1 reached a real company during a cyber eval because the test sandbox had internet access it was never supposed to have. The disclosure landed August 5. The failure was the harness, not a frontier model inventing…
Mistral introduces Shieldstral for text and image safety classification
Mistral open-weights Shieldstral 1.0 under Apache 2.0: a 3B multimodal safety classifier that takes policy as a question and returns calibrated scores.
Liquid AI ships LFM2.5-2.6B — on-device agents with 128K context and open weights
Liquid AI’s LFM2.5-2.6B runs planning and tool-calling agents on phones and laptops, with ~34T pretraining tokens and open weights for on-device stacks.
Alibaba introduces Qwen 3.8 Max with API access
Alibaba launched Qwen 3.8 Max with global API access and plans for model weights. The launch announcement did not verify the pricing quoted in the earlier article.
ByteDance ships Seedance 2.5 — single-take video generation doubles to 30 seconds
ByteDance Seed’s Seedance 2.5 extends single-take video generation to 30 seconds with multimodal referencing that changes short-form production math.
Tether QVAC open-sources VisionPsy-Nano — company claims top sub-0.5B on-device scores
Another sub-half-billion vision-language model just hit Apache 2.0 with a phone-first pitch. Tether Data’s QVAC initiative open-sourced VisionPsy-Nano, a ~460M-parameter VLM aimed at on-device and edge runs, and used a PR Newswire release to claim the top normalized score among…
Google DeepMind launches Gemini Robotics ER 2 for embodied reasoning
Robot “brains” just got a version that watches the job finish instead of guessing from a still. Google DeepMind launched Gemini Robotics ER 2, its most capable embodied-reasoning model: continuous video progress tracking, sub-second moment finding, native tool use (including…
DeepMind introduces Gemini Robotics 2 for full body robot control
DeepMind’s Gemini Robotics 2 is a vision-language-action model for full-humanoid control, shipping as a distinct product from Gemini Robotics ER 2.
xAI ships Grok Voice Think Fast 2.0 — time-to-first-audio hits 0.70s
xAI’s Grok Voice Think Fast 2.0 ships speech-to-speech at 0.70s TTFA and $0.08/min, becoming the latest grok-voice alias for low-latency voice agents.
Google brings Lyria 3.5 music generation to free Flow Music
DeepMind’s Lyria 3.5 music model lands in free Flow Music with richer melodies and longer songs, pushing AI music further into everyday listening habits.
Microsoft introduces MAI Cyber 1 Flash through MDASH
Microsoft introduced MAI Cyber 1 Flash inside MDASH. Its reported CyberGym score covers a combined system, with access limited to verified defenders.
Anthropic launches Claude Opus 5 with a lower cost per task
Anthropic just made the expensive frontier feel optional for a lot of daily work. Claude Opus 5 is live, pitched as near–Fable 5 intelligence on coding and knowledge-work evals at roughly half the per-task cost, while keeping the same $5…
Black Forest Labs ships FLUX 3 — multimodal video, image, and robotics in one backbone
FLUX 3 expands Black Forest Labs’ stack into multimodal video-with-audio, image editing, and robotics in one backbone for creative and industrial work.
Google introduces Gemini 3.6 Flash and two new 3.5 Flash variants
Google just cut the bill for agent loops and kept the sharpest knife behind a velvet rope. On July 21, the company launched Gemini 3.6 Flash, plus 3.5 Flash-Lite and a gated 3.5 Flash Cyber stack inside CodeMender. The pitch…
Alibaba previews Qwen 3.8 Max ahead of planned August release
Alibaba put a Max-class model on the WAIC stage before the scoreboard existed. On July 19 the Qwen team previewed Qwen3.8-Max as a sparse multimodal MoE the company pegs near 2.4 trillion total parameters — with coding and “cowork” claims…
Google Vids adds Gemini Omni editing and personal avatars
Google just made Workspace video feel less like a slide deck with motion and more like a generative editor. In a Workspace blog post, the company says Gemini Omni and personal avatars are rolling into Google Vids, so teams can…
Moonshot AI launches Kimi K3 — 2.8T open frontier MoE with 1M context
The open-weight ceiling just jumped a class. Moonshot AI’s Kimi K3 is live as a 2.8-trillion-parameter Mixture-of-Experts model with about 104 billion parameters active per token, native vision, and a 1-million-token context window — and Moonshot is promising the full…
Thinking Machines Lab releases open-weights Inkling — 975B MoE enters the arena
Thinking Machines Lab just put a frontier-scale multimodal MoE on Hugging Face and said the quiet part out loud: it is not the strongest model on the board. The bet is customization, not crown. In a July 15 announcement, the…
xAI introduces Grok 4.5 for coding and agent tasks
The July 16 Grok 4.5 announcement details launch pricing and reported coding results. This update corrects the earlier date and separates benchmarks from deployment decisions.
Mistral ships Robostral Navigate — 8B embodied model steers robots from one RGB camera
Mistral’s first embodied model, Robostral Navigate, is an 8B system that follows plain-language instructions to steer robots from one RGB camera.
Cohere releases open-weight Transcribe Arabic — dialect ASR that beats Whisper baselines
Cohere’s 2B open-weight Transcribe Arabic targets dialect and code-switching speech, with claims it beats Whisper baselines on hard Arabic ASR cases.