AI Week, Aug 10–16, 2026: Pixel 11, a Three-Week Gemini Refresh and Meta's Single-GPU Agent
Google launched the Pixel 11 series and Gemini 3.7 Flash, Meta open-sourced its 30-billion-parameter Muse Glimmer agent model, and reports put Anthropic in talks to buy Decart and Apple behind its own model for the China market.
Written on November 13, 2026, covering news from Aug 10–16, 2026.
Agents ran through this week’s news, from a model Meta built to run locally to Google’s phones and its new Flash model. Meta released Muse Glimmer, an open-weight agent model sized for a single consumer GPU, on Monday. Google followed with the Pixel 11 series on Wednesday and Gemini 3.7 Flash on Thursday, while separate reports covered Anthropic’s talks to acquire Decart and Apple’s own language model for China.
Monday, August 10
Meta Superintelligence Labs released Muse Glimmer and published its weights under the permissive Apache 2.0 license. It is a 30-billion-parameter model that targets always-on agent workflows running locally; Meta says it fits on a Mac or PC equipped with one consumer GPU. The weights went up on Hugging Face together with developer documentation.
Meta describes a staged training process. Pre-training used logit distillation on outputs from Muse Spark, the teacher model; mid-training added longer-context, agent-heavy data; post-training mixed supervised fine-tuning, on-policy distillation and reinforcement learning.
The memory budget is the centre of the pitch. At full precision the model would need more than 55 GB, so Meta quantizes the weights to roughly 4-bit precision, bringing the language model below 20 GB. Meta says that leaves room for the KV cache, the image perception encoder and a speculative-decoding drafter inside a 24 GB or 32 GB envelope, with “minimal to no degradation” on agentic tasks. The drafter is based on DFlash and proposes blocks of tokens that the main model then verifies in parallel. Meta measured speed on a MacBook M4-Max, an M5-Max and an RTX-5090.
On capability, Meta compares Muse Glimmer with Gemma4-31B and Qwen3.6-27B and says it performs strongly for its size class. It lists DeepSearch QA, MCP-Atlas, 𝛕-Bench and SWE-Bench among the full-task benchmarks, and says the model takes interleaved text and images, works with OpenClaw and was trained on data from more than 100 languages. The release was assessed under Meta’s Advanced AI Scaling Framework.
Distribution was partly staged. Optimized llama.cpp, MLX and ExecuTorch integrations were promised “in the coming days,” along with local use through Ollama, LM Studio and Unsloth, serving via vLLM and SGLang, and hosted access through Together AI, Fireworks AI and OpenRouter. Meta named AMD, Arm, Dell, Intel and NVIDIA as partners on device optimization.
Tuesday, August 11
Spotify announced an “AI Persona” badge for artist profiles. The badge tells listeners that an artist’s identity may be AI-generated and not a real person. Artists could self-disclose as an AI Persona through Spotify for Artists starting August 11. The badge is due to appear from mid-September on artist profiles, in Search and on track rows across playlists.
Spotify said it will not rely on self-disclosure alone. It will also review profiles that appear to present photorealistic AI-generated identities, starting with artists who meet defined audience thresholds. Deadline reports that a human reviews every Spotify-applied badge before it goes live. Artists who receive a badge from Spotify will be notified and can self-disclose or appeal.
By default, AI Personas are excluded from editorial and algorithmic recommendations, unless, for example, a listener follows them. Spotify said the badge concerns an artist’s public identity, not how the music was made. It plans to let listeners report suspected AI Persona profiles in the coming months.
Google said more than 1 billion people now use the Gemini app every month. CEO Sundar Pichai called Gemini the company’s “fastest growing product ever” and its 14th product to reach that mark. Gemini app vice president Josh Woodward said more than a tenth of those users are on iOS. Google said more Android updates would follow the next day at Made by Google 2026.
Wednesday, August 12
Google unveiled Pixel 11, Pixel 11 Pro and Pixel 11 Pro XL, all running its new Tensor G6 chip, with 9To5Google noting the full details went out hours before the Made by Google 2026 event itself. Pre-orders opened that day at $899, $1,099 and $1,299, with the phones due on shelves at the Google Store and retail partners on August 20. All three start at 256GB of storage.
TechCrunch framed the lineup as carrying few design and internal changes but considerably more Gemini features. It noted that the Pixel 11 and Pixel 11 Pro match last year’s 256GB pricing, although a 128GB model used to be available for $100 less, and that the top model now costs $100 more than the Pixel 10 XL, which TechCrunch linked to rising component prices.
On the base Pixel 11, Google says the camera bar is 40% thinner, the new 48-megapixel main sensor delivers 56% more light sensitivity, and the 5x telephoto now supports 30x Super Zoom. The Pro models get a new main sensor, which 9To5Google identifies as 50MP, plus an upgraded 48-megapixel telephoto that Google says has 30% more light sensitivity, Portrait Mode at 5x and Pro Zoom at 120x. TechCrunch reported the telephoto light figure as 32%. On the Pro phones, Instant Night Sight is up to 4.5x faster.
Google says the Pro displays reach up to 3,600 nits and that a new coating more than doubles scratch resistance compared with the previous generation. 9To5Google reported 12 GB of RAM on the 256 GB Pro models, a 4 GB reduction, with 16 GB available on the 512 GB or 1 TB versions. Pixel 11 and Pixel 11 Pro add Qi 2.2 25W compatibility, and Google says the Pro XL gains 15 hours of battery life from 15 minutes of wired charging.
For Tensor G6, Google cites 25% faster web browsing, 15% quicker app launches and 50% more TPU compute, with on-device AI work running up to 3.5 times faster on up to 3.5 times less energy. TechCrunch, citing Google, put the chip’s power-efficiency gain at 20%. 9To5Google added that the TPU gets 2x memory bandwidth and runs the new Gemini Nano 4 model. A new Titan M3 security chip works with Tensor G6 to bring post-quantum cryptography to secure boot.
The Pro phones also get HiLight, colored LEDs around the flash that glow in patterns while Gemini listens, thinks or responds, and in custom colors for calls from chosen contacts; Google says more capabilities, such as message notifications, will follow. 9To5Google noted that HiLight replaces the temperature sensor and offers 2 features at launch.
On software, TechCrunch reported that US-based users will be able to have Gemini order groceries, book rides or get coffee, and that Gemini can call businesses for reservations or appointments, with Google saying users can take over or stop a task and review transcripts of AI-operated calls. More connected apps, including OpenTable (U.K.), Ticketmaster and Zocdoc, were due over the following few weeks, per TechCrunch. Rambler, the dictation product Google announced in May, is coming to Pixel. 9To5Google reported that Magic Cue is now called Proactive Assistance, that Live Translate extends to podcasts and videos in YouTube and other supported apps, and that Magic Capture analyzes around 500 frames to pick the best shot. TechCrunch also described a camera accessibility feature that uses Gemini to turn sign language into text.
All three phones get seven years of OS, security and Pixel Drop updates, and the Pro models include six months of Google AI Pro. 9To5Google’s event roundup also listed the Pixel 11 Pro Fold at $1,899, a $100 increase, the Pixel Watch 5 from $399, and the $29 Pixel Tag, which it said would not be available until November 11.
Thursday, August 13
Google released Gemini 3.7 Flash, which 9To5Google said arrived three weeks after the previous release. Android Authority noted that Gemini 3.6 Flash had been introduced in late July. Google pitches the new model at software engineering, web development and knowledge work.
Against 3.6 Flash, 9To5Google reported DeepSWE v1.1 rising from 49.0% to 65.3% and FrontierCode 1.1 Main from 34.4% to 43.6%, with the WebDev Arena Elo score moving from 1538 to 1588. It also cited 34.0% versus 22.0% on GDP.pdf and 30.4% versus 17.0% on AutomationBench.
Google says the model “better adapts to roadblocks, clarifies intent when needed, and follows instructions with greater fidelity”. It also carries updated safeguards covering CBRN and cyber offense misuse.
The introductory price is $0.75 per 1M input tokens and $3.75 per 1M output tokens until the end of 2026. 9To5Google described that as half the previous model’s launch price, while Android Authority said it matches the introductory price 3.6 Flash came with. Android Authority reported that after December 31, 2026, the rates rise to $1.50 and $7.50 respectively.
In the Gemini app, the model is rolling out to Spark, which requires an AI Pro or Ultra subscription. It is also available in Google Antigravity, AI Studio, Android Studio, Gemini Enterprise Agent Platform and the Gemini Enterprise app.
Also this week
Anthropic and Decart (reported on August 13). Fortune reported, citing people familiar with the matter, that Anthropic is in talks to buy Decart AI for about $6 billion. The deal had not been finalized and could fall through; if completed, it would be Anthropic’s largest known acquisition, coming ahead of a hotly anticipated initial public offering. Decart builds world models as well as software that helps chips run AI training more efficiently, and Fortune cited a person familiar with the matter saying the latter could help Anthropic’s infrastructure absorb more demand; one of the people said Decart’s team would join Anthropic’s inference and performance organization. Decart said in May that it had raised $300 million in a round led by Radical Ventures, which the Wall Street Journal reported valued it at almost $4 billion, up from $3.1 billion in August 2025. Fortune also noted a Calcalistech report of SpaceX talks with Decart, which Elon Musk called “fake news” on X. Representatives for both companies declined to comment.
Apple in China (reported on August 14). MacRumors, citing Reuters, reported that Apple trained an in-house large language model aimed at China, with help from Alibaba, rather than relying on a third-party model for Apple Intelligence there. The report said the rollout is expected in the coming months and described it as a “dual-track strategy”. MacRumors writes that this puts Apple first among foreign companies to win Chinese government approval for offering a proprietary AI model there. Apple had earlier agreed with Chinese regulators to incorporate Alibaba’s Qwen into Apple Intelligence, and briefly posted a support guide on connecting Qwen to Siri and Writing Tools on Mac before removing it in under a day.
What to Watch
- Gemini 3.7 Flash after December 31, 2026. Android Authority says the price doubles to $1.50 and $7.50 per 1M tokens then. My read: that date is the real test of whether developers who adopted at the introductory rate stay on.
- Apple’s China launch. The rollout was expected in the coming months as of the August 14 report. The thing to check is which model actually powers it: Apple’s own, Qwen, or both under the dual-track setup.
- Anthropic and Decart. Fortune’s sources said talks could fall through. Watch for any confirmation from either company, and whether a deal lands before Anthropic’s initial public offering.
- Muse Glimmer’s compression claim. Meta says quantizing to roughly 4-bit precision, with everything fitting in a 24 GB or 32 GB envelope, brings minimal to no degradation on agentic tasks. I think independent benchmarks on consumer cards are what will make or break the local-agent pitch.
This article was written by AI from the sources listed on this page. Figures, names and quotes were checked against those sources by software before publishing.