Apple and Meta have enabled Meta's 816-billion-parameter Muse Spark Open model to run locally on clustered Mac Studios. CNBC reported that each 256GB M5 Ultra machine costs at least $10,799, so a two-machine cluster comes to $21,598 or more.
Perspective Coverage
32 publishers
- Builder
- Builder 39%
- Operator
- Operator 31%
- Investor
- Investor 30%
Reality
- Evidence58
- Adoption
- Not scored
- Hype gap+35
- Incentives70
- Confidence68
Shoppers at Nvidia's RTX Spark launch told PCMag they were holding off over the $2,599 price
Over 100 people lined up at a San Francisco Best Buy for Nvidia's RTX Spark launch, but PCMag says it looked like only a handful bought one. Many came for free swag or out of curiosity, so the line says little about demand at $2,599.
Reality
- Evidence35
- Adoption
- Not scored
- Hype gap
- Not scored
- Incentives
- Not scored
- Confidence40
Google will roll out an opt-in Android Messages setting in the coming weeks that has Gemini assess the intent of unknown senders. It adds a cloud check on top of existing on-device detection, so the new layer reaches only people who switch it on.
Perspective Coverage
6 publishers
- Builder
- Builder 21%
- Operator
- Operator 67%
- Investor
- Investor 12%
Reality
- Evidence40
- Adoption
- Not scored
- Hype gap+15
- Incentives
- Not scored
- Confidence72
Reviewers can judge RTX Spark's chip today but not the laptops around it until October 16
Reviewers of NVIDIA's RTX Spark may publish chip results now but not laptop-specific verdicts before October 16, Windows Central reports. GPU performance varies by laptop, it says, so early scores belong to the machines tested.
Reality
- Evidence35
- Adoption
- Not scored
- Hype gap+20
- Incentives
- Not scored
- Confidence55
MacPaw launched Eney for Mac and iPhone, an assistant that says it spots what needs attention and asks for approval before it acts. 9to5Mac credits local models now capable of complex tasks, though demanding work still goes to the cloud.
Perspective Coverage
3 publishers
- Builder
- Builder 30%
- Operator
- Operator 47%
- Investor
- Investor 23%
Reality
- Evidence40
- Adoption
- Not scored
- Hype gap+20
- Incentives65
- Confidence55
MIIT filings completed September 29 list three AI systems approved for iPhones in China, two of them Alibaba's Tongyi models. Of Apple's three global tiers, TechTimes reports, only the on-device one stays Apple's own.
Reality
- Evidence50
- Adoption
- Not scored
- Hype gap+20
- Incentives
- Not scored
- Confidence55
Google DeepMind and Singapore's GovTech announced a Gemma-based scam detector on October 14 that runs on the phone and reads whole conversations. It has no public release date, and Google has not named the messaging apps that will support it.
Perspective Coverage
3 publishers
- Builder
- Builder 34%
- Operator
- Operator 33%
- Investor
- Investor 33%
Reality
- Evidence35
- Adoption3
- Hype gap+25
- Incentives
- Not scored
- Confidence70
buildConfirmed4 publishers Liquid AI says its LFM2.5-24B-A2B scores 48.09 on LiveClaw against 38.19 for the larger Qwen3.6-35B-A3B. The 184 tok/s it reports on an Apple M5 Max is a decode rate, so agent latency depends on how many tokens the model spends thinking.
Perspective Coverage
4 publishers
- Builder
- Builder 45%
- Operator
- Operator 33%
- Investor
- Investor 22%
Reality
- Evidence42
- Adoption
- Not scored
- Hype gap+30
- Incentives65
- Confidence62
Seven of eight Surface Laptop Ultra configurations lack the memory in Microsoft's local coding test
Microsoft's on-device coding model peaked at 75.5GB of memory at full context in the company's own test, according to a Command Line post. Only the 128GB model, at $5,899 and not yet on sale, has more, TweakTown reported.
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+55
- Incentives60
- Confidence50
Microsoft makes Execution Containers for Windows 11 AI agents generally available
According to CRN, Microsoft's Execution Containers for securing AI agents on Windows 11 are now generally available, and Intune policies manage them. Windows teams now have a supported place to govern agents, though the report says little about what a container actually isolates.
Reality
- Evidence28
- Adoption
- Not scored
- Hype gap+35
- Incentives65
- Confidence35
Tencent open-sources a 33-language translation model that fits on a device in 440 MB
Tencent has open-sourced its Hy-MT2 translation models for 33 languages, with the smallest compressed to 440 MB to run on devices. The claim that the small model beats Microsoft's API is Tencent's own, so teams paying per call have a candidate to test.
Reality
- Evidence30
- Adoption
- Not scored
- Hype gap+35
- Incentives70
- Confidence40
buildOne report1 publisher Google DeepMind's EmbeddingGemma 2 embeds four modalities with 740M Apache-licensed parameters
Google DeepMind's EmbeddingGemma 2 maps text, images, video and audio into one vector space using 740M Apache-licensed parameters. It is built to run on phones and laptops, so the case for a hosted embedding API now depends on whether its quality holds up on a buyer's own data.
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+15
- Incentives60
- Confidence50
buildOne report1 publisher EmbeddingGemma 2 maps five modalities into one 768-dimension vector space
Google DeepMind released EmbeddingGemma 2, a 740M-parameter open model that runs on a phone and embeds text, code, images, video and audio in one space. Teams running a separate embedder per modality can consolidate on it if Google's reported benchmark numbers hold on their own data.
Reality
- Evidence50
- Adoption20
- Hype gap+25
- Incentives60
- Confidence55
AI firms want 2028 phones to run models more than three times what Qualcomm's 2nm chips handle
Qualcomm CEO Cristiano Amon says AI firms want phones running 100-billion-parameter models continuously by 2028, more than triple his newest chips' limit. Most of Qualcomm's plan for closing that gap is memory, the cost Amon says already deters phone upgrades.
Reality
- Evidence35
- Adoption
- Not scored
- Hype gap+25
- Incentives75
- Confidence40
Physicists' formula for when chatbots turn bad got 15 of 16 small-model tests right
George Washington University physicists published a formula predicting when chatbots turn bad, after it called 15 of 16 preprint test cases. That evidence comes from six small open models and 300-token chats, a narrow base for the offline safety monitor the authors propose.
Reality
- Evidence40
- Adoption
- Not scored
- Hype gap+30
- Incentives
- Not scored
- Confidence45
George Washington University formula predicts when a chatbot's attention tips into harmful output
George Washington University researchers say a formula predicts how many good outputs a chatbot gives before its first bad one, across seven open-weight models. Their proposed check sits inside the model, so offline AI companions on personal devices, with no cloud filter and no way to patch weights, would need it before they ship.
Reality
- Evidence35
- Adoption5
- Hype gap+35
- Incentives
- Not scored
- Confidence35
Google has released AI Edge Foresight, a Granola-style Mac meeting note-taker that works fully offline on its on-device EmbeddingGemma 2 model. The Verge reported it is free and experimental, so teams barred from cloud transcription can pilot it cheaply.
Perspective Coverage
3 publishers
- Builder
- Builder 27%
- Operator
- Operator 55%
- Investor
- Investor 18%
Reality
- Evidence55
- Adoption
- Not scored
- Hype gap+10
- Incentives40
- Confidence60
Microsoft's $2,600 Surface Laptop Ultra is a complicated, unproven bet on performance claims
Microsoft's Surface Laptop Ultra, its first RTX Spark laptop, runs on the Nvidia-MediaTek N1X ARM chip and starts at $2,600, Gizmodo reported. Its showcase game ran under emulation and its battery claim is untested, so for now it is a laptop to pilot before ordering in volume.
Reality
- Evidence35
- Adoption
- Not scored
- Hype gap+40
- Incentives60
- Confidence40
Asus launches ProArt laptops in Australia that run 120-billion-parameter AI models offline
Asus has launched ProArt P14 and P16 laptops in Australia whose 128 GB models can run 120-billion-parameter AI offline. Teams wary of how cloud services handle their data now have local hardware to price, though only the top memory tier supports that claim.
Reality
- Evidence35
- Adoption
- Not scored
- Hype gap+30
- Incentives70
- Confidence40
buildConfirmed13 publishers Google released EmbeddingGemma 2, an Apache 2.0 model that maps text, code, images, video and audio into one embedding space with 740M parameters. Phone apps get offline cross-media search from one set of weights, within limits set by a shared 8K-token window and lossy vector truncation.
Perspective Coverage
17 publishers
- Builder
- Builder 59%
- Operator
- Operator 29%
- Investor
- Investor 12%
Reality
- Evidence55
- Adoption22
- Hype gap+20
- Incentives70
- Confidence62
Earlier coverage
- Free RemoveMacAI tool deletes 12 GB of Apple Intelligence models from macOS 27
Security · October 5, 2026 · Confirmed7 publishers
- Microsoft drops the Copilot+ PC badge from Surface machines that still clear its 40 TOPS bar
Product · September 27, 2026 · Confirmed11 publishers
- Qualcomm's earbud silicon skips the phone and talks to the cloud itself
Product · September 23, 2026 · Confirmed8 publishers
- Paint writes a server-issued ID into locally generated pixels, and egress logs will never see it
Build · August 24, 2026 · Confirmed3 publishers
- Eufy bets an on-device chip beats the cloud round trip for home AI
Product · September 3, 2026 · Confirmed5 publishers
- Apple gates every new Siri AI feature to the iPhone 15 Pro and later
Product · September 11, 2026 · Confirmed2 publishers
- Every app on a Googlebook arrives through the Play Store or the browser
Build · September 21, 2026 · Confirmed6 publishers
- Motorola builds its Snapdragon Extreme flagship around a 200MP periscope camera
Product · September 22, 2026 · Confirmed7 publishers
- Qualcomm unveils two new 2nm chips: Snapdragon 8 Elite Gen 6 and 8 Elite Extreme Gen 6
Product · September 22, 2026 · Confirmed16 publishers
- Google prices Googlebooks from $899 for an October consumer launch
Product · September 21, 2026 · Confirmed22 publishers
- Google set 16GB of RAM as the floor on all five launch Googlebooks
Product · September 21, 2026 · One report1 publisher
- Google shipped Gmail on Googlebook OS as a Play Store web wrapper
Product · September 21, 2026 · One report1 publisher
- PrismML shrinks Qwen3.8 27B model to 5.9GB by storing each weight as +1, -1 or zero
Product · September 17, 2026 · Confirmed2 publishers