PrismML fits 27B model on iPhone: Inside OpenAI’s first AI-powered consumer device

In today’s edition of Data Points, you’ll learn about our top headlines, and more:

• Claude models treat you differently by language
• Cognition’s SWE-1.7 boosts Devin’s coding power
• Nvidia’s Audex handles audio and text with one transformer
• OpenAI updates its prompting guide for GPT-5.6

But first:

Startup shrinks a 27 billion parameter model onto recent iPhones

PrismML, a Caltech spinout backed by Khosla Ventures, publicly released compressed versions of Alibaba’s Qwen model on Tuesday, shrinking it from 54 GB to under 4 GB so all 27 billion parameters run on an iPhone (15 or newer). The startup’s CEO told CNBC that Apple is currently evaluating the technology, though he characterized the discussions as very early. PrismML achieves the compression by reducing the models to ternary or binary quantization, reducing each value from 16 bits to just two or three possible values. Although they they lose a few percentage points of overall performance, particularly in factual recall, the compressed models use up to 14 times less memory, generate responses up to eight times faster, and consume three to six times less energy than conventional versions. For Apple, running more capable AI directly on iPhones would reduce latency, lower cloud costs, support privacy claims, and enable features to work offline, all especially valuable for health data and computational photography. Analysts cautioned that claims need real-world validation at scale, particularly around power consumption during continuous use and performance across millions of device combinations. (CNBC)

Report reveals details of OpenAI’s first hardware project

OpenAI’s first consumer hardware product will be a screenless, battery-powered smart speaker designed to be carried around and act as a humanlike AI companion in the home, according to Bloomberg. The device uses mechanical elements to shift and orient itself, paired with a camera and sensors so it can read its surroundings and context, and it can draw on personal data like emails to personalize its responses, using ChatGPT's capabilities for smart-home control, media playback, and messaging. The project traces back to OpenAI's $6.5 billion acquisition of Jony Ive's io Products, and the team includes numerous former Apple hardware and industrial designers who worked on the iPhone and Mac. Pricing is expected to land between $200 and $300, with a reveal targeted for later this year and a 2027 release. For developers, this signals OpenAI's ambition to move beyond software APIs into AI’s physical interaction, and perhaps a shift into voice, ambient sensing, and always-on context, drawing more on personal user information than general knowledge. (Bloomberg and TechCrunch)

Anthropic finds model values vary by language

Anthropic compressed more than 3,300 distinct values identified in prior research into four behavioral axes by analyzing 309,815 Claude conversations across three models and 20 languages. The axes (Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, Candor vs. Execution) measure how Claude responds to subjective questions with no universal right answer. The patterns are clear: Sonnet leans warm and deferential; Opus leans rigorous and cautious. Language produces even larger shifts. Claude expresses warmth-related values most in Arabic and Hindi, rigor-related values most in English and Russian. One practical implication is that two people asking the same model for business feedback, one in Hindi and one in Russian, may receive substantively different advice not because they’re using different models, but because the same model calibrates its character differently depending on the language it’s processing. The differences are small relative to conversation-level noise, but structured enough that Anthropic can trace them back to training decisions. (Anthropic)

Devin updates its internal low-cost coding model

Cognition released SWE-1.7, a code model trained on Kimi K2.7 that scores 42.3 percent on FrontierCode 1.1 Main, a leap from the previous version’s 9.4 percent. The company trained it across four data centers on three continents, syncing compressed weight updates through object storage every few gradient steps, with cross-continental rounds completing in one to two minutes for a one-trillion-parameter model. Three technical moves enabled this: Top-p sampling replay prevents entropy collapse during long reinforcement learning runs by excluding low-probability tokens from training. A self-compaction system lets the model summarize its working state mid-task and resume from those summaries, stretching rollouts to six hours despite a smaller context window. An alternating length penalty compresses responses on easier tasks while preserving long-horizon reasoning on hard ones. The infrastructure treats failures pragmatically; for example, inference engine crashes restart cheaply and statelessly, while trainer failures recover fast from local checkpoints, letting runs persist through continuous hardware problems at scale. The model runs at one thousand tokens per second in Devin via Cerebras. (Cognition)

Audex uses a single decoder for audio and text

Nvidia introduced Nemotron-Labs-Audex-30B-A3B (Audex), a unified audio-text large language model that handles audio and text tasks through a single shared transformer decoder. The model encodes audio into the text embedding space and treats both text and quantized audio tokens uniformly during generation, enabling seamless multimodal processing without requiring separate audio-specific components. Training involved 157.4 billion audio tokens and 320.5 billion text tokens across curated datasets, followed by reinforcement learning and distillation stages. Audex achieves state-of-the-art performance on audio understanding, speech recognition and translation, text-to-speech, and audio generation tasks while maintaining the reasoning and knowledge capabilities of its text-only backbone with minimal regression. The architecture’s simplicity makes it compatible with existing LLM training and inference systems, and Nvidia released model checkpoints for open research. (Hugging Face)

Short, clear prompts work best for most people on latest models

OpenAI published an updated prompting guide for GPT-5.6 Sol that inverts the advice from its GPT-5 playbook: stop writing long system prompts. The core shift is toward outcome-first prompting—define success criteria and constraints, then let the model work. The company backs this with internal data from coding-agent tests showing that leaner prompts improved evaluation scores by 10–15 percent while cutting tokens by 41–66 percent and costs by 33–67 percent. The old GPT-5 approach relied on scaffolding, including XML blocks, detailed templates, step-by-step narration. For GPT-5.6, these detailed instructions now act as noise the model has to parse around rather than helpful structure. For example, the new guide has a text verbosity parameter to control output length globally, and a section on programmatic tool calling to offload filtering and aggregation to code rather than asking the model to reason through it. The guide also warns against giving the model overlapping directions; GPT-5.6 wastes reasoning tokens trying to reconcile conflicting instructions rather than picking one and moving forward, making prompt clarity more critical than before. (Decrypt)

Want to know more about what matters in AI right now? 

Read the latest issue of The Batch for in-depth analysis of news and research. 

Last week, Andrew discussed the importance of agentic loops for efficiently building applications from scratch, the value of using coding agents iteratively  to refine product specifications, and the goal of treating human input as valuable “gold” in the development process. 

“As a software project matures, coding agents will work longer (maybe hours) to build to more complex specs, But at this early stage, it is fine to throw away the entire codebase and restart from scratch, so nothing is hard to change. All that code was cheap, in human time and token costs.” 

Read Andrew’s letter here. 

Other top AI news and research stories covered in depth:

• Fable’s Return and Fallout describes how Anthropic’s Claude Fable 5 was banned by the U.S. government and its subsequent return to the market.
• Google enhances its AI capabilities by pairing its Nano Banana update with Omni Flash’s API, integrating Gemini’s image and video models.
• DeepSeek’s DSpark Gains Velocity as the company open sourced a speculative decoding module that accelerates text generation while maintaining accuracy.
• Text Without Typing highlights how researchers at Meta and other institutions developed Brain2Qwerty v2 to generate sentences directly from brain waves.

Start today with DeepLearning.AI Pro!

DeepLearning.AI’s first-ever subscription plan for our entire course catalog includes foundational classics like the Machine Learning and Deep Learning specialization plus seminars on the latest tools and frameworks you need.

As a Pro Member, you’ll immediately enjoy access to:

• Nearly 200 AI short and long courses from Andrew Ng and industry experts
• Labs and quizzes to test your knowledge
• Projects to share with employers
• Certificates to testify to your new skills
• A community to help you advance at the speed of AI

Enroll now to lock in a year of full access for $25 per month paid upfront, or opt for month-to-month payments at just $30 per month. Both payment options begin with a one-week free trial.

Explore Pro’s benefits and start building today!

Try Pro Now

Data Points is produced by human editors with AI assistance.