AI Weekly #10 — speed wars, open models, and stolen thoughts
GPT-5.6 Sol hits 750 tok/s on Cerebras silicon, Meta drops a 30B local agent model, DeepSeek ships two updates, and researchers expose a reasoning-trace leak.
This week the model release cadence stayed brutal: OpenAI previewed 14× inference speedups powered by Cerebras, Meta quietly dropped a 30B open-weights model built for always-on local agents, and DeepSeek pushed two updates in as many days. Meanwhile a security paper showed that proprietary chain-of-thought traces aren’t as hidden as vendors assume, and Google shipped Gemini 3.7 Flash. The throughput and openness gaps are closing faster than anyone’s roadmap predicted.
GPT-5.6 Sol Ultrafast: 750 tokens/sec via Cerebras partnership
OpenAI previewed a new ‘Ultrafast’ API tier running GPT-5.6 Sol at up to 750 output tokens per second—roughly 14× the standard speed—powered by Cerebras wafer-scale chips. The tier is in preview and targets latency-sensitive agent workloads where generation speed is the bottleneck. Cerebras published a companion post detailing the hardware side of the integration.
Why it matters: If you’re building agents that chain multiple LLM calls, a 14× throughput jump materially changes what’s feasible in a single user-facing request. Watch pricing and availability before re-architecting anything. (Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed)
Gemini 3.7 Flash lands with a builder-focused release
Google DeepMind released Gemini 3.7 Flash, continuing the Flash line’s focus on speed and cost efficiency. The release follows the pattern of iterating the Flash tier more aggressively than the Pro/Ultra line. HN engagement was high, suggesting the developer community is paying attention to the Flash-tier price-performance curve.
Why it matters: Flash-class models are where most production API spend actually lives. A new Gemini Flash release means a fresh set of benchmarks to run against your workload before assuming GPT-5.6 Sol is the default. (Introducing Gemini 3.7 Flash)
Meta’s Muse Glimmer: 30B open-weights model designed for local agents
Meta released Muse Glimmer, a 30B-parameter model explicitly optimized for always-on local agent workflows, with multimodal capabilities and agentic task framing baked into training. It’s open-source and runnable locally, positioning it as a serious option for teams that want agent infrastructure without API egress. Zuckerberg also made public comments attacking closed-model competitors the same week.
Why it matters: A 30B model purpose-built for agentic use running entirely on local hardware changes the threat model for cost and privacy. If your agents are doing sensitive work, this is worth a benchmark run. (Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows)
DeepSeek V4 Pro 0813: another mid-cycle update worth tracking
DeepSeek pushed V4 Pro 0813 to OpenRouter just days after V4 Flash 0731 appeared on ARC Prize leaderboards. The rapid iteration cadence—two distinct model updates in under two weeks—signals DeepSeek is treating continuous improvement of production models as a deployment strategy. Simon Willison flagged both; HN gave V4 Pro over 1000 points.
Why it matters: DeepSeek’s update pace means capabilities you benchmarked last month may be stale. If you’re using DeepSeek in production or as an eval reference, add their model changelog to your weekly check. (DeepSeek V4 Pro 0813 (on OpenRouter))
Researchers demonstrate extraction of hidden reasoning traces from LLM APIs
A paper and accompanying site (stolen-thoughts.com) showed that chain-of-thought reasoning traces from proprietary LLM APIs can be recovered even when vendors suppress them from the visible response. The attack exploits timing and token-probability side channels rather than any direct API leak. Simon Willison covered it; HN gave it 694 points.
Why it matters: If you’re building on top of a reasoning model and your prompts or intermediate thoughts contain sensitive business logic, assume those traces are not fully hidden. Treat the reasoning scratchpad as potentially observable. (Stealing Reasoning Traces from Proprietary LLM APIs)
Hugging Face’s ICML reproduction study: 2,200 papers, hard lessons
Hugging Face published findings from an effort to reproduce 2,200 papers from ICML 2026, surfacing systematic patterns in what reproduces cleanly versus what doesn’t. The scale of the project makes it one of the largest ML reproducibility audits to date. Results have direct implications for which benchmark claims engineers should weight when evaluating models.
Why it matters: Reproducibility failure at scale means a non-trivial fraction of published benchmark numbers are unreliable. Before you adopt a model based on a paper’s claims, this study is useful calibration for how skeptical to be. (What We Learned by Reproducing 2,200 papers from ICML)