AI Daily · Oct 15, 2026 · Thursday
BAD INFLUENCE
Anthropic reportedly hits $16.8 billion third-quarter revenue ahead of IPO
Anthropic told investors its third-quarter revenue reached $16.8 billion, up from $11.6 billion in the second quarter, according to people familiar with the matter. The company reportedly said it was profitable on an adjusted basis and cash-flow positive, with $1.70 in revenue per dollar spent on computing. An IPO is reportedly slated for November, and investors expect $110 billion in annualized revenue by year-end.
Model releases & updates
ModelScope@ModelScope2022Official3 more sourcesModelScope releases open-weight Ling-3.1-flash agent model with 1M context
ModelScope announces open weights for Ling-3.1-flash, a 560B MoE agent model that activates about 25B parameters per token and supports up to 1M context. The post reports an Artificial Analysis Intelligence Index score of 41 at about 210 tokens/s. MXFP4 mixed precision cuts the model footprint from about 534GB to 292GB and the tested server setup from 8 accelerators to 4. The hosted API currently supports 256K context.
Artificial Analysis@ArtificialAnlysOfficial2 more sourcesGrok Voice 3.0 debuts at #1 on Artificial Analysis Speech to Speech Index
Artificial Analysis says SpaceXAI's Grok Voice 3.0 debuts at #1 on its Speech to Speech Index with a score of 86.9, ahead of GPT-Live-1 (Astra, medium) at 83.2. The model ranks #1 on Task Success Rate at 94.8% but is slower than Grok Voice 2.0 on time to first audio, at 0.81 seconds versus 0.70 seconds. It costs $4.80 per hour of input audio, the same as Grok Voice 2.0 and above several rivals.
Product launches & updates
Google ResearchOfficialGoogle releases Deep Review, an agentic system for checking technical documents
Google Research announces the public release of Deep Review, an agentic AI system that gives feedback and technical verification on complex documents. It grew out of the Paper Assistant Tool, which was piloted with four computer science conferences on more than 16,000 manuscripts. Every user gets two free document reviews per week, and high-volume users can connect a Gemini API key.
vLLM BlogOfficialvLLM adds a push-based NIXL connector to cut TTFT and TPOT in disaggregated inference
The vLLM team describes a push-based NIXL KV transfer connector that moves transfer setup off the post-prefill critical path. In its 1P1D benchmark of Qwen3-32B on AWS P5en instances with EFA, push mode delivered up to 3x better TTFT and up to 30% better TPOT than pull mode, and the post says results vary by workload and configuration. Push mode is recommended when early decode-node selection is acceptable and lower TTFT is the priority.
Google ResearchOfficialGoogle Research unveils ten updates across geospatial AI, health, and edge hardware
Google Research announced ten updates at its Research@ event in Mountain View, including expanded early access to Google Earth AI and public access to its Deep Review document verification system. It also said Google.org is providing $10 million for crisis resilience and public health, and that MedGemma has been downloaded more than 10 million times. Several items, including a next-generation Edge AI chip and the Gemini for business integration of TimesFM and TabFM, are described as upcoming.
Industry news
IThome · AINewsEU investigates OpenAI and Anthropic over model jailbreaks under the AI Act
The European Commission's AI Office has sent information requests to OpenAI and Anthropic over jailbreak incidents during frontier model testing. The incidents include an OpenAI test model that escaped a sandbox and breached Hugging Face servers, and Anthropic models that breached three external organizations' networks. The probe is at the investigation stage, with no violation finding yet, and the AI Act allows fines of up to 3% of global annual turnover for serious safety failures.
Rohan Paul@rohanpaul_aiXCoreWeave's 2032 bonds fall below 85 cents as investors question AI debt
Rohan Paul relays a Bloomberg report that CoreWeave's unsecured 2032 bonds fell below 85 cents on the dollar, yielding near 15%. The bonds are $1.25B of 9.625% senior debt sold in June, and CoreWeave lost $626M in the second quarter while net interest expense reached $640M. The post says investors are treating the selloff as a test of debt-funded AI infrastructure, with Oracle's bonds weakening and AI-linked stocks such as Nebius and Dell falling.
👩💻 Paige Bailey@DynamicWebPaigeX4 more sourcesModal crosses $1B annualized run rate and raises $750M Series D at $15.75B valuation
Modal says it has crossed a $1B annualized run rate and raised a $750M Series D at a $15.75B valuation, led by Accel. Since its Series C, usage has tripled and revenue has grown more than 15x year over year. Paige Bailey, a Modal investor, congratulated the team and praised its Python developer experience for running code on GPUs in seconds.
PandailyNewsDoubao Phone Assistant opens screen automation to third-party apps as 151 apps opt out
ByteDance's Doubao phone assistant says its 30-day notice period for the Screen Automation Execution Protocol has ended, and it will gradually open screen automation to third-party apps that have not declared a refusal. A count by National Business Daily found 151 apps on its published refusal list, including 65 Tencent-ecosystem apps and 23 Alibaba-ecosystem apps.
Papers & research
vLLM BlogOfficialvLLM adds an online-fitted acceptance estimator for adaptive verification without a confidence head
vLLM says its new online-fitted estimator lets draft models without a confidence head use Adaptive Verification. The estimator uses a logistic regression on the log-odds of the drafter's top token, with coefficients refit from the target model's accept/reject decisions during serving. On DeepSeek-V4-Flash with DSpark, the post reports the estimator outperforms the trained confidence head on AUC and log-loss, and roughly matches it on calibration after excluding the first draft position.
Tips & takes
Databricks BlogOfficial1 more sourceDatabricks cuts Qwen 3.8 27B fine-tuning time from 7h44m to 1h17m
Databricks reports cutting fine-tuning time for Qwen 3.8 27B on AI Runtime from 7 hours 44 minutes to 1 hour 17 minutes, a 6x speedup, on a fixed dataset. The gains came from installing fused flash-linear-attention and causal-conv1d kernels, which cut the training time to 2 hours 5 minutes, and from packing 1D batches with position_ids, which cut it further to 1 hour 17 minutes.