Ling 3.0 Flash combines 124B total parameters with 5.1B active per token, and Artificial Analysis scored it 38 at 334 output tokens per second. Its $0.07/$0.22 per million input/output-token pricing is attractive, but the independent measurements do not validate the model card's broader benchmark claims.
OpenAI says preliminary internal evaluations of the upcoming Astra model leave Critical cybersecurity capability in scope under its Preparedness Framework. It implemented universal monitoring, isolated testing, restricted access, and paused work that did not meet stronger requirements; Astra was not involved in the earlier Hugging Face incident.
AMD agreed to acquire Taalas for an undisclosed amount, aiming to integrate its specialized inference silicon into AMD's accelerator roadmap and pair it with Instinct GPUs. The deal has not closed and remains subject to customary conditions and regulatory approvals.
Together ran 900 rollouts across 113 DeepSWE tasks with hidden-test grading. DeepSeek V4 Flash reached 53.3% pass@1 at $0.10 per rollout versus GPT-5.6 Luna's 67.2% at $0.61; sending DeepSeek failures to Luna raised success to 78.9% at $0.385 per attempted task. This is a first-party provider evaluation.
Excited to see bitwise train/inference parity on TorchTitan RL + vLLM extend to linear attention ✨ Gated DeltaNet's recurrent kernel is batch-invariant: one sequence's own state, fixed order, no cross-sequence reductions. Logprob diff = exactly 0 ✅ prefix caching works as-is.
vLLM reports exactly zero log-probability difference between TorchTitan RL training and vLLM inference for Gated DeltaNet. Its recurrent kernel keeps each sequence's state and order independent of the batch, and existing prefix caching still works; the update does not report the throughput cost of this extension.
32 posts · 86% confidence
No themes have been published for this stream yet.
No must-reads have been published yet.
The preliminary July release showed nonfarm payrolls down 23,000, unemployment at 4.1%, and labor-force participation at 61.4%. May was revised down by 66,000 and June by 37,000, cutting the two prior gains by 103,000 combined; later revisions can still change the July estimate.
The initial August 7 release put July goods trade at CNY4.66tn, up 19.2% year over year. Exports rose 17.8% to CNY2.71tn and imports rose 21.2% to CNY1.95tn; the figures are nominal yuan values, with no revision disclosed in the release.
USTR said S&P Global's August 5 reports for gallium, germanium, tungsten, antimony, and NdPr will inform negotiations over border-adjusted critical-mineral price floors. The statement sets no floor levels or implementation timetable.
232 posts · 92% confidence