Tencent lists Hy4 as available through TokenHub, WorkBuddy, CodeBuddy and other company products. Its announcement describes new command-line and graphical computer-control capabilities, including training on Blender.
Model weights, the numbers learned during training, and system cards documenting the model are promised in the coming days.
Standard API prices are approximately $0.90 per million input tokens, $2.70 per million output tokens and $0.045 per million cached input tokens. Tokens are the units of information a model processes and generates.
Tencent advertises half-price TokenHub access through December 31. WorkBuddy and CodeBuddy are free November 4–15.
The lower-cost Hy4 Flash is still scheduled for a future release, despite a section heading calling it recently released.
This is the full Hy4 release following the August preview, with newly announced visual and computer-control capabilities.
Caveat: The performance and safety comparisons in the announcement are Tencent’s internal results, rather than independent evaluations.
(Techmeme discovery · Tencent release announcement)
Early reporting counts restrictions passing in at least 14 municipalities. Across Ohio, 18 municipalities voted on 21 data center measures.
In Trenton, roughly 80% supported banning data centers using more than 25 megawatts of electricity, a threshold for large facilities.
In Perry, more than 87% supported requiring financial guarantees from developers seeking tax incentives. Wilmington voters rejected project-related rezoning in four separate referendums.
The local reporting distinguishes jurisdiction from project cancellation: Trenton’s ban applies inside city limits, while an adjacent proposed campus could face separate township rules.
Caveat: These are early local results with some municipalities still unreported, and passing a restriction does not automatically stop every existing or nearby project.
(Techmeme · Fortune, cross-checked against Cincinnati Enquirer reporting)
An agent is a model running a multi-step task with tools. The method rewrites earlier reasoning while preserving user instructions, program requests and results.
After filtering 72 candidate histories to 58, rewrites shortened later reasoning in 151 of 172 comparable continuations. The main analysis retries three capped continuations and one provider error.
On a fixed 50-task coding sample, six of eight model-and-compressor configurations cost less. Only two reduced combined model-request time after including rewriting.
DeepSeek V4 Flash with external rewriting cost $1.59 versus $2.43 across those tasks. Both settings solved 32 of 50, while combined request time rose from 3.37 to 3.45 hours.
Those time sums exclude tools and other execution work. The released code and frozen records expose the comparisons and accounting.
Caveat: Costs exclude abandoned attempts, and the small fixed task sample with provider changes leaves preservation of task quality unproven.
(arXiv 2611.03852 · full paper, accounting appendix and released records)
A judge model is a model grading another model’s output. Five judges from three providers assessed 4,000 human-voted answer comparisons covering nine systems and 33 system pairs.
Each comparison was judged in both answer orders. Agreement with human system rankings varied sharply across judges, even though they saw the same answers.
Haiku and Gemini Flash-Lite each matched the human conclusion on 22 of 23 statistically clear system comparisons. Requiring their agreement preserved the same shared mistake.
Requiring agreement with GPT-6 Luna left conclusions for only eight of those 23 comparisons. Seven of the eight matched the human conclusion.
The released analysis bundle includes settings and deviations. Luna’s reasoning was disabled, larger Claude judges used low effort, and a planned sixth judge was dropped after partial results were seen.
Caveat: Each reference label is one person’s preference, so agreement measures alignment with those voters rather than independently verified answer correctness.
(arXiv 2611.04090 · full paper, saved judgments and disclosed design changes)
Measuring local model speed under your actual workload#
The review covers downloadable models with nine to 20 billion parameters, the numbers learned during training. It compares lower-precision model storage with the usual 16-bit baseline.
In an older Qwen2.5-14B test on one A100, four-bit storage raised generation speed from 46.30 to 98.02 tokens per second. It used vLLM 0.6.3 model-serving software, one request, a one-token prompt and 2,048 output tokens.
In a separate Gemma 3 12B test on an H100, processing 256 requests together took 14.07 seconds at 16-bit precision versus 13.81 with four-bit storage. Each request had 256 input and 256 output tokens.
Only two reviewed model builds had both serving timings and accuracy results. Their accuracy test covered one philosophy subject, and one build was adjusted using questions from that subject.
The reporting checklist asks for the exact model revision, software, hardware, workload, baseline, quality-test settings and calibration data. Calibration means adjusting compression using example inputs.
Caveat: This is a review of older public measurements checked through October 5, not a new controlled comparison of current serving software.
(arXiv 2611.04380 · full paper, original Qwen measurements and PyTorch model card)
Google’s November 3 clarification says schools can individually enable Gemini, Gemini Notebook and Search, choosing one, two or all three.
Its education announcement says access will default to off for school-managed kindergarten-through-twelfth-grade student accounts until administrators enable it. More details will precede rollout in the coming months.
Guided Learning, which coaches students through problems, became globally available in Classroom on November 3. A Search pilot starts in selected US schools later this month.
Google says Search’s AI Mode and AI Overviews will provide guided learning. Its education chief told reporters that Search and AI are deeply integrated.
Caveat: The clarification promises separate product controls, not an AI-free version of Search.
(Techmeme · CNBC, Google announcement and full company statement)
The Agent Board coordination paper → THIN_GUARD: relevant design, but its small pilots also find a plain external watcher cheaper; the selected studies offer stronger immediate testing guidance.
The engineering-output essay → LOW_UTILITY: its pull-request increase spans different employers, team sizes and processes, so it supplies no controlled estimate of AI’s productivity effect.