AI efficiency
Latest stories covering AI efficiency, with summaries and links to related coverage.
2 stories · 2 reports · page 1 of 1
1 story · 1 report in the last 7 days
Latest report · Counts as of (UTC)
Models in this coverage
Up to six tracked models linked to this topic’s reports in the last 30 days. Counts measure coverage on Themata, not usage.
- Qwen3.8 27B
Qwen
· Open weights
A 27-billion-parameter vision-language model for text, images, and video. Supports coding and multi-step tasks, with optional thinking and configurable reasoning effort.
1 story · 1 report on this topic
- Qwen3.8 Flash
Qwen
· Open weights
An open-weight preview of Qwen’s next architecture, combining sparse and linear attention. A vision-language model designed to make long-context agent workloads more efficient.
1 story · 1 report on this topic
No established topics share multiple stories with this topic yet.
Latest stories
Gemma-4-26B-A4B with 39-69% less thinking at same accuracy
research
8 min read
Bottlecap AI released ThinkingCap-gemma-4-26B-A4B-it, a token-efficient version of Google’s Gemma-4-26B-A4B-it reasoning model. Gemma-4-26B-A4B is a sparse mixture-of-experts model with 26 billion parameters and about 4 billion active parameters per token. Bottlecap says its model uses 38.9% fewer thinking tokens on average across 12 benchmarks while accuracy declined by 0.15 percentage points, from 75.18% to 75.03%, under identical vLLM serving and sampling settings. Thinking-token reductions ranged from 13.7% on IFBench to 69.1% on LongBench v2. The model used 54.4% fewer tokens across long-context and multimodal tasks, 44.4% fewer on knowledge and reasoning, 36.1% fewer on math, and 27.6% fewer on code. Accuracy increased on some tests, including IFBench and LongBench v2, but fell 4.00 percentage points on AA-LCR. Bottlecap measured looping in 0.76% of the optimized model’s responses, down from 3.64% for the base model. Bottlecap evaluated both models with thinking enabled, temperature 1.0, top-p 0.95, top-k 64, and a 65,536-token generation cap. The open model and GGUF builds are available under the bottlecapai/ThinkingCap-gemma-4-26B-A4B-it repository. Bottlecap also offers a commercial PRO version, which it says cuts thinking tokens by up to 87.6% and 62% on average.
Timestral: Mistral release time-series foundation models
tool
7 min read
Mistral has released Timestral, a family of four time-series foundation models for probabilistic multivariate forecasting. The models range from 5 million to 209 million parameters, accept multivariate targets plus past-only and known-future covariates, and produce forecasts across 21 quantiles. Mistral says its 209M-parameter model ranks first among pretrained zero-shot models on the GIFT-Eval, TIME and fev-bench public benchmarks, using a single checkpoint without ensembling. Timestral is designed for uses such as sensor monitoring, capacity planning, energy, supply chains, finance and predictive maintenance. It can jointly model related signals, use categorical inputs such as machine states, and incorporate information known in advance, including production schedules, holidays or weather forecasts. The models have a 32,768-step context length and can run inference on CPUs: Mistral reports that forecasting one day from three weeks of 15-minute data takes 10 milliseconds for the 5M model and 140 milliseconds for the 209M model on an eight-thread server CPU. Training uses mostly synthetic data, followed by post-training on mixed real and synthetic data. Mistral offers zero-shot use, fine-tuning on proprietary data and deployment through Mistral Forge, including on-premises, VPC and air-gapped environments. A GitHub repository is available under a modified MIT license.