Qiskit SDK v2.6
Qiskit SDK v2.6 adds C API pass-manager framework, QPY serialization, and custom operations.
The Deploy Log | Free, no signup
Breakthroughs
45 rows in Breakthroughs in The Deploy Log, newest first. They are listed from to , each by its publisher's date or, where the record carries none, its edition's date. Every item sourced to the publisher's own page.
MLX on Mac Studio
Latest MLX runs Muse Spark Open (816B) on Mac Studio, generating over 36 tokens per second on four Mac Studios.
AstaBrief 8B
Ai2 open-sourced AstaBrief 8B, a model that turns a research question and retrieved literature excerpts into a cited report, with Fast mode averaging 51.1 seconds per report compared with 178.5 seconds for Thinking mode, about 3.5x faster.
Olmo-core 3
Ai2 released Olmo-core 3, an open MoE training system benchmarked at over one trillion total parameters, with a 47-billion-parameter MoE processing 52,000 tokens per second per GPU on eight NVIDIA B300 GPUs, about 2.7x the throughput of the earlier implementation.
Skala 1.1
Skala 1.1, trained on 2.5x more data than its predecessor, achieves a weighted average error of 2.8 kcal/mol on GMTKN55 and is now available in CP2K with integrations underway for Psi4, FHI-aims, ORCA, and VASP.
Aurora 1.5
Aurora 1.5 adds 22 new weather variables, hourly temporal resolution, and probabilistic ensemble forecasting, outperforming ECMWF ENS on 88.9% of evaluated variable-and-lead-time targets.
LeRobot v0.6.0
LeRobot v0.6.0 ships world model policies VLA-JEPA, FastWAM, and LingBot-VA, new VLAs including GR00T N1.7 and MolmoAct2, reward models Robometer and TOPReward, and six new simulation benchmarks under lerobot-eval.
MolmoAct 2
MolmoAct 2, Ai2's vision-language-action model, is available with model weights, data, training code, and fine-tuning scripts, and won an embodied AI hackathon with a voice-controlled robot ten days after release.
SensorFM health model
SensorFM is a foundation model for wearable health pre-trained on over one trillion minutes of sensor data from five million people, transferring to 35 health prediction tasks.
Flint visualization language
Flint is a visualization intermediate language that compiles compact specs into polished charts across Vega-Lite, Apache ECharts, and Chart.js, with higher LLM-judge scores than DirectVL across three models.
Show the other 35 rows
TabFM
TabFM, a zero-shot foundation model for tabular data, is now available on Hugging Face, GitHub, and within Google Cloud BigQuery.
MolmoMotion
MolmoMotion forecasts 3D point trajectories from video, instructions, and query points, and a control policy built on it succeeds on 76.3% of pick-and-place tasks versus 56.0% for Molmo 2.
Earth AI vectorized dataset
Google Research released a vectorized dataset mapping hedgerows, stone walls, and copses across England, using a deep learning framework.
olmo-eval
Ai2 released olmo-eval, an evaluation workbench that builds on OLMES and supports agentic and multi-turn evaluation, pairwise model comparison, and per-question analysis with standard errors and minimum detectable effects.
EMO
Ai2 releases EMO, a 1B-active, 14B-total-parameter MoE trained on 1 trillion tokens, where modular structure emerges from data, allowing selective use of just 12.5% of experts with near full-model performance.
Auto frame in Google Photos
Google introduced a new approach for editing images, now live in the Auto frame feature in Google Photos, allowing users to re-imagine photos from a new perspective after they have been taken.
ReasoningBank
Google introduced ReasoningBank, a novel agent memory framework that uses successful and failed experiences to distill generalizable reasoning strategies, enabling an agent to continuously learn from experience after deployment.
MoGen synthetic neurons
Training PATHFINDER with 10% simulated data from MoGen reduced reconstruction error on reserved mouse axons by 4.4%, equivalent to saving a single expert 157 years of manual work at the scale of a complete mouse brain.
Simula
Simula generated datasets of up to 512K data points per domain, and high complexity yielded a 10% accuracy gain in math reasoning on GSM8k while hurting performance in legal reasoning on LEXam.
Vibe Coding XR
Google Research's Vibe Coding XR uses Gemini with the XR Blocks framework to translate prompts into physics-aware WebXR apps in under 60 seconds, with a one-shot success rate around 70% initially, now improved after 11 releases.
S2Vec
Google Research introduced S2Vec, a self-supervised framework that learns embeddings of the built environment, performing best for zero-shot geographic adaptation in socioeconomic prediction, but weaker on environmental tasks.
TurboQuant
Google Research introduced TurboQuant, a quantization algorithm that compresses key-value cache to 3 bits without accuracy loss, achieving up to 8x performance increase over 32-bit unquantized keys on H100 GPUs.
Groundsource flood dataset
Groundsource uses Gemini to extract 2.6 million flood events from news reports across 150 countries, enabling urban flash flood forecasts.
Urban flash flood forecasts
Google's Flood Hub now provides urban flash flood forecasts up to 24 hours in advance, using a model trained on news-derived data.
AMIE clinical study
A prospective study with Beth Israel Deaconess Medical Center found AMIE required zero safety stops and matched PCPs on diagnosis quality.
SpeciesNet wildlife identification
SpeciesNet is an open-source AI model that classifies 2,498 animal categories in camera trap images, trained on over 65 million labeled images, and it finds 99.4% of images containing animals with 94.5% of species-level predictions correct.
WAXAL speech dataset
WAXAL is a large-scale, openly accessible speech dataset covering 27 Sub-Saharan African languages spoken by over 100 million speakers, with approximately 1,846 hours of transcribed natural speech for ASR and over 565 hours of high-fidelity recordings for TTS, released under CC-BY-4.0.
DialogLab
Google Research introduced DialogLab, an open-source prototyping framework for authoring, simulating, and testing dynamic human-AI group conversations, evaluated with 14 participants who rated human control mode significantly more engaging.
Time-varying scheduling
Google Research introduced the first constant-factor approximation algorithms for non-preemptive throughput maximization under time-varying capacity, with a 1/2-approximation offline and a 1/11 competitive ratio online with common deadlines.
GIST algorithm
Google Research introduced GIST, a data subset selection algorithm with a provable guarantee of at least half the optimal value, outperforming benchmarks.
Intent decomposition
Google Research showed that decomposing intent extraction into screen summaries and then intent prediction lets small models match larger ones.
Smartwatch gait metrics
Google validated smartwatch-based gait metrics with 246 participants and 70,000 walking segments, showing strong validity (Pearson r>0.80) and reliability (ICC>0.80) for most metrics.
Hard-braking events as crash indicators
Google Research found hard-braking events from Android Auto correlate with crash rates, with 18x more road segments covered than crash data, and a case study showing a 70x higher HBE rate on a high-risk merge.
MedGemma 1.5
MedGemma 1.5 4B improves accuracy on CT, MRI, and histopathology classification, with gains like 14% on MRI findings, and is free for research and commercial use.
EV charging port prediction
Google Research deployed a lightweight linear regression model that predicts EV charging port availability and reduces bad predictions by approximately 20% in morning peak times and approximately 40% in evening peak times.
Real-time speech translation
Google Research introduced an end-to-end speech-to-speech translation model that enables real-time translation in the original speaker's voice with only a 2-second delay, now available in Google Meet on servers and as an on-device feature for Pixel 10 devices.
Generative UI
Google introduced generative UI, where AI models create interactive interfaces on the fly, rolling out in the Gemini app as dynamic view and in AI Mode in Google Search.
Decoded Quantum Interferometry
Google Quantum AI introduces Decoded Quantum Interferometry, a quantum algorithm that converts certain optimization problems into decoding problems and could solve OPI examples with a few million quantum operations versus over 10^23 classical operations.
Nested Learning
Google introduced Nested Learning, a new ML paradigm that treats models as nested optimization problems to mitigate catastrophic forgetting.
ForestCast benchmark
Google released ForestCast, the first deep learning-powered benchmark for proactive deforestation risk forecasting, using pure satellite data.
Gemini astronomy
Gemini achieved 93% accuracy across three astronomical datasets by learning from just 15 annotated examples per survey, and improved to 96.7% on MeerLICHT through an iterative human-in-the-loop process.
Google Earth AI
Google Earth AI introduces new Remote Sensing Foundations and Population Dynamics Foundations models, and a Gemini-powered Geospatial Reasoning agent that achieved 0.82 accuracy on a Q&A benchmark versus 0.50 for Gemini 2.5 Pro.
Private synthetic albums
Google Research introduced a method for generating differentially private synthetic photo albums using an intermediate text representation and hierarchical generation, tested on the YFCC100M dataset.
Quantum Echoes
Google Quantum AI demonstrated a verifiable quantum advantage with the Quantum Echoes algorithm on the Willow chip, where experiments took approximately 2 hours versus an estimated 13,000 times longer on a classical supercomputer.
DeepSomatic
DeepSomatic identifies tumor variants with higher accuracy, scoring 90% F1 on Indels with Illumina data versus 80% for the next-best method.
One line every time AI actually ships, around the clock, free on Threads and Bluesky. Every one is in The Deploy Log with its source.