Models AI News

Google Adds Agentic Video Understanding to Gemini, Claiming 88% Lower Token Use

The API feature shifts Gemini from fixed-rate video sampling to targeted inspection of frames, audio and transcripts. Its value for long recordings will depend on whether Google’s reported efficiency and accuracy gains carry into production workloads.

Yesterday

Open release Models

Phonely Opens Alma, a Voice Model Trained on 10 Million Calls, to Outside Builders

The new external offering turns Phonely’s own call traffic into a model-development pitch, combining voice-specific training with company-reported speed and pricing advantages. Whether those results transfer to other companies’ call flows remains the practical test.

Yesterday

Last week

10 stories

Open release Models

Gnani Launches 30B Evon Model and Self-Hosted Agent Stack for Indian Institutions

The open-weight release couples Indian-language support with tools intended to keep sensitive data inside a customer’s own infrastructure.

Aug 28, 2026

Benchmark Models

Meta’s EvoHarness-RL Takes Qwen3-8B to 96.9% on ALFWorld, Near Claude Opus 4.5

The result is a strong benchmark showing for learned state management, not yet evidence that the framework improves the long-running enterprise workflows it is designed to target.

Aug 28, 2026

Research Models

JAMA Paper Says Autonomous AI Could Surpass Physicians and AI-Assisted Care by 2030

The contested forecast recasts clinical AI from a physician tool into a possible replacement, while real-world patient use and medical training remain unsettled.

Aug 28, 2026

Cohere’s Parse 5 Turns Enterprise Documents Into Markdown for $1.50 per 1,000 Pages

The compact vision model is built for high-volume retrieval and document-processing pipelines, but its headline benchmark result leaves out chart and visual-grounding tests.

Enterprise adoption Models

Model-Serving Platforms Reach 6.1% of AI-Spending Businesses, While Frontier Spend Holds

The early shift is showing up in specialized deployments and model-serving platforms, while direct business spending still favors Anthropic and OpenAI.

Benchmark Models

Databricks Says 300M Chart-JSON Pipeline Tops Four Multimodal Baselines on Answer Accuracy

The company’s test suggests chart values can become a useful retrieval layer instead of relying solely on page-image embeddings, though its comparison measures answer correctness and includes a synthetic benchmark it built.

Research Models

Mammogram AI Found Prior Stroke at 86%, but Clinical Use Still Needs Validation

The study points to a possible way to extract cardiovascular signals from breast scans already taken in routine care. Its reported results are promising, but the path to clinical use still runs through accuracy, reliability and false-result reduction.

Z.ai Says 100,000 Chinese Chips Serve GLM-5.3-Flash; Shares Rise More Than 8%

The low-cost model is a test of whether China-made hardware can support a public AI service at scale, but Z.ai has not named the chipmakers behind the system.

Aug 26, 2026

Google Puts Gemini 3.5 Transcribe in Preview With Live and Recorded-Audio APIs

The release turns speech recognition into a Gemini product layer for developers and Google surfaces, but a product-specific price remains unavailable for teams weighing production use.

Aug 26, 2026

DeepSeek V4 Pro Lands on Fireworks With $2.50 CyberGym Solves, but Kimi K3 Scores Higher

The provider’s results frame security-agent selection as a three-way choice among raw solve rate, cost per completed task, and whether a model will reliably execute the required tool workflow.

Aug 25, 2026

Enterprise adoption Models

August 2026

5 stories

Ant Puts Its FX Forecasting Model Into Tools Used by Major Banks

The model targets a specific treasury problem in cross-border payments: predicting the timing, size and currency of cash needs. Its reported bank integrations are meaningful, but their scale—and one headline accuracy measure—remain unclear.

Generalist’s GEN-1.5 Lets Robots Try a Task After Watching Once

Generalist’s new model treats a demonstration as temporary instruction, potentially reducing task-specific training—but its reported one-shot results remain far from dependable execution.

Benchmark Models

Stanford Won Databricks’ Agent Cup, but 18.8% of Questions Stumped Every Team

The live contest tested whether agents developed on one benchmark could handle an unfamiliar Treasury archive. Stanford led the field, but the unsolved questions show the limits of reliable document reasoning.

Open release Models

Meta’s 30B Muse Glimmer Tries to Make 131K Context Fit in 24 GB

Quantized weights and a narrow attention cache make Meta’s local-hardware claim plausible. But fitting a 131,072-token prompt is different from processing one quickly, and runtime implementation will decide whether the memory savings appear in practice.

DeepSeek Gives V4-Flash a Separate Vision API With a 384-Token Image Cap

The experimental model gives developers several familiar ways to submit images and a fixed upper bound on image-token use. It leaves the harder questions—visual accuracy, latency and reliability—for users to test.