Models AI News

Yesterday

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. Google has launched agentic video understanding for Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite.

Open release Models

Yesterday

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. Phonely has launched Alma, a language model built for voice agents and trained on more than 10 million phone conversations.

Last week

10 stories

Open release Models

Aug 28, 2026

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. Gnani has launched Artha, a sovereign AI stack aimed at Indian enterprises and public institutions. It combines Evon v3.3, a 30-billion-parameter open-weight language model supporting more than 11 Indian languages, with Plexus, a platform for building enterprise AI agents.

Benchmark Models

Aug 28, 2026

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. Meta AI and the University of Illinois Urbana–Champaign say their EvoHarness-RL framework let Qwen3-8B reach 96.9% average success on ALFWorld, edging Claude Opus 4.5’s reported 96.4% score. Source: venturebeat.com.

Research Models

Aug 28, 2026

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. The clinical-AI debate is moving beyond whether software can help doctors. The case for autonomy The authors reviewed AI-in-medicine studies published from January 1, 2024 onward.

Aug 27, 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. Cohere has released Parse 5, a 2.3-billion-parameter vision-language model that converts enterprise document pages into structured Markdown for $1.50 per 1,000 pages. Those limits define the boundary of Cohere’s pitch.

Enterprise adoption Models

Aug 27, 2026

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. The share of AI-spending businesses paying for platforms that offer open-source and Chinese-developed models rose to 6.1% in July. Ramp’s AI Index put the share at 4.5% in January 2026.

Benchmark Models

Aug 27, 2026

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. The company plans to add the extraction automatically to ai_parse_document, turning chart values into structured text without changing that tool’s interface. Source: databricks.com.

Research Models

Aug 27, 2026

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. Mammography is designed to find breast cancer, but researchers in Israel say its images also hold signals associated with cardiovascular conditions.

Aug 26, 2026

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. Z.ai says its new GLM-5.3-Flash model is handling every online request on 100,000 China-made chips, putting a public model service on wholly domestic hardware at substantial claimed scale. Its Hong Kong-listed shares rose more than 8% after the launch.

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. Google has released Gemini 3.5 Transcribe in public preview, giving developers separate APIs for real-time voice interaction and recorded-audio processing. But Google’s pricing page does not list a price for Gemini 3.5 Transcribe.

Aug 25, 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. Fireworks AI has made DeepSeek-V4-Pro-0813 available on its platform, pairing the release with security-benchmark results that favor the model on cost and tool reliability rather than outright solve rate.

Enterprise adoption Models

August 2026

5 stories

Aug 22, 2026

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. The deployment puts AI into decisions over how much cash to prepare for cross-border payments and which currencies to hold.

Aug 22, 2026

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. Generalist AI has released GEN-1.5, a robot foundation model that can use one brief demonstration to attempt some physical tasks without changing its underlying parameters. Language, sensor data and the robot’s position information.

Benchmark Models

Aug 22, 2026

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. Stanford won Databricks’ inaugural Grounded Reasoning Cup with 63.3% accuracy, but 18.8% of questions were unsolved by every team.

Open release Models

Aug 22, 2026

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. The design puts the main constraint of long-running agents in focus: not just model weights, but the memory retained for every prior token.

Aug 22, 2026

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. DeepSeek has launched deepseek-v4-flash-vision-exp, an experimental model for requests that combine text and images.