New AI Models, Tools, and Cloud Updates You Can Use Today

Model releases you can start testing

NVIDIA Gemma 3 n series

NVIDIA just opened its Gemma 3 n multimodal family. Each model accepts text, images, video, and audio, and it supports 30+ languages out of the box. The smallest variant runs on a single RTX 4090 with about 7 GB VRAM, while the largest needs 48 GB of memory. You can pull the weights from the NVIDIA NGC catalog and run inference with the nvidia‑model‑serve CLI.

# Install the server
pip install nvidia-model-serve
# Pull the 7B multilingual model
model-serve download gemma-3n-7b
# Start a local endpoint
model-serve serve --model gemma-3n-7b --port 8080

Microsoft MAI Code 1 Flash

Microsoft released MAI Code 1 Flash, its first model built specifically for code generation. The model runs on Azure NC T4 GPUs and claims a 30 % cost reduction compared with using generic OpenAI endpoints. You can access it through the new azure‑ai‑code Python package.

from azure_ai_code import CodeClient
client = CodeClient(endpoint="https://codeflash.azure.com")
response = client.generate(prompt="Write a fastapi endpoint that returns the current UTC time")
print(response.text)

Google DeepMind Gemini 3.5 and Omni

DeepMind announced Gemini 3.5 and Omni. Gemini 3.5 adds chain‑of‑thought reasoning that improves performance on logic puzzles by 12 %. Omni merges vision and language into a single transformer, letting you ask questions about a video clip without separate models. The models are available via the deepmind‑api beta.

import deepmind_api as dm
model = dm.GeminiOmni()
result = model.analyze(video="demo.mp4", query="What objects are moving in the foreground?")
print(result)

Open‑source LLM benchmark

Fireworks AI published a benchmark that compares DeepSeek v3.2, Kimi K2.5, and Qwen3 VL on latency, token cost, and accuracy on the MMLU suite. For a 2‑GB GPU, DeepSeek v3.2 costs $0.0008 per 1 K tokens, while Qwen3 VL is $0.0012 per 1 K tokens but scores 4 % higher on reasoning tasks. Use these numbers to decide which model fits your budget.

---

Developer tools that speed up your workflow

DeepEval – LLM evaluation framework

DeepEval lets you run automated quality checks on your LLM prompts. Define a YAML test suite, point the tool at your API key, and it will report precision, hallucination rate, and latency.

# tests.yaml
- name: factual‑recall
  prompt: "Who founded OpenAI?"
  expected: "Sam Altman"
- name: math‑add
  prompt: "What is 123 + 456?"
  expected: "579"

deeval run --suite tests.yaml --model gpt-4o

The report outputs a CSV you can feed into your CI pipeline.

Amazon Q for AWS development

Amazon Q adds an AI‑powered assistant directly into the AWS console and CloudShell. You can ask it to generate CloudFormation snippets, debug Lambda errors, or rewrite a Step Functions state machine.

Tip: When you type q generate s3 bucket my‑data‑store --public, Q returns a ready‑to‑paste CloudFormation block.

Integrating these tools

All three tools expose HTTP endpoints, so you can chain them. For example, generate code with MAI Code 1 Flash, validate it with DeepEval, then let Amazon Q provision the required resources.

---

Cloud and infrastructure updates you should know

Google AI Studio now supports Kotlin

Google AI Studio added Kotlin bindings, letting you write Android AI features without leaving the language you use for UI. The new ai‑studio‑kotlin library includes a ChatModel class that streams responses directly to a Compose UI element.

val model = ChatModel("gemini-3.5")
model.send("Summarize this article:") { response ->
    textView.text = response.text
}

Apple Core ML 2 for iOS

Apple released Core ML 2, which adds on‑device quantization and a new MLComputeDevice API for selecting GPU or Neural Engine at runtime. The update reduces model size by up to 40 % and improves inference latency on iPhone 15 Pro by 22 %.

let config = MLModelConfiguration()
config.computeUnits = .all // use GPU + Neural Engine
let model = try MyModel(configuration: config)
let result = try model.prediction(input: inputData)

Why these changes matter for you

• Multimodal models let you add image or audio understanding without stitching separate APIs.
• Evaluation frameworks catch regressions before they reach production.
• Cloud‑native assistants like Amazon Q keep you in the console, saving context switches.
• Native language support (Kotlin, Swift) reduces glue code and improves performance.

Ready to try any of these? Check out our AI services or see recent projects in the portfolio.

---

Stay ahead of the curve by integrating the newest models, testing tools, and platform features into your stack today.