AI engineering, November 12, 2026

Top News

Gemini 4 Argon is now generally available, giving you a new Google model to evaluate for complex coding and agentic tasks. Google positions Argon as its most capable model for those workloads, as well as math and science.

Here's what the launch covers:

• For complex coding, Google positions Gemini 4 Argon as its most capable model in that category.
• Agentic tasks are part of the same release, alongside the model's coding focus.
• Math and science round out the tasks Google highlights for its most capable model.

Why care? You can now access a generally available model that Google positions as its most capable option for complex coding, agentic tasks, math, and science.

Try it: Use model ID gemini-4-argon.

Sources: ai.google.dev

Top News · 344 HN points

Need classification scores without generating an answer? Gemini Decisions API scores candidate options in one forward pass and is available through the Gemini API in Google AI Studio, with a no-cost tier for prototyping.

Here's what you get:

• Input costs $0.10 per million tokens, with $0 output-token charges. Shared input is billed once per request, while question instructions and criteria add token costs.
• Responses include probabilities and confidence scores. Google reports 150-250 ms end-to-end latency.
• Requests accept text, images, audio, video, and PDFs, with up to 1,000 questions and 256K input tokens.
• Google says the API is nearly five times cheaper than Gemini 3.5 Flash-Lite with the Interactions API across its shared benchmark suite.

One catch: The nearly flat latency result covers up to 100 questions and 16K-token contexts. Google Cloud support is coming soon.

Try it: Get an API key in Google AI Studio and use gemini-decision-1 with the Decisions API.

Sources: blog.google

Top News · Reported · tomshardware.com

MEXT is designed to reduce server DRAM requirements by moving cold memory pages to flash. AMD has partnered with HPE to bring it to customers, starting with selected EPYC Venice servers carrying a “MEXT Ready” badge and a dedicated NVMe drive or bay.

Here's what changes:

• AMD and MEXT's founder say DRAM needs can fall by 20-75%, depending on the application and performance target.
• MEXT predicts when pages are needed and returns them to DRAM. Offloaded memory appears to the OS as installed DRAM.
• AMD's internal five-agent test retained 99% of full-server performance at 50% memory capacity and 85% at 25%.
• The Linux kernel module trains in under a second at deployment, uses one EPYC CPU core without a GPU, and keeps training as workloads run.

One catch: MEXT is EPYC-exclusive, and the license rollout with RAM View has no start date. Performance and suitability vary by workload.

Sources: tomshardware.com

1. Runway releases Praxis-1, ranked #1 among open-source robot foundation models on RoboLab-120
2. vLLM adds metadata-only requests for DeepSeek V4.1 Flash to avoid repeated media preprocessing
3. ExecuTorch runs PyTorch models on browser GPUs, decoding 2-10x faster than its CPU path
4. NVIDIA releases NVFP4-quantized Kimi-K3.1-Pro, a multimodal model with 2.8T total parameters · 8 Likes
5. Open-HumanWAM releases 50,000 hours of annotated first-person human video for robot pretraining
6. Unsloth cuts FLUX.2 and FLUX.3 decoding VRAM use by 36-43% without reducing speed · 5 Stars today