Google shipped a cheaper model for agent loops.
NVIDIA open-sourced its GPU Kubernetes stack.
▲ AI DEV SIGNALS
by Marktechpost · November 17, 2026

Hi there,
"Does my AI provider keep my data?" is now a standard question in security reviews, and it's often answered wrong. The real answer depends on which endpoint you call, which flags you set and what you upload. This week we explain the 2 API designs behind data retention. Also: Google priced a frontier-class agent model at $1.50 per 1M input tokens, Alibaba shipped a 2.4T MoE, and India got a local inference cloud.
01 · Featured: zero data retention applies to endpoints, not vendors
Sponsored by Fireworks AI
Fireworks AI explains on how every AI API follows 1 of 2 designs: upload, use, delete, or send, compute, return. Which one you're using decides who manages your data's lifecycle. Teams that assume a vendor-wide "ZDR" label covers them are often wrong.
How it works: In pattern 1, you upload a file or dataset for batch jobs or fine-tuning. It stays stored until you delete it, so deletion is your job. In pattern 2, you send a request and get a result back. Nothing stays on the platform, and your app owns the history. Zero data retention only applies to pattern 2.
The numbers: OpenAI keeps abuse-monitoring logs for up to 30 days by default. Its batch files expire after 30 days, and Anthropic keeps Message Batches for 29 days. Anthropic can keep flagged content for up to 2 years. Stored interactions on Google's Gemini API last 55 days on the paid tier. Fireworks runs ZDR by default for open models on Chat Completions, its Training SDK and Nexus.
The fine print: 1 flag can switch you from one pattern to the other. Google's Interactions API sets store to true by default. At OpenAI, setting store to false doesn't give you ZDR; that needs prior approval. /v1/files and /v1/batches aren't ZDR-eligible. Amazon Bedrock's "none" mode blocks models that require retention.
Our take: Treat ZDR like an endpoint setting you check yourself, not a promise on a vendor's slide. List every call you make under pattern 1 or pattern 2. Then write the delete call for every upload. Read the full breakdown.
02 · 4 more drops: a cheaper agent model, a 2.4T MoE, and India-hosted tokens
Google calls it its "daily driver for agentic workflows," priced at $1.50/$7.50 per 1M tokens with a 1M-token window. It scores 57% on Terminal-Bench 4.1, vs 18.2% for Gemini 3.8 Flash. The catch: it's API-only, ranks #5 on Agent Arena, and synchronous output caps at 65,536 tokens.
It has native vision and costs $1.50 per 1M input tokens. Alibaba says it cuts Work360 cost per task by 62%. The catch: unlike most Qwen releases, this one is proprietary.
NVKE is the control plane for a managed Kubernetes service on your own bare-metal GPUs. NVKR builds the images it runs, with signatures and SBOMs. Both are Apache 2.0 and conformance tested on Kubernetes 1.36. The catch: after install, the NVKE API serves plain HTTP with authentication off.
It serves Sarvam 105B, GLM 5.3, DeepSeek V4.1 Flash, Kimi K3.1 Pro and Gemma 4 31B from GPU clusters in India. Everything runs through 1 OpenAI-compatible API, and it's now open to everyone. It's also another way to decide where your data lives.
03 · Signals: decision models multiply, dominoes beat frontier models, and Rubin doubles up
Shanghai AI Lab ships Sage S2.1-Flash. It's a 321B MIT-licensed scientific multimodal model with an FP8 variant. It self-reports 90.3 on MLE-bench Lite, vs 89.2 for DeepSeek-V4.1-Flash.
Liquid AI's d1.2-ultra tops the AA Decision Index. It reads text, images and audio and returns calibrated probabilities with 0 output tokens. It scores 75 on Decision Index v2.1.
Perplexity open-weights Decider v1.3. A new Locate feature adds span extraction. It reports 91.6% mean accuracy on 63 query-classification tasks.
Frontier models fail at falling dominoes. DominoBench has 854 items in 427 minimal pairs. Top models get at most 34% of pairs right unaided, and reach 82–92% when told which calculation to run. Data on Hugging Face.
Vera Rubin NVL72 debuts at MLPerf Training v6.1. NVIDIA reports up to 2x per-GPU performance vs GB300 NVL72. DeepSeek-V3 training dropped from about 16 to about 4 minutes at 97% scaling efficiency.
Stanford's Memorilla squeezes a corpus into 16 tokens. It feeds a frozen Qwen3-8B and reaches 84.33% on FactKG, using 108x fewer document tokens than full context.
Meta's JiTTests caught 74 privacy SEVs before landing. The LLM writes tests designed to fail on risky diffs. Engineers report a 60% true-positive rate.
Cmpnd's dspy-wsp edits model weights without gradients. A frontier LLM edits a small model's weights in a sandbox and tests every change. Each run costs about $5 to $10.
OVHcloud announces Olala Mini. It's a 7B MoE with about 1B active parameters, mixing Differential-TPA and Mamba-3 attention. It runs at 160 tok/s on an RTX A4500. Apache 2.0 weights are promised but not out yet.
InclusionMed scores frontier models from 17.5 to 49.1 out of 100. It combines 4 public clinical benchmarks, picked from 388 candidates. Every top-10 model did better at patient communication than at clinical decisions.
1 question
How many endpoints in your stack could you sort into pattern 1 or pattern 2 without opening the docs?
Have a great Tuesday.
until next week,
the marktechpost team
REACH 70K+ AI BUILDERS
Was this email forwarded to you? Sign up here
