The AI log, agents and coding · open weights · frontier
17 Nov 2026
MAI Code 2, SenseNova-Loop-9B, sparse attention, Gemini tasks
Microsoft ships MAI Code 2 models, SenseNova opens a 9B looped model, NVIDIA maps sparse attention costs, and Gemini gets a Task mode.
By Ali Asif, published
What shipped
- Microsoft AI
Introducing MAI Code 2 and Code 2 Flash
Two coding models in GitHub Copilot; Microsoft reports MAI Code 2 at 98% of GPT 6.1 Sol's average benchmark score at 40% of the cost, with costs shown at a 50% promotional price.
- SenseNova
sensenova/SenseNova-Loop-9B
An Apache-2.0 image-text-to-text model with about 9B parameters, a looped middle block and a 512K-token context, reporting 59.55 on Terminal-Bench 2.1.
- NVIDIA
Co-Designing Sparse Attention for Efficient Long-Context Inference
Sweeps the indexer knobs in sparse attention and ends with eight design guidelines, including a Top-k method 2.1x to 2.4x faster than radix select on DeepSeek V3.2 and V4 decode.
- Google
Put Gemini to work for you
Tasks in Gemini, a Task toggle in the app running on Gemini 4 Hydrogen, is free in six countries and can pay through Google Pay and AP2 with a spending cap.
Four items today, all with something concrete to read. Every number below comes from the pages I opened. I have not run any of it. OpenAI also posted about GPT-6.2 Astra and Sol, but its page would not load for me, so I am leaving it out until I can read it.
MAI Code 2 and Code 2 Flash
Microsoft announced two coding models at Ignite, and both are available in GitHub Copilot. MAI Code 2 is aimed at enterprise work such as deep debugging, multi-file implementation and repo-wide reasoning. Code 2 Flash is for high-volume jobs like explanations, targeted edits and routine fixes. Microsoft says plainly that neither is tuned for very long-horizon agentic work.
The headline claim is an average over DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1: MAI Code 2 reaches 98% of GPT 6.1 Sol's score at 40% of its cost, and 96% of Claude Opus 5.5's score at 5% of its cost. In a blind frontend evaluation with 258 annotators, reviewers preferred its output over GPT-6 Sol 64% of the time. The cost figures are shown at a 50% promotional price, and all models ran at max reasoning effort. Flash is in limited preview in Microsoft Foundry.
SenseNova-Loop-9B
This open model has about 9B parameters and an inference-equivalent size of about 14.6B. It has 32 decoder layers but runs 56 layer executions per forward pass, because layers 4 to 27 form a block that loops twice. It takes text and images, outputs text, and supports a 512K-token context under Apache 2.0.
The card reports 59.55 on Terminal-Bench 2.1, 69.90 on GAIA and 65.31 on AIDABench. On NL2Repo it scores 41.36 against 44.60 for Qwen3.8-27B. With 4-bit weights of about 4.9 GB, a 256K context needs about 15 GiB with an 8-bit KV cache. The repo is gated, so you accept access terms before downloading.
NVIDIA on sparse attention design
At 128K tokens, NVIDIA measures dense attention at 91% of a layer's prefill time. In DeepSeek V3.2 the indexer and Top-k take 47%, core attention 24% and the feed-forward block 25%. The post sweeps indexer settings and finds, for example, that raising key heads from 1 to 8 makes decode 7.6x slower, while doubling head dimension costs only 1.1x to 1.2x in prefill. Its Guess-Verify-Refine Top-k in TensorRT LLM is 2.1x to 2.4x faster than radix select on DeepSeek V3.2 and V4 decode.
Tasks in Gemini
Google added a Task toggle to the Gemini app, running on Gemini 4 Hydrogen. It handles multi-step jobs such as booking, calling businesses and shopping, connects to more than 60 third-party apps, and can pay through Google Pay and AP2 with a hard spending cap. It is free in the U.S., Japan, Korea, Canada, Australia and India. Google notes tasks can fail, and some sites hand the final click back to you.
What I'd watch next
Independent runs of MAI Code 2 outside Microsoft's own benchmark set, and whether the Loop-9B numbers hold once people can load the gated weights.