|
|
MODEL
MAJOR
2026-09-15
Gemini 3.8 Live — Google's voice models talk while they think
Google's new Live API models hold a real-time voice conversation and keep reasoning in the background while they speak.
What is it?
Gemini 3.8 Live replaces gemini-3.1-flash-live-preview as Google's audio-to-audio model for the Gemini Live API. It ships in two flavours: a base model for cheap, fast voice agents, and an Extended Thinking variant for harder tasks.
How does it work?
The Extended Thinking variant runs its reasoning in parallel with speech, filling silence with natural spoken acknowledgements so tool calls happen in the background and the audio stream never stalls. Language detection happens mid-conversation across 97 languages.
Why does it matter?
Voice agents usually break in the pause between a question and an answer. Gemini 3.8 Live Extended Thinking scores 82.6 on the Artificial Analysis Speech to Speech Quality Index and 97.7% on Big Bench Audio, so support flows can use a stronger model without sounding slower.
Who is it for?
Developers building voice agents and phone support flows.
|
|
|
|
TOOL
MAJOR
2026-09-16
Firefox Smart Window runs on Mistral Small 4 — and opens in France
Mozilla's browser assistant adds Mistral Small 4 as a model option and opens its beta to France.
What is it?
Mistral Small 4 now powers Firefox Smart Window, the AI assistant built into Mozilla's desktop browser, with the beta now open in France alongside the U.S. and Canada. Smart Window groups tabs, replays history visually, compares options, and summarizes long pages.
How does it work?
Smart Window supports three built-in models and lets you bring your own; Mistral agreed to zero data retention for all traffic it handles, and the Memories feature keeps context on your own device rather than Mozilla's servers.
Why does it matter?
The deal puts a European, open-weight-focused lab in front of Firefox desktop users without requiring a product switch. France gets official French-language support for the first time, with the UK and Germany named for later in 2026.
Who is it for?
Firefox desktop users in the U.S., Canada, and now France.
|
|
|
|
MODEL
MAJOR
2026-09-14
Atria Dawn Preview — the report behind Shanghai AI Lab's 744B agent model
The 744B open-weight agent model that showed up on GitHub with no announcement now has a 140-author paper explaining how it was built.
What is it?
Atria Dawn Preview is Shanghai AI Laboratory's 744B mixture-of-experts model with a 256K context window, released under MIT with no blog post — now documented by a 140-author arXiv report covering 16 benchmarks.
How does it work?
Training used a Verifiable Experience Pipeline that connects tool-mediated interactions to executable environments, so the model learns from actions that can be checked rather than text alone. The GLM-5.2 MoE backbone supports FP8 checkpoints for reduced memory.
Why does it matter?
An open-weight agent at this scale under MIT terms is rare. Atria Dawn leads on DeepSearchQA (96.0 vs 95.9 for Kimi K3) and BFCL v4 tool use (77.0 vs 74.1 for GLM 5.3), matching closed API quality you'd otherwise pay for.
Who is it for?
ML engineers and researchers running self-hosted agents.
|
|
|
|
MODEL
MAJOR
2026-09-15
Salesforce Koa — a CRM reasoning model built on NVIDIA Nemotron
Salesforce Koa is a CRM-focused reasoning model that Salesforce post-trained from NVIDIA's open-weight Nemotron 3 Super.
What is it?
Koa is Salesforce's first reasoning model, post-trained from NVIDIA Nemotron 3 Super (120B open-weight) on synthetic data from 27 years of CRM deployments across 14 industries — no customer data used.
How does it work?
Training runs supervised fine-tuning then GRPO reinforcement learning on a simulation-to-reward pipeline: Agent Script task specs expand into persona-conditioned multi-turn conversations, and the reward is resolved tasks with correct tool calls, not fluent wording.
Why does it matter?
Multi-turn tool use is where enterprise agents break, and that's exactly where Koa beats its Nemotron base on CRM Bench (0.86 vs 0.84). It shows an application vendor specialising an open-weight model rather than buying a frontier API.
Who is it for?
Salesforce admins and enterprise agent builders; in pilot inside Agentforce now, GA expected winter 2026 in U.S. regions.
|
|
|
|
MODEL
MAJOR
2026-09-15
Jev — TypeSafe's model returns typed decisions, not text
A model that answers typed questions with calibrated probabilities instead of writing prose you have to parse.
What is it?
Jev is TypeSafe AI's first System One Model: send it unstructured state plus typed questions, and it returns typed answers — Choice (pick one), Score (rate against levels), or Noul (yes/no probability) — each with a calibrated confidence score your code can act on directly.
How does it work?
Trained with Reinforcement Learning for Calibrated Decisions (RLCD) rather than RLHF, Jev answers in 70–500 ms and can handle many questions at once because no answer is written token by token.
Why does it matter?
At $0.042 per million input tokens with output free, triage, routing and review tasks that were too expensive to run on every record become cheap enough to run on all of them — and the returned value already matches your type, with no JSON-parsing guesswork.
Who is it for?
Backend and automation engineers building decision pipelines.
|
|
|
|
TOOL
MAJOR
2026-09-15
Claude Code 2.1.273 — a subshell could hide a dangerous rm in bypass mode
Claude Code 2.1.273 closes two permission-checker gaps and stops the context meter double-counting advisor-tool turns.
What is it?
Two permission bugs are fixed: Bash commands the checker can't fully analyze no longer silently skip the prompt, and a subshell can no longer hide a dangerous rm while bypass mode is on.
How does it work?
The checker now falls through to ask rather than allow unanalyzable lines. A 2.1.268 change that blocked eval and env -C outright is also reverted, restoring commands like time -p make build.
Why does it matter?
Teams running in bypass mode or with blockReadsOutsideWorkingDirectories were relying on a guard with a hole. A second fix corrects the context meter, which was firing auto-compact at about half the real context window.
Who is it for?
Claude Code users and platform admins — update immediately if running bypass mode.
|
|
|
|
MODEL
MAJOR
2026-09-14
PhysBrain 1.5 report — an open 8B model tops 28 embodied AI tests
One open 8B network reads a scene, plans the arm's next move, and predicts what the scene looks like a second later.
What is it?
PhysBrain 1.5 folds three robotics jobs into one open model: scene understanding, motion generation, and future-state prediction. DeepCybo published 2B and 8B checkpoints on Hugging Face under an open license alongside the technical report.
How does it work?
Built on a Qwen3-VL backbone, it encodes language, end-effector motion, and visual targets as discrete tokens over one shared vocabulary, trained end-to-end with next-token prediction from human interaction videos — no separate perception, control, or world-model components.
Why does it matter?
Averaging 72.5 across 28 embodied benchmarks puts it just under GPT-6-Astra (73.3) and Gemini 3.6 Flash (73.0) — the best open-source result reported — so robotics labs can download weights instead of renting a closed API.
Who is it for?
Robotics and embodied-AI researchers; runs under Transformers, vLLM, or SGLang.
|
|
|
All releases at ai-tldr.dev
Simple explanations • No jargon • Updated daily
|
|
