🚨 Six releases. One theme. The machine stopped waiting

Presented by MongoDB  Event Venue Operator — an agentic demo with three-layer memory on Atlas. Open the repo →

Most weeks in AI are noise. This one is not. Every release below removed a wait — a robot waiting for a script, a caller waiting for a transcript, a GPU waiting for its slowest neighbor. That is the whole story. Here it is in six parts.

<300ms

PolyAI voice reply

32–92%

Gemini finger dexterity

99.4%

tokens saved, 233pp

01 · Google DeepMind · Robotics

Google DeepMind Ships Three Physical AI Models For Whole Body Control, Dexterity And Multi Robot Collaboration

A robot used to be a machine that repeated itself. Gemini Robotics 2 ships three models — a VLA, an embodied reasoning VLM, and an on-device VLA — and one checkpoint drives Apollo 2 with two different hands plus a Franka Duo gripper. Whole body control, not table-top. The robot walks to the shelf, then places the object.

The honesty is the best part: multi-finger dexterity still runs 32% to 92% depending on the task. Unscrewing a bulb is solved. Tying a trash bag is not.

Read on Marktechpost →  · Announcement  · ER 2 model card  · Hugging Face dataset  · GitHub samples

02 · PolyAI · Voice AI

PolyAI Releases Dialog-RSN-1: An Audio-Native Dialog Model That Fuses Turn-Taking, Speech Recognition, Function Calling, And Response

Every voice stack throws away the thing that matters. The ASR hands over its best guess, and the tone, the hesitation, the doubt — all gone before the model reads a word. Dialog-RSN-1 listens to the audio itself.

Turn-taking is the first output token: EMPTY, ONGOING or COMPLETE. TTS stays separate, so you still own the voice. Sub-300ms responses, +11% relative containment at a restaurant group, −37% latency at an insurer. English only, and only through PolyAI — no open weights, no public API.

Read on Marktechpost →  · Technical blog  · Early access

03 · Tencent · Inference

Tencent Open-Sources AngelSpec: A Unified Training Framework for MTP and Block-Parallel Speculative Decoding on Hy3 Models

Everyone hunts for the one drafter that wins the average benchmark. Real traffic is not an average. Open-ended chat and code are different problems, so AngelSpec ships two drafters instead of one compromise — autoregressive MTP for conversation, block-parallel DFly for code and math.

On Hy3-A21B, DFly reaches 4.79 mean accepted length against 3.69 for DFlash and 3.00 for MTP. On HY3-295B-A21B at TP=8, DFly-8 delivers a 1.98–2.40× speedup across concurrency 4 to 64. Seven checkpoints, one config-driven pipeline, six draft architectures.

Read on Marktechpost →  · Paper  · GitHub  · Docs  · Hugging Face  · ModelScope

Sponsored · MongoDB

Event Venue Operator: an agentic demo you can actually run

Most agent demos are a screenshot. This one is a repo. An AI operator agent works a tennis tournament on Day 6, rain ninety minutes out, and has to decide what to do for two very different visitors.

Under it: three-layer memory — long-term, short-term, shared — on a single MongoDB Atlas cluster. Hybrid retrieval with Voyage AI multimodal embeddings. A LangGraph chain that runs perceive → plan → hitl_gate → act → reflect. Optional Langfuse traces. MIT licensed.

GitHub repo → · Live demo · Architecture docs

04 · Moonshot AI · Training infra

Moonshot AI Open-Sources MoonEP: A Perfectly Balanced Expert Parallelism Library for MoE Training

A collective runs at the speed of its slowest rank. So the hottest expert sets the price of every training step, and the shapes change every step, and your memory fragments. MoonEP makes one hard promise instead: every EP rank receives exactly S × K tokens, no matter how skewed the router is.

It gets there with redundant experts planned online on-GPU, then prefetched before expert compute. Static shapes kill the per-layer host sync. MIT licensed, shipped on Kimi K3 Open Day alongside FlashKDA and AgentEnv.

Read on Marktechpost →  · GitHub  · Announcement  · Benchmark script

05 · Liquid AI · Open weights

Liquid AI Releases LFM2.5-Encoder-230M and LFM2.5-Encoder-350M: Bidirectional Encoders That Stay Fast at 8K Context on CPU

Nobody demos an encoder. Everybody runs one — under the classifier, the intent router, the safety filter, the PII detector. All day. Usually with no GPU. These two carry 8,192 tokens of context, which is roughly 13 to 15 pages: one contract, one patient record, one forward pass.

The 350M posts an 81.02 mean across 17 tasks and ranks fourth of 14 — the three models above it are all larger, the top one nearly ten times its size. At 8K on CPU, the 230M runs a forward pass in about 28s against over 90s for ModernBERT-base. LFM Open License v1.0, 15 languages, evaluation harness open under Apache-2.0.

Read on Marktechpost →  · Liquid AI blog  · Hugging Face write-up  · 230M  · 350M  · Eval code

One more thing…

Marktechpost · Open source

Meet Token Saver: An Open-Source MCP Extension Using Local Hybrid RAG to Cut Claude PDF Token Costs 90-99%

Drop a 200-page PDF into a chat and you do not pay once. You pay again on every follow-up, forever, because the history gets re-sent. Token Saver ends that. The PDF never leaves your machine.

A local MCP server runs hybrid retrieval — BM25 over SQLite FTS5 weighted 0.4, plus a local all-MiniLM-L6-v2 embedding at 0.6 — and hands Claude only the passages that answer the question, with page numbers attached. On a 233-page filing: 133,349 tokens down to 740. That is 99.4%. No Python, no terminal. One .mcpb file, MIT licensed. Built at Marktechpost.

Read on Marktechpost →  · MCP Bundle format  · all-MiniLM-L6-v2

Sponsor AI Dev Signals

Put your product in front of the people who build with it.

This list is engineers, researchers, and founders who read a release note before they read a press release. No banner farms, no interstitials. One sponsor per issue, written in your voice, placed where people are already reading.

Newsletter placements, article sponsorship, product launches, GitHub and Hugging Face repo promotion, and webinars. Tell us the goal and we will send the media kit and available dates.

See one in action — this issue’s placement for MongoDB →