🚨 Meta's 4-bit 30B in 24 GB of VRAM, NVIDIA's 448 ms full-duplex turns, and an agent runtime that forks 5× faster than Docker

Presented by Mend.io  Securing AI Agents, MCP Servers & LLM Apps — a practical see-fix-protect framework for agents already in your codebase. Get the guide →

No benchmark decided anything this week. VRAM ceilings, license text and deployment targets did. Meta got a 30B under 20 GB with 4-bit weights. webAI shipped a 1.78 GiB checkpoint you may not deploy for revenue. NVIDIA published permissive weights and then tagged them research-only. ByteDance published no weights, no parameter count, and no API. Read the spec line before the eval table.

24 GB

Muse Glimmer VRAM envelope

448 ms

VoiceChat turn-taking latency

93.3%

Isaac RULER at 10M

01 · Meta · Open weights

Meta AI Releases Muse Glimmer: A 30B Open-Weights Agentic Model That Runs on One Consumer GPU

A 30B model used to be a data-center line item. At full precision Muse Glimmer needs over 55 GB. Meta compresses the weights to roughly 4-bit, which brings the language model under 20 GB, and the whole thing fits a 24 GB card — KV cache, perception encoder and drafter sharing the headroom. No network call.

Two quantized builds ship. K-Quant-17GB targets 24 GB VRAM at 1.0% average degradation across 15 benchmarks; K-Quant-Dynamic targets 32 GB at 0.2%. Speed comes from DFlash, a block-diffusion drafter predicting 16 tokens per forward pass: on an RTX 5090 throughput moves from 74.9 to 233.4 tok/s, a 3.1x speedup. It leads MCP Atlas at 75.5 against Gemma4-31B’s 54.2 and Qwen3.6-27B’s 62.5, and trails Qwen3.6-27B on OSWorld-Verified, 65.9 to 75.6. Apache 2.0.

Read on Marktechpost →  · Hugging Face  · Meta blog  · Docs

02 · webAI · Formal logic

webAI Releases TwIL-LM: A 1.7B and 3B Formal-Logic Model Family for Autoformalization on Local Hardware

The small-model pitch is usually cost. webAI’s is verification. TwIL-LM translates English into first-order logic and checks whether a conclusion follows from its premises — a job that wants a checker, not a conversationalist. The 3B Q4_K_M GGUF is 1.78 GiB and runs on CPU or 4 GB of VRAM.

The announcement frames the release around beating gpt-oss-120b on four of five formal-reasoning lanes. The shipped checkpoint trails it on the six-lane average, 0.4488 to 0.5192. The unambiguous win is throughput: 482-token generations and 32.9 answers per second against the 120B’s 4.2. Both checkpoints ship under the webAI Non-Commercial License 1.0 — revenue-generating deployment needs a separate agreement.

Read on Marktechpost →  · Model weights  · Announcement

03 · NVIDIA · Voice AI

NVIDIA Releases NemotronLabs VoiceChat 11B: An Open Full-Duplex Speech-to-Speech Model with ~450 ms Turn-Taking and Live Tool Calling

A voice agent has been three models in a trench coat: ASR, then an LLM, then TTS, every handoff adding latency and losing information. VoiceChat 11B does streaming speech understanding and speech generation in one hybrid Mamba/Transformer network. It listens while it speaks, so a user can barge in mid-turn and the agent yields.

Measured smooth turn-taking latency is 448 ms on Full-Duplex-Bench 1.0 at TOR 0.82, with a take-over rate of 1.00 at 480 ms on user interruption. It is the first open full-duplex model to call tools while the conversation keeps flowing, emitting on a side channel while an operator-defined on-hold line covers the silence. Weights are permissive OpenMDW-1.1, but NVIDIA labels the checkpoint research-only, caps it at five tools per session, and wants one 80 GB GPU. There is no hosted API.

Read on Marktechpost →  · Model card  · GitHub  · NGC container

Sponsored · Mend.io

Agent behavior isn’t defined by code alone.

Agents, MCP integrations, and LLM-powered apps are entering codebases faster than security programs can track them. This guide gives AI/ML engineers, platform engineers, and AppSec teams a practical see–fix–protect framework, with the checklists and templates to start today.

Inside: the agentic AI attack surface map, five layers from prompt injection to poisoned MCP tools. A 12-point misconfiguration checklist for agent and MCP permissions. Triage at AI speed — what to automate, what to keep human. Runtime guardrail architectures and prompt hardening. And an agentic AI maturity roadmap, self-assessed against NIST AI RMF, OWASP AIMA, ISO/IEC 42001, and the EU AI Act.

Download the white paper → · From Mend.io

04 · ByteDance Seed · Multimodal

ByteDance Seed Introduces SeedRealtime: a Native Audio-Visual Full-Duplex LLM That Watches, Listens and Speaks in One Model

Every other release here has somewhere to run. This one has nowhere, and that is why it is in the issue. SeedRealtime fuses audio, video and text in one end-to-end architecture and moves turn-taking inside the model, replacing the external voice-activity detector most real-time stacks still depend on.

The demos are the evidence. At a noisy group dinner it binds names to faces and keeps each voice tied to its identity. At a museum it holds an instruction and speaks up unprompted when the object enters frame. Watching an espresso workflow it interrupts when whole beans go into the portafilter. ByteDance’s own human eval reports pacing issues halved versus cascaded stacks — no benchmark, no latency numbers. It is live in the Doubao app. There is no technical report, no parameter count, no weights, and no announced API.

Read on Marktechpost →  · Launch post  · Seed models

05 · Pokee AI · Long context

Pokee AI Releases Pokee-Isaac 28B: A 10M-Token Context Agentic Model Built to Run Inside the Customer Boundary

Long context has been a cloud feature, which excludes everyone whose data is not permitted to leave the boundary at all. Isaac is a 28B text-only model with a 10M-token window, built to be deployed inside that boundary — VPC, on-premises, or on-device.

It stays above 93.3% on RULER at every tested length, ending at 93.3% at 10M, where GPT-5.6 Luna and Gemini 3.5 Flash Lite track it to 512K and then overflow at 1M. It leads BFCL v4 at 70.94 to Luna’s 70.61 — the report calls that parity, not a lead — and loses Terminal-Bench 2.1, resolving 56 of 86 text-compatible tasks against Luna’s 60. Weights are not published; deployment is licensed. Pokee publishes measurements only from a single B200-class GPU, so the single-GPU-from-an-RTX-4090 claim is vendor guidance, not a reported result.

Read on Marktechpost →  · Blog  · Technical report

One more thing…

Northeastern & Stanford · Agent runtime

Meet Shepherd: An Open-Source Python Substrate That Lets Meta-Agents Fork, Replay, and Revert Any Agent Run

A coding agent at step 10 is holding edited files, a running dev server, installed packages, and a warm prompt cache. It misreads a traceback and rewrites a file that was already correct. Now you have two bad options: patch forward, growing the context and the token bill, or restart, re-paying every model and tool call while reproducing nothing exactly, because runs are non-deterministic.

What you want is step eight. Git versions files, not a live process or a cache, so no existing runtime offers it. Shepherd records an agent run as a Git-like trace of typed events, where each commit covers the agent process and the filesystem together, copy-on-write. A branch therefore carries live state, not just files. Core operations are mechanized in Lean, and permissions are declared in the signature — a May[GitRepo, ReadOnly] binding compiles down to that run’s writable roots and is enforced at the syscall jail.

Reported: forks 5× faster than Docker and over 95% prompt-cache reuse on replay. Once a run is forkable, a meta-agent can sit on top of it. Live supervision raised CooperBench pair-coding pass rates from 28.8% to 54.7%. Branching exploration beat baselines across four benchmarks by up to 11 points while cutting wall-clock time by up to 58%. Forking rollouts for Tree-RL moved TerminalBench-2 from 34.2% to 39.4%. MIT-licensed, pip install shepherd-ai, Python 3.11+, early alpha.

Read on Marktechpost → · GitHub · PyPI · Project page

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 Mend.io →