Sakhanda Wire
NVDA $153.62 -1.62% MSFT $506.90 -2.14% GOOGL $279.81 -1.38% META $619.78 -1.06% AMZN $256.58 -0.45%

MarkTechPost

396 stories · all sources

MarkTechPost

Microsoft and Berkeley Researchers Release TokenEconomizer for Coding Agents

Researchers from Microsoft Research and UC Berkeley have released TokenEconomizer, an open-source plugin designed to reduce coding-agent costs without changing the agent or its model weights. It works with Claude Code, Codex, GitHub Copilot CLI and Pi on machines running Python 3.10 or later, and does not require a GPU or its own API key.

MarkTechPost

Four AI security agents converge on a verify-before-patching design

Anthropic, OpenAI, Cloudflare and Google have released AI agents that search source code for security vulnerabilities. Their systems use a common workflow: one group of agents investigates, an independent agent challenges each candidate, and only findings that survive validation proceed to reporting or remediation. Google’s security team has warned that unvalidated AI scanning can produce true-positive rates below 7%, making false alarms costly for reviewers.

MarkTechPost

Samaya AI releases Sterling for lower-cost financial research

Samaya AI has released Sterling, an agent model designed for financial research. It is a LoRA post-trained version of NVIDIA’s Nemotron 3.5 Lightning, a mixture-of-experts model with 30B total parameters and 3B active parameters. Sterling plans research, uses finance tools, collects evidence and passes the results to a separate report writer.

MarkTechPost

Edge0 Releases 0.6B Model for Multilingual Voice Cloning

Edge0 has released Audio8-TTS-Realistic, an open-weight text-to-speech model designed for zero-shot voice cloning. It can reproduce a voice from one reference recording and its transcript, while also supporting speech generation without a reference. The 0.6B model supports 23 languages, including 12 added in this release, and its weights and code are available under Apache 2.0. The package includes a tokenizer, processor, custom code and a 44.1 kHz neural audio codec.

MarkTechPost

Joey-1 launches as a 2B decision model for low-cost classification

Baseten has released joey-1, a 2B-parameter open-weight model developed by its Base Labs research team. Instead of producing free-form text, it returns labels and calibrated probabilities for questions involving choices, yes/no answers or ordered scores. The model is designed for high-volume routing, classification and guardrail checks, and can process text or images in one forward pass.

MarkTechPost

Prime Mega MoE adds trainable backward pass to fused MoE kernel

Prime Intellect has released Prime Mega MoE, a training extension for DeepSeek’s fused Mega MoE kernel. The original implementation supported only the forward pass; the new overlapping backward pass makes the kernel usable during model training. It is available as the mega_moe dispatch backend in prime-rl.

MarkTechPost

Infini-AI-Lab Releases Open-Source Harness Adapter Framework Hermit

Infini-AI-Lab has released Hermit, an Apache-2.0 Python framework for adapting AI agents without changing their underlying model or harness. Version 0.1.0 was released on October 25, 2026, and CMU postdoc Haizhong Zheng announced it on October 27. The framework trains an external adapter that can contain instructions, skills, memory, tools, executable code or coordination logic.

MarkTechPost

Block Unveils Dessa, a Speech-Only Full-Duplex Model

Block’s AI research lab has presented Dessa as a research preview for a full-duplex speech-to-speech system. The model takes in audio and produces audio while listening and speaking simultaneously. Block says it was trained solely on human conversation recordings, without text data or pretrained models.

MarkTechPost

AssemblyAI launches TTS-1 for live voice-agent calls

AssemblyAI has made AssemblyAI-TTS-1 generally available, targeting voice agents that operate during live phone calls. The hosted streaming model converts agent responses into speech and costs $15 per 1 million characters. The price includes unlimited concurrency and does not require an annual contract. It runs in AssemblyAI’s cloud in the US region us-east-1 or the EU region eu-north-1, with no announced on-premises option.

MarkTechPost

NVIDIA releases open 40-language Nemotron 3.5 Transcribe model

NVIDIA has released Nemotron 3.5 Transcribe, an open-weights speech language model designed for transcription. It supports 40 languages and audio recordings lasting up to one hour. Text instructions can select verbatim, edited, speaker-attributed, contextual or translated output, combining transcription, diarization and word alignment in one model.

MarkTechPost

Google DeepMind unveils Gemini 4 Carbon for coding and cyber defense

Google DeepMind has introduced Gemini 4 Carbon, a frontier model designed for long-running agentic coding and cybersecurity work. The company says it ranks first on the v5.0 Artificial Analysis Intelligence Index with a score of 59 and leads 15 of 17 published comparisons. Its parameter count and context window have not been disclosed, and the model is available only through Google-hosted systems.

MarkTechPost

MiniMax open-sources MCode Stack for coding agents

MiniMax has released MCode Stack, an MIT-licensed repository containing the workflows and agent skills used by its MiniMax Code team. Published on October 28, 2026, the repository includes 17 SKILL.md skills—15 from Ronny and two from Jiachen—along with five checklists and five companion notes. It contains no MiniMax Code source code and requires no GPU. The skills run on machines with MiniMax Code and Node.js 22.19+ or with Claude Code.