Sakhanda Wire
NVDA $153.38 -1.77% MSFT $505.32 -2.45% GOOGL $279.76 -1.40% META $619.10 -1.17% AMZN $257.21 -0.21%

MarkTechPost

396 stories · all sources

MarkTechPost

Jev 1.14 Adds Experimental Images, Bounding Boxes and Longer Context

TypeSafe AI has released Jev 1.14, a new version of its System One typed decision model, under the jev-preview alias. The model returns typed answers such as Choice, Score or Noul, along with probabilities and confidence values, rather than generated text. The update is available only through TypeSafe’s hosted API, and its parameter count has not been disclosed.

MarkTechPost

SpaceXAI releases Grok Voice 3.0 for realtime voice agents

SpaceXAI has released Grok Voice 3.0, a speech-to-speech model designed to listen, reason, use tools and respond with audio. It is available only through the company’s hosted Voice Agent API, under the grok-voice-3.0 name on the /v1/realtime WebSocket endpoint. The grok-voice-latest alias already points to the new version, while Grok Voice 2.0 is deprecated. SpaceXAI has not disclosed parameter counts or architectural details.

MarkTechPost

Datalab Releases Open OCR Benchmark With 16,288 Pass/Fail Tests

Datalab has released OmniParseBench, an open benchmark for document parsers. It evaluates outputs through 16,288 yes-or-no tests covering 2,937 pages from 2,343 documents in 92 languages. The dataset and code are public, and the project is designed to show how a parser performs on specific document characteristics rather than only provide a single leaderboard position.

MarkTechPost

Fish Audio launches Drama 3 open-weight text-to-speech model

Fish Audio has released Drama 3, a 35B mixture-of-experts text-to-speech model with about 3B active parameters per token. It converts scripts and plain-language acting instructions into performed speech, allowing cues such as whispering or emotional direction to be applied to selected text. The model supports 83 languages and up to 16 speakers in one request, with separate voice IDs for each character.

MarkTechPost

Bifrost Leads 2026 Ranking of AI Gateways

A review of nine AI gateways for 2026 places Bifrost first among self-hosted options, followed by Portkey, LiteLLM, Kong AI Gateway, TrueFoundry, Cloudflare AI Gateway, Vercel AI Gateway, OpenRouter and Orq.ai Router. The comparison considers performance, provider coverage, governance, deployment choices, pricing and vendor or supply-chain risk.

MarkTechPost

Liquid AI releases 24B MoE model for local AI agents

Liquid AI has released LFM2.5-24B-A2B, an open-weight, text-only reasoning model designed for agents running on laptops, PCs and mini PCs. The Mixture-of-Experts model contains 24B parameters but activates about 2B per token, or 2.45B according to its model card. It has a 131,072-token context window and is the largest and final model in the LFM2.5 family.

MarkTechPost

Google Makes Gemini 4 Argon Generally Available

Google DeepMind has made Gemini 4 Argon generally available to developers, enterprises and paid Gemini subscribers through the Gemini API, AI Studio and the Gemini platform. The sparse mixture-of-experts model is designed for extended, multi-step tasks in coding, finance, legal work, science and cybersecurity. Google first introduced it on September 30 to a limited group of cyber defenders.

MarkTechPost

Meta releases Muse Wonder models for long-running AI agents

Meta has released Muse Wonder and Muse Wonder Full, two models designed for extended agent tasks such as software development, legal and financial work, and computer use. Both support a 1M-token context window and accept text, images, video and audio as input, while producing text. They run on Meta’s cloud through the Meta Model API or the Muse Code command-line tool.

MarkTechPost

Tutorial examines Meta’s Rebalancer solvers and assignment checks

A coding tutorial examines Meta’s Rebalancer, an assignment solver designed to place tasks across datacenter resources. The implementation combines a C++ core with Python bindings and introduces the main concepts used to describe models, including objects, containers, dimensions, scopes and partitions. It also shows how to define constraints and prioritize goals.

MarkTechPost

Z.ai introduces GLM-Legal for agentic legal work

Z.ai has released GLM-Legal, a legal-specialized model built on GLM-5.3-Flash for contract review, evidence analysis and compliance research. The underlying model has 320B total parameters, 18B active parameters, a 1M-token context window and a maximum output of 128K tokens. Z.ai has not disclosed a separate parameter count for GLM-Legal.

MarkTechPost

Alibaba’s Qwen team presents QwenGyre for long-horizon agent training

Alibaba’s Qwen team has introduced QwenGyre, an online reinforcement learning framework for training coding agents on tasks that can run for hours and approach 1M tokens. It is a training system rather than a model and works with unmodified, black-box harnesses. Tests used Qwen 3.6 122B and Qwen 3.8 2.4T models, with rollouts of about 700K tokens.

MarkTechPost

Twelve AI tools target real-time speech translation in 2026

A review of 12 AI live translation products compared their stated languages, latency, platforms and pricing using official product and documentation pages. These services generally combine speech recognition, machine translation and text-to-speech, producing captions or translated audio while a conversation is in progress.