Sakhanda Wire
NVDA $153.55 -1.66% MSFT $504.27 -2.65% GOOGL $281.47 -0.79% META $620.64 -0.92% AMZN $256.56 -0.46%

MarkTechPost

396 stories · all sources

MarkTechPost

Sarvam opens India-hosted inference platform to all developers

Sarvam has made Sarvam Token Factory available to everyone. The managed inference platform serves Sarvam and open-weight models from NVIDIA B200 GPU clusters in India, using one OpenAI-compatible API. The eight models listed at launch include Sarvam 105B, GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash 1017, DeepSeek V4 Pro, Kimi K3, Kimi K3.1 Pro and Gemma 4 31B. Enterprises can also request on-premise deployment.

MarkTechPost

Liquid AI launches d1.2 decision models with text, image and audio support

Liquid AI has released d1.2 and d1.2-ultra, hosted decision models that process text, images and audio. Instead of generating prose, they return probabilities for predefined answers and typed results, producing zero output tokens. The models support up to 65,536 input tokens per question and 128 questions in one call. Their parameters are undisclosed, and both are currently available through the hosted Liquid API only.

MarkTechPost

Meta Deploys JiTTests to Catch Risky Code Changes Before Merge

Meta has moved its Catching Just-in-Time Tests (JiTTests) system from a research prototype into production infrastructure. The LLM-based tool generates executable tests intended to fail when a code diff introduces an unintended behavior change. It runs overnight on spare internal capacity and is designed to target risky changes before they are merged.

MarkTechPost

Alibaba releases Qwen4-Max with 2.4T parameters and 1M-token context

Alibaba’s Qwen team has released Qwen4-Max, the first model in the Qwen4 generation. It is a proprietary sparse mixture-of-experts model with a 2.4T-parameter backbone and 71B parameters activated per token. The multimodal system accepts text, images and video, supports a 1M-token context window, and can produce up to 393K output tokens. It is available only through hosted services, including QwenCloud API, Token Plan, Qwen Studio, QwenWork and Qoder; Alibaba has not released weights for self-hosting.

MarkTechPost

Hark launches Handoff² computer-use agent with faster web task execution

Hark has released Handoff², the second-generation computer-use agent included in Hark Pro. The company says it completes web tasks about twice as fast as the first version on average, with reported gains of up to 3.8 times for flight bookings and more than five times for long research tasks. Handoff² is available through Hark’s cloud on the web, iOS and Android, with each request running on a dedicated virtual computer.

MarkTechPost

Bland Releases Tofu-1 Full-Duplex Speech Model for Phone Calls

Bland has released Tofu-1, a full-duplex speech-to-speech model now offered in an enterprise research preview. The San Francisco voice-AI company says it is designed for difficult phone environments, including moving cars, hospital rooms and weak-signal connections. Tofu-1 processes incoming and outgoing audio directly rather than passing a transcript between speech recognition, a language model and speech synthesis systems.

MarkTechPost

Prototype brings weight-space programming to DSPy

Cmpnd has released dspy-wsp, a prototype optimizer that adds weight-space programming to DSPy. Its WeightSpace component lets a frontier language model inspect and directly modify the weights of a smaller model, then evaluate each change. The process uses no gradients, training loop or fine-tuning dataset. The package is MIT-licensed, can be installed with pip, and requires DSPy 3.5.

MarkTechPost

Stanford Researchers Release Memorilla Memory System for Frozen LLMs

Stanford researchers have released Memorilla, an open-source memory module designed to help frozen large language models use large document collections. The system compresses a collection into 16 learned memory tokens, which are then added to a question for Qwen3-8B. The decoder and Qwen3-Embedding-4B encoder remain frozen, while the 194,123,776-parameter memory module is trained.

MarkTechPost

Vercel Labs launches beta UI SDK for AI coding agents

Vercel Labs has released Vercel UI SDK, an open-source beta library containing more than 50 interface behaviors. It is headless: instead of finished components, it provides primitives for keyboard navigation, focus, selection, popups, dismissal, drag and drop, and related ARIA requirements. Coding agents can combine them into accessible interfaces for React, Vue, Svelte, Solid and Web Components. The library is distributed across six npm packages, with 51 component modules in the repository.

MarkTechPost

Wan 3.0 leads 2026 AI video model rankings

A ranking dated November 13, 2026 places Alibaba’s Wan 3.0 first on Artificial Analysis’s text-to-video board with audio, scoring 1175 Elo. The comparison uses blind human pairwise preferences and sets Kling 3.0 1080p (Pro) at 1000 Elo as the anchor. Wan 3.0 costs $0.200 per second at 1080p, while Utopai X ranks second at 1170 Elo but has no public API.

MarkTechPost

Perplexity Releases Decider v1.3 With Span-Extraction Feature

Perplexity has released pplx-decider-v1.3 and pplx-decider-flash-v1.3, open-weight decision models that return probabilities for labeled options rather than generated text. They support yes/no, multiple-choice and scoring tasks involving text or images. The weights are available under Apache 2.0 on Hugging Face, and both models can be accessed through the Perplexity Decisions API and OpenRouter.