Sakhanda Wire
NVDA $153.75 -1.53% MSFT $506.06 -2.30% GOOGL $279.99 -1.31% META $621.17 -0.84% AMZN $257.88 +0.05%

MarkTechPost

396 stories · all sources

MarkTechPost

NVIDIA Library Enables Zero-Training Predictions for Tabular Data

NVIDIA’s Structured Data Models library provides a PyTorch interface for several pretrained foundation models that work with tabular and relational data. Instead of fitting a model, users supply labeled context rows together with query rows, and the system produces class probabilities or a predictive distribution in a single forward pass.

MarkTechPost

Voxel51 3.0 combines multimodal data workflows for physical AI

Voxel51 has released Voxel51 3.0, an end-to-end platform that brings search, curation, annotation and evaluation into one system. It is designed for teams working with sensor data from robots, vehicles and factories, and supports MCAP, images, video, 3D point clouds, meshes, DICOM, NIfTI, geospatial data and audio.

MarkTechPost

Kyutai releases 2.4B-parameter REKI robot manipulation model

Kyutai, together with Inria and École des Ponts, has released REKI, a 2.4B-parameter vision-language-action model for robot manipulation. It combines a Qwen3.5-2B vision-language backbone with a 0.2B flow-matching action expert and was pretrained exclusively on public data. The model’s weights are available in BF16 under the CC BY-NC-SA 4.0 license.

MarkTechPost

DeepSeek releases V4.1-Flash-1017 and opens Search and Fetch APIs

DeepSeek has released DeepSeek-V4.1-Flash-1017, an updated checkpoint of its V4.1-Flash model. It is available through the company’s website, app and API under the model name deepseek-flash. DeepSeek says the release improves performance, completes tasks faster and produces shorter outputs than the September version. The model has a 1M-token context window and supports up to 384K output tokens; its hardware requirements and 1017 weights have not been disclosed.

MarkTechPost

Twenty decision models emerge as alternatives to TypeSafe’s Jev

TypeSafe launched Jev on September 15, 2026, as a decision model rather than a chatbot. Developers provide a state, such as a ticket or log line, together with typed questions, and Jev returns labels, scores or yes/no probabilities without generating text. Its documented price is $0.042 per 1M input tokens, with free output and a 64k-token context, but its weights are closed and access is waitlisted.

MarkTechPost

StepFun opens weights for 600B-parameter Step 5 Preview model

StepFun has released Step 5 Preview with open weights in FP8 and BF16 formats through Hugging Face and ModelScope. The sparse Mixture-of-Experts model has 600B total parameters, while 27B are active for each token. Its context window reaches up to 1M tokens, and StepFun presents it for software engineering, document analysis, finance and other long-horizon agentic tasks.

MarkTechPost

Perplexity open-sources SPACE sandbox for AI agents

Perplexity has released SPACE, an Apache 2.0-licensed platform for running AI-agent tasks in isolated Firecracker microVMs. The Rust project is available on GitHub and provides an HTTP API with server-sent events. Each sandbox has its own guest kernel, writable filesystem and network gateway, while the host remains outside the agent’s control.

MarkTechPost

Liquid AI Releases d1-ultra Decision Model With Multimodal Input

Liquid AI has released d1-ultra, a decision model that accepts text, images, or both. Instead of producing free-form text, it answers typed questions with calibrated probabilities in a single forward pass. Responses contain no output tokens, which supports classification, routing, moderation, triage and scoring without generation charges.

MarkTechPost

Meta releases Muse Spark Open with 816B parameters and 1M context

Meta Superintelligence Labs has released Muse Spark Open, an open-weight Mixture-of-Experts model with 816B total parameters and 42B activated for each token. It accepts text, images, video and audio, produces text, and supports a 1M-token context window. The model uses 256 experts, with eight selected per token, and includes a 1.8B-parameter perception encoder for visual input. Meta describes it as the most capable US open-weight model so far.

MarkTechPost

NVIDIA Releases 120B Open Multimodal Agent Model

NVIDIA has released Nemotron 3.5 Super VL, an open vision-language model designed for enterprise agents. The mixture-of-experts system contains 120B parameters and activates 12B for each token. It accepts text, images, documents and video, and can produce text, reason, call tools and write code. Its context window extends to 1M tokens, while MP4 video inputs are limited to clips of up to two minutes.

MarkTechPost

CopilotKit launches CLiF for learning from agent and user trajectories

CopilotKit has released CLiF in early access, a continual-learning platform designed to help teams improve AI agents from activity inside their products. CLiF records the agent’s runs, tool calls and state changes together with users’ clicks, screen context, server requests and conversations as a single full-stack trajectory through the AG-UI protocol. CopilotKit says this combines information that agent tracing and product analytics typically keep separate.

MarkTechPost

openKylin Presents AI-Focused Linux Architecture at Open Source Summit Europe

At Open Source Summit Europe 2026 in Prague, held from October 7 to 9, openKylin presented its plans for adapting Linux to AI-oriented computing. On October 8, Executive Director Min Liu introduced KART, a unified AI runtime, and an environment for executing system-level agents. Technical Committee member Jianfeng Li separately discussed an AI-ready Linux desktop on RISC-V.