Sakhanda Wire
NVDA $153.73 -1.54% MSFT $506.30 -2.26% GOOGL $280.09 -1.28% META $619.28 -1.14% AMZN $256.62 -0.44%

MarkTechPost

396 stories · all sources

MarkTechPost

Three optimizations raise Ling-3.0-flash throughput on DGX Spark

NVIDIA tested three combined techniques for mixture-of-experts inference on a single DGX Spark: Humming MXFP4 MoE kernels, an FP8 version of the language-model head, and built-in NEXTN speculative decoding. The test used Ant Group inclusionAI’s Ling-3.0-flash-fp4, a 124B-parameter model with 5.1B active parameters per token, running at batch size one on a system with 128 GB of unified memory.

MarkTechPost

Soniox STT v6 adds agent context and voice echo filtering

Soniox has released Soniox STT v6, a hosted speech-to-text API for voice agents. The service is available through the cloud as stt-rt-v6 for streaming and stt-async-v6 for files, supporting more than 60 languages. Soniox says the new model is significantly more accurate than v5, although it has not published independent v6 measurements or a parameter count.

MarkTechPost

OpenAI unveils Pando Git backend for high-volume monorepo traffic

OpenAI has introduced Pando, a Git-compatible backend serving development in its monorepo while preserving standard Git clients and protocols. Engineers and Codex agents continue to use commands such as git push and git fetch. Written in Rust, Pando reached production in 10 weeks and has handled production traffic since April 2026.

MarkTechPost

Zyphra presents tree-based sparse attention for long-context models

Zyphra has introduced Zyphra Hierarchical Attention (ZHA), a sparse attention method that uses a tree-structured index to select relevant entries from a model’s key-value cache. Unlike approaches whose indexer scores every previous token, ZHA makes the search hierarchical. It groups keys into chunks, creates summary keys level by level and uses top-p routing to expand only branches carrying the estimated attention mass. The complete KV cache remains stored, while the tree determines which entries are read.

MarkTechPost

The Letter Company Releases Letterpress-1 Growth Engineering Agent

The Letter Company has released the beta of Letterpress-1, a coding agent, harness and IDE designed for growth engineering. It analyzes a git-based website alongside market signals, then proposes and builds changes in a sandbox before opening a pull request with a live preview. The hosted platform works with a team’s git repository and does not require a local GPU; Letter says setup takes less than five minutes.

MarkTechPost

Deep Principle releases 753B-parameter Hydrogen-1.0 science model

Deep Principle has released Hydrogen-1.0, an open-weight model aimed at chemistry, materials science and life sciences. The 753B-parameter mixture-of-experts system is a post-trained version of Z.ai’s GLM-5.3, with about 40B active parameters, a 1M-token context window and up to 128K output tokens. Its weights are available under the MIT License.

MarkTechPost

NVIDIA Releases Open Auto Ontology for Grounded Agent SQL

NVIDIA has released Auto Ontology, an Apache 2.0 library that generates a semantic layer from enterprise metadata. The layer describes where data is stored, what fields and metrics mean, how they should be used, and their lineage. It is intended to help agents produce SQL based on certified business definitions rather than raw schema information.

MarkTechPost

qwenGPU embeds a four-lane CUDA-capable GPU in Qwen3.5-2B

qwenGPU is an experimental version of Qwen3.5-2B that places 45 compiled layers ahead of the model’s original 24, creating 69 layers in total. A compiler generates the added weights rather than training them. The embedded machine is a four-lane SIMT GPU with registers, program ROM, memory and support for integer operations, predicates, branches, loops and selected CUDA primitives.

MarkTechPost

Nace AI releases open 27B Drex Max decision model

Nace AI has released Drex Max v1 with open weights, making its 27B decision model available for local deployment. The package includes bf16 weights, a Q8_0 GGUF file, an affine Q4 checkpoint, and modified versions of llama.cpp and Ollama. The model accepts text, JSON or images together with named questions and returns probabilities for possible answers rather than generating a response.

MarkTechPost

DeepSeek releases dsh-superagent for parallel coding tasks

DeepSeek has released dsh-superagent, an MIT-licensed research prototype designed to handle requests for a single GitHub repository. The TypeScript bundle, version 0.1.0, listens to GitHub and optionally Feishu, then converts authorized requests into Tickets. Each Ticket runs in a separate agent session and directory, with results returned to the original conversation, often through a pull request.

MarkTechPost

Factory releases model-independent Droid 1.0 coding agent harness

Factory has released Droid 1.0, an update to its model-independent coding agent harness. Its main change is programmatic tool calling, enabled by default through a JavaScript-based Script tool. Instead of returning every tool result to the model, the system runs scripts in a sandbox and sends back only printed output. The sandbox itself has no file, network or system access, while independent calls can run concurrently.

MarkTechPost

Vercel Labs releases open-source desktop tool for AI agents

Vercel Labs has released agent-desktop, an open-source Rust CLI and C library that enables AI agents to operate native desktop applications on macOS, Windows and Linux. Version 1.0.2 was released on October 30, 2026, under the Apache-2.0 license. The tool has 60 command names, 56 of them operational, and runs on standard CPUs without a GPU.