Sakhanda Wire
NVDA $152.86 -2.10% MSFT $505.56 -2.40% GOOGL $280.34 -1.19% META $618.69 -1.23% AMZN $256.63 -0.43%

MarkTechPost

396 stories · all sources

MarkTechPost

TokenGate Speeds Long-Context LLM Prefill by Up to 11.3x

Researchers from Stanford and UC Berkeley introduced TokenGate, a drop-in module that adds learned token pruning to pretrained large language models. The system lets an LLM process all input tokens through lower layers, then uses a small MLP to score them. Only the highest-scoring document tokens continue through upper layers; the question portion is retained. The researchers combined the gate with LoRA (r=16) on Qwen3-1.7B and Qwen3.5-2B.

MarkTechPost

Anthropic launches free AI scanner for open-source vulnerabilities

Anthropic has introduced OSS Scanner, a free, opt-in service that periodically examines selected critical open-source repositories with its strongest models, including Claude Mythos. Eleven days after the October 8 launch, the company reported 77 critical and 803 high-severity findings rated by its models. By October 19, 114 projects had joined from about 650 applications.

MarkTechPost

Microsoft Research open-sources decentralized harness for 1,024 coding agents

Microsoft Research has released Agensh, an open-source harness designed to coordinate large groups of coding agents without a central orchestrator. Each worker selects a sub-task, communicates with peers and contributes changes to a shared Git repository. Gitea, Mattermost and a shared context board provide the main collaboration infrastructure.

MarkTechPost

Cognition releases open-source agent skills for three critical vulnerabilities

Cognition has released Devin Vulnerability Fixes, a free, MIT-licensed plugin containing three agent skills: fix-php-vuln, fix-node-vuln and fix-gitlab-vuln. They address PHP2Shell (CVE-2026-109618 and CVE-2026-109619), vulnerabilities covered by Node.js’s October 21, 2026 security release, and GitLab CVE-2026-110264. The GitLab issue affects self-managed GitLab CE/EE from version 7.0 and has a CVSS score of 10.0. GitLab.com and GitLab Dedicated are already patched.

MarkTechPost

Guide compares 2-bit Saluki 27B with ISTA quantization

A coding tutorial examines Underdog Saluki 27B, a 7.9 GB version of Qwen3.8-27B designed to preserve tool-calling capabilities at roughly two bits per weight. The model’s makers say it slightly surpasses the full 54 GB model on their own benchmark of 120 tasks. Those results are presented as the makers’ claim rather than as an independent conclusion.

MarkTechPost

Anthropic open-sources OS-level sandbox for AI agents

Anthropic has released sandbox-runtime, or srt, an open-source tool that runs processes under operating-system-level file and network restrictions without using containers. The package is available as @anthropic-ai/sandbox-runtime v0.0.81 under the Apache-2.0 license and requires Node.js 22.12+. It can be used with Claude Code, other agents, local MCP servers and ordinary command-line programs.

MarkTechPost

Letta tutorial demonstrates stateful engineering copilots

A tutorial presents a stateful engineering copilot built with Letta, showing how agent memory can extend beyond the contents of individual prompts. The example uses inspectable core-memory blocks that the agent can update autonomously. These blocks can also be read and synchronized by application code, allowing persistent context to be managed outside a single conversation.

MarkTechPost

NVIDIA releases open 4GB Nemotron model for local agents

NVIDIA has released Nemotron 3.5 Bolt E4B 4GB, an open small language model designed for locally running voice assistants, game characters and robot interfaces. It is post-trained from Google’s Gemma 4 E4B and has 8B total parameters, with 4B effective parameters and a 64K context window. NVIDIA targets GeForce RTX, Jetson, RTX Spark and DGX Spark systems, with the Q4_K_M version using 3.1–4.3GB of VRAM in its RTX tests.

MarkTechPost

Kyutai releases CADRE video autoencoders for low-latency world models

Kyutai has released CADRE, a family of seven causal diffusion video autoencoders designed for interactive world models. The approach targets action delay caused by conventional video autoencoders, which often compress several frames into one latent representation. Kyutai reports that CADRE-based world models generate 20 frames per second on one H100 GPU and respond in less than 230 milliseconds at 288×512 resolution.

MarkTechPost

Moonshot AI releases Kimi K3.1 Pro with a one-million-token context

Moonshot AI has released Kimi K3.1 Pro, the first model in its K3.1 series. It is a 3T-class mixture-of-experts model with 2.8T total parameters, 104B active parameters and a one-million-token context window. The model is available through Kimi.ai, Kimi Work, Kimi Code and the Kimi API, while open weights are expected in November 2026.

MarkTechPost

TinyFish outlines a production architecture for agentic search

TinyFish has published two technical guides on building production systems for agentic search and deep research. Its proposed architecture places a base model inside a seven-phase control loop: define an evidence contract, plan retrieval, select tools, discover and retrieve sources, evaluate evidence, recover from failures, and validate the final status. The design separates model reasoning from deterministic execution and uses limits for tokens, time, cost, and tool calls.

MarkTechPost

Google Cloud introduces isolated microVMs for AI agents

Google Cloud has announced Agent Compute Engine, a service designed to run AI agents in dedicated, hardware-isolated microVMs. Each session receives its own guest kernel, root access and Hyperdisk, while the sandbox contains no credentials. The service is already used by Gemini for work tasks, as well as by Antigravity and internal evaluation and reinforcement-learning pipelines. Customer preview is planned for the coming months.