Sakhanda Wire
NVDA $155.46 -0.44% MSFT $506.24 -2.27% GOOGL $280.20 -1.24% META $618.87 -1.20% AMZN $257.89 +0.05%

MarkTechPost

396 stories · all sources

MarkTechPost

Guide Demonstrates Safe Selection for Self-Improving AI Agents

A coding tutorial demonstrates RRSI, or Regularized Recursive Self-Improvement, a method for changing an AI agent’s harness while leaving its underlying language model frozen. The harness can include prompts, tools, memory, control flow and sub-agents. The approach is designed to limit overfitting to the tasks used during evolution.

MarkTechPost

Mellum2.1 launches as an open 12B model for coding agents

JetBrains has released Mellum2.1, an Apache 2.0-licensed mixture-of-experts reasoning model designed for coding agents, fast sub-agents and self-hosted use. It contains 12B total parameters but activates 2.5B per token through eight of 64 experts. The model supports a 131,072-token context and can run on local GPUs with vLLM or SGLang. GGUF versions range from 7.0 GB to 24.3 GB, depending on format.

MarkTechPost

Architect launches Liquid Inference auction router for LLM requests

Architect Financial Technologies has introduced Liquid Inference, an exchange-style router that lets inference providers compete for individual LLM requests. Providers submit offers for specific models, and the lowest-priced offer that satisfies the buyer’s conditions is selected. The maximum price is fixed before generation begins, while billing covers measured usage.

MarkTechPost

NVIDIA Researchers Introduce PivotOPD for Recovering AI Agents

Researchers from NVIDIA, Princeton University and the University of Maryland introduced PivotOPD, an on-policy distillation method for multi-turn language-model agents. It is a training technique rather than a new model and adds no inference cost. Tests used Qwen3-1.7B and Qwen3-8B students, with additional evaluation of a Nemotron-3.5-SFT student on SWE-Bench Verified.

MarkTechPost

Perplexity releases two multimodal embedding models for local deployment

Perplexity has released pplx-embed-v2-late, a pair of ColBERT-style multimodal embedding models available on Hugging Face under the MIT license. The 0.6B version is intended for low-cost, fast query encoding on laptops, edge devices or small GPUs, while the 9B model is designed to build higher-quality document indexes. Both handle text, images and rendered PDF pages in a shared embedding space. A hosted Perplexity API endpoint is planned but is not yet available.

MarkTechPost

Unsloth Studio adds layered checks before running model code

Unsloth has launched the beta desktop app Unsloth Studio after more than 500 million downloads of its projects and years of requests from the open-source community. The app brings model installation, fine-tuning and local execution into a single dashboard. Its security design was influenced by a compromised LiteLLM release on PyPI and an infostealer hidden in a Hugging Face repository that imitated an OpenAI Privacy Filter model.

MarkTechPost

Anthropic launches Claude Haiku 5.5 with one-million-token context

Anthropic has released Claude Haiku 5.5, a small model designed for high-volume tasks such as summarization, classification, document question answering and subagent work. It accepts text and images, produces text, supports a context window of 1 million tokens and can generate up to 128K tokens. The model is generally available through the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS.

MarkTechPost

Liquid AI releases open-weight multimodal decision models

Liquid AI has released Open d1, a pair of open-weight models designed to make fixed-option decisions rather than generate text. The d1-3B model processes text and images, while d1-omni-600M accepts text with an image or with audio. Each model returns typed, calibrated probabilities in a single forward pass and reports zero output tokens.

MarkTechPost

Meta open-sources Rebalancer for large-scale assignment problems

Meta has released Rebalancer, an Apache 2.0-licensed C++ library with a Python interface for assigning objects to bins under defined constraints and objectives. The company says the system has been used across its infrastructure for more than 9 years and now solves about 40 million assignment problems per day across more than 30 formulations. The release includes documentation, a PyPI package and Rebalancer Explorer, a Dockerized interface for examining solver behavior.

MarkTechPost

Guide examines Laya for zero-shot decisions and calibration

A developer tutorial examines Laya, an open-source decision engine from Convai Innovations that became one of the most-starred machine-learning repositories in September 2026. Laya is described as a non-autoregressive System 1 model: its 421-million-parameter encoder reads text together with typed questions and returns probabilities for the available answers in one forward pass, without generating output tokens.

MarkTechPost

Google DeepMind releases open multimodal EmbeddingGemma 2

Google DeepMind has released EmbeddingGemma 2, an open embedding model that represents text, code, images, video and audio in a shared 768-dimensional vector space. Built on the Gemma 4 architecture, it has 740M parameters, an 8,192-token context window and an Apache 2.0 license. The model is intended for applications including on-device search, classification and retrieval-augmented generation where data can remain local.

MarkTechPost

Mistral AI previews 1.05-trillion-parameter Mistral Large 4

Mistral AI has released Mistral Large 4, also known internally as “Le Chonk,” as a public preview. The Mixture of Experts model contains 1.05 trillion total parameters, with 49 billion active for each token. It accepts images natively through a 1.6-billion-parameter vision encoder and supports a context window of up to 1 million tokens.