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

397 stories · all sources

MarkTechPost

Google Research releases RRSI for self-improving AI agent harnesses

Google Cloud AI Research, together with UNC-Chapel Hill, Stanford and Washington University in St. Louis, has released RRSI (Regularized Recursive Self-Improvement), an Apache 2.0 research framework for adapting an LLM agent’s harness. The system can modify prompts, tools, memory, control flow and sub-agents while keeping the model weights frozen. It requires Python 3.10+ and accepts LiteLLM model strings; its defaults assume Claude Opus 4.8 on Vertex AI.

MarkTechPost

H Company releases Holo4 models for computer-use agents

H Company has introduced Holo4, a family of vision-language models designed for agents that operate software through screens, write code and invoke MCP or API tools. The release includes Holo4 27B, a dense model, and Holo4 35B-A3B, a mixture-of-experts model with 3B active parameters. Both support a 256K context window through the H Models API and work with the open hai-agents harness.

MarkTechPost

Alibaba releases Qwen-Audio-3.1 voice model for realtime interaction

Alibaba’s Qwen team has released Qwen-Audio-3.1, an audio stack covering automatic speech recognition, text-to-speech and realtime interaction. Its main component, Qwen-Audio-3.1-Realtime, is a full-duplex model designed for voice agents that can reason, use tools and manage when to speak. It is available as a managed QwenCloud API through WebSocket under the name qwen-audio-3.1-realtime-plus; no open weights were announced.

MarkTechPost

Anthropic launches Claude Sonnet 5.5 with faster, cheaper task execution

Anthropic has released Claude Sonnet 5.5, the second model in its Claude 5.5 line after Claude Opus 5.5. The company describes it as a faster, less expensive option for defined everyday work, debugging, and producing documents, presentations, and spreadsheets. The model is available through the Claude Platform under claude-sonnet-5-5, as well as AWS, Google Cloud, and Microsoft Azure. Its weights are closed, so it cannot be self-hosted.

MarkTechPost

NVIDIA introduces an open platform for securing AI agents

NVIDIA has introduced the Open Agent Safety Platform, an open software platform and reference design intended to secure AI agents through controls outside the agents themselves. It combines OpenShell, an Apache 2.0 runtime, with Sentry, an out-of-band watchdog running on BlueField-4 DPUs.

MarkTechPost

Fireworks AI releases Ember-1 for shorter Kimi K3 reasoning

Fireworks AI has introduced Ember-1, a specialized model created by post-training Moonshot AI’s open-weight Kimi K3. The model is designed to preserve useful reflection and feedback-driven reasoning while removing repetitive traces and unproductive loops. Fireworks says this approach differs from simply reducing Kimi K3’s reasoning-effort setting, which lowered quality too substantially in its tests.

MarkTechPost

Jev turns agent decisions into typed probabilities

TypeSafe AI has released Jev, its first System One model, for making structured decisions inside AI-agent workflows. Unlike a conversational model, Jev does not generate text, write code or summarize documents. It receives unstructured text or JSON together with typed questions, then returns decisions, probabilities and confidence values.

MarkTechPost

Google Research presents four frameworks for coherent long-form video

Google Research has introduced an AI video co-director designed to turn short generated clips into coherent, multi-shot stories lasting several minutes. The system addresses identity drift, such as changing clothing or scenery, and cascading failures in which an early faulty asset affects later scenes. It operates as a model-agnostic orchestration layer above Gemini and Veo, whose outputs retain SynthID watermarking.

MarkTechPost

AI coding agents differ sharply on indemnity, data residency and enterprise cost

A contract review of GitHub Copilot, AWS Kiro, Cursor, Devin and Windsurf found major differences in intellectual-property protection, data handling and administrative controls. Cognition acquired Windsurf and renamed the editor Devin Desktop on June 2, 2026, so the two products now share pricing and terms. The review examined vendor documents available on September 26, 2026.

MarkTechPost

Tutorial demonstrates custom encoders on Google Research’s MSEB

A coding tutorial examines Google Research’s Massive Sound Embedding Benchmark (MSEB) through the evaluator interface that turns embeddings into leaderboard results. It explains the package structure and the roles of its main layers: data types for sounds and embeddings, the encoder interface, and task-specific evaluators.

MarkTechPost

Sarvam AI launches Saaras V4 for 22 Indian languages and English

Sarvam AI has released Saaras V4, a speech-to-text model covering all 22 scheduled Indian languages and English, including global English accents. The company says it achieved state-of-the-art accuracy across the Indian-language set. The model is available through Sarvam’s API with model="saaras:v4"; its weights have not been released, and current SageMaker self-hosting documentation applies only to Saaras v3.

MarkTechPost

Supersonic Labs releases CPU-based Julia 1 decision model

Supersonic Labs has released Julia 1, a 144.3M-parameter open decision model designed to select among supplied alternatives rather than generate text. Users provide context, a question and between 2 and 20 options; the model returns one selection along with probabilities for every option. It supports choice classification, ordered scoring and yes-or-no truth estimates, while preserving caller-supplied IDs.