Sakhanda Wire
NVDA $154.51 -1.04% MSFT $506.49 -2.22% GOOGL $281.27 -0.86% META $621.51 -0.78% AMZN $256.90 -0.33%

MarkTechPost

396 stories · all sources

MarkTechPost

Zyphra launches open-source ZONOS3-Mini speech model

Zyphra has released ZONOS3-Mini, an open-source text-to-speech model designed for real-time operation on local devices. Its sparse Mixture of Experts architecture contains 800M parameters, with about 100M active for each token through 64 experts and top-2 routing. The weights are available under the Apache 2.0 license, and the model supports zero-shot voice cloning and word-level control of delivery.

MarkTechPost

Anthropic releases Claude Fable 5.2 with a lower-safeguard Mythos variant

Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2 as two safeguard configurations of the same underlying model. Fable targets long-horizon coding, research and agentic tasks, while Mythos removes safeguards covering high-risk cybersecurity work and most research biology. Access to Mythos is restricted to Anthropic’s Cyber Verification and Life Sciences Verification programs.

MarkTechPost

Meta launches Muse Ember for high-volume agent tasks

Meta Superintelligence Labs has released Muse Ember, a hosted multimodal reasoning model designed for repetitive work inside AI agents. Meta targets it at tool calls, coding subtasks, routing, extraction and long-document parsing. It is the smallest model in the Muse line, positioned below Muse Wonder and Muse Spark.

MarkTechPost

Google adds dual-chip TPU 8i Max and confidential TPU 8tx

Google has expanded its eighth-generation TPU lineup with TPU 8i Max and TPU 8tx. The company describes them as new system configurations rather than new chip designs. Both are intended for Google Cloud TPU systems, alongside the previously announced TPU 8t and TPU 8i.

MarkTechPost

Baidu’s ERNIE Evolve reports gains in AI code optimization

Baidu has released ERNIE Evolve, an automated research system designed to improve AI training and inference code. Given a repository, an objective and evaluation criteria, it selects research directions, modifies code and runs experiments. A hidden evaluator and an independent reviewer must approve a change before it is retained. Baidu has not disclosed which language models power the agents, and the system itself has not been released.

MarkTechPost

Reason Machines Open-Sources Aspen Git Server for AI Agents

Reason Machines has released Aspen, an open-source, self-hosted Git server designed for workloads in which hundreds of AI agents push to repositories concurrently. Aspen implements the design OpenAI described for Pando and supports standard Git over HTTPS and SSH. It uses PostgreSQL or SQLite for branch history and S3-compatible object storage for Git packs. The software can run on a laptop with aspen dev, through Docker Compose, or in cloud environments using Helm and Terraform.

MarkTechPost

0sec Labs Releases Open-Source AI Security Harness

Zurich-based 0sec Labs has released 0, an open-source harness for AI-assisted vulnerability investigations, code review, penetration testing, reverse engineering and fix generation. The software is distributed under the MIT OR Apache-2.0 license as a research preview. It is a control layer rather than an AI model: users connect their own provider, with 14 providers listed, while inference runs externally.

MarkTechPost

Google Open-Sources Elephant-Goldfish Plugin for Spec-First Coding

Google has open-sourced the Elephant-Goldfish Model (EGM), an Apache 2.0 plugin that adds a design and review layer to AI coding agents. EGM is not a language model or an officially supported Google product. Instead, an “Elephant” agent retains project context, drafts a specification and does not write code, while fresh “Goldfish” subagents review only the draft and referenced files.

MarkTechPost

Comparison maps routing choices across 10 open MoE models

A configuration-verified comparison examines how 10 open mixture-of-experts (MoE) models assign tokens to experts. The models range from Qwen3.6-35B-A3B, with 35B total and 3B active parameters, to DeepSeek-V4-Pro, with 1.6T total and 49B active. Their routed expert counts span 8 to 384, while top-k selection ranges from one expert in Llama 4 Maverick to eight in several models.

MarkTechPost

MoE routing strategies shape model quality, balance and serving costs

Mixture-of-Experts models use a router for each token, selecting a small number of expert networks instead of running all of them. The choice affects model quality, workload balance and the communication costs of multi-GPU serving. Mixtral 8x7B, for example, exposes 47B parameters but activates 13B for each token.

MarkTechPost

Arete-27B cuts propaganda responses in Qwen3.8-27B fine-tune

Nous Research has released Arete-27B, a 27B dense open-weight model created by supervised fine-tuning Alibaba’s Qwen3.8-27B. The model keeps the base architecture and Apache 2.0 license. The training used no reinforcement learning or preference optimization, and the vision encoder remained frozen because the dataset was text-only.

MarkTechPost

Guide details how to run NVIDIA Dory 1.0 4B with llama.cpp

A coding tutorial examines Dory 1.0 4B, a compact reasoning model developed by NVIDIA’s Nemotron Labs. Its middle layers form a shared block that can be executed one, two, three or four times for each token. Increasing the loop count raises computation per token without increasing the parameter count, allowing one checkpoint to operate at four compute levels without retraining.