Mistral Large 4 Weights Due Oct 31; Local Use Would Need Roughly 400GB of Memory

·3 views

Mistral has announced Large 4, a 1.05 trillion parameter mixture-of-experts model with 52 billion active parameters, which it describes as open-weight. As of October 22 the weights are not yet downloadable, and the Hugging Face page points to an October 31 release. Mistral's documentation leaves the license unnamed, unlike its other recent models, which are listed under Apache 2.0. Self-reported benchmarks show mixed results, and commenters have questioned performance and pricing relative to rival models. Running the model locally would require memory for all of its parameters, with quantized estimates falling in the range of about 400GB or more.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Where Wealth Is Moving in America

A Hacker News post links to an article titled "Where Wealth Is Moving in America" published on thestatedtruth.com. The submission has drawn 6 points and 2 comments on the Hacker News discussion thread. The provided material contains only the link and engagement metrics, without the article's body text. As a result, the specific findings, data, and conclusions about wealth migration across the United States cannot be confirmed from this source.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

Microsoft Links ConnectWise Cloud Backup Supply-Chain Attack to Storm-0940

Microsoft has attributed a supply-chain attack on ConnectWise customers' Microsoft 365 environments to a threat group it tracks as Storm-0940. The intrusion path ran through ConnectWise's hosted Cloud Backup service. ConnectWise temporarily shut down services and began restoring them in phases, advising customers to keep connections disabled during the process, which it said would take several days. Microsoft said successful authentication does not necessarily mean the group accessed or stole data. Reporting does not state how many tenants were compromised.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Defense Contractors Urged to Shift From Snapshot to Continuous NIST 800-171 Evidence

Ahead of a Cyber AB Town Hall on October 27, the CMMC Reform Task Force's findings remain unpublished. Presenters at September's Town Hall speculated that the Department of Defense may move away from point-in-time assessments that quickly become outdated. Cybersecurity advisors recommend that contractors generate evidence on a regular schedule, such as monthly MFA and access reports, dated vulnerability scans, and centralized audit logs. They also suggest mapping each of NIST SP 800-171's 110 requirements to its proof, refresh cadence, and owner within the System Security Plan. Regardless of reform outcomes, contractors handling CUI under DFARS 7012 must still meet NIST SP 800-171 and maintain an accurate SPRS score.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

ECUPSE Runs a Company Where Every Operational Role Is an Autonomous AI Agent

ECUPSE is a company in which every operational function, including content creation, market analysis, creative writing, and publishing, is handled by autonomous AI agents producing public output on a weekly schedule. The article explains the methodology, starting with breaking each business process into discrete, automatable stages with defined inputs, outputs, and success criteria. Each stage runs as a Lambda function orchestrated by AWS Step Functions, which halts and alerts on failure rather than failing silently. Each agent is shaped by a persona configuration covering identity, constraints, guardrails, and visual theme, which lets the same underlying model produce distinctly different output. The author argues that architecture and process engineering matter more than the model itself, with Bedrock guardrails acting as hard post-generation filters.

0 comments Read more at DEV Community