Alibaba's Qwen3.8-Max Tops Some Benchmarks But Trails Claude on Key Coding Tests

·1 views

Alibaba has launched Qwen3.8-Max, its largest AI model to date, featuring 2.4 trillion total parameters, 95 billion active per token, and a one-million-token context window. The mixture-of-experts model supports multimodal input and agentic workflows, and Alibaba plans to release its weights publicly within a week of the API going live. The company claims the model autonomously built a software project over roughly 16 days, generating 265 commits and 127 pull requests in a publicly viewable GitHub repository. Qwen3.8-Max outperforms Claude on three benchmarks — Terminal Bench 2.1, PaperBench, and OSWorld-Verified — but trails Claude Fable 5 on harder repository-level engineering tests like SWE-bench Pro and FrontierSWE. Alibaba's benchmark results have not yet been independently replicated, and the model weights remain unavailable for download at the time of reporting.

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)

Developer Tests Four RAG Retrieval Methods on 312-Doc Knowledge Base, Only One Holds Up

A developer built a personal knowledge database called context-forge, containing 312 documents including tweets, research abstracts, blog posts, and YouTube transcripts, managed daily using Claude Code via a CLI tool. The system uses a SQLite index with markdown files and an automated pipeline that fetches, summarizes, scores, and categorizes each new document from a URL. To find the best retrieval strategy, the developer ran four different approaches over two-week trials each against real daily tasks such as writing, research, and decision-making. Vector embedding search degraded around 200 documents because topically dense, short docs produced near-identical similarity scores, while keyword-based BM25 search via SQLite FTS5 later broke down due to ambiguous query terms returning unrelated results. Of the four strategies tested, three failed in distinct and reproducible ways that the developer notes are commonly recommended in mainstream tutorials.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Most Companies Have a Dangerously Flawed Backup Strategy

Most organizations lack a proper data inventory, making it impossible to build an effective backup strategy since they do not know what data they hold, where it lives, or who owns it. Not all data carries equal value — raw inputs may be recreatable from upstream sources, while processed insights and pipelines often require far greater protection. Annual backup costs can run into hundreds of thousands of dollars, yet many companies never formally decide how much data loss or downtime they can actually tolerate. Backup strategies often drift into place based on initial defaults rather than deliberate executive decisions, leaving dangerous gaps between what teams assume is protected and what actually is. Experts recommend cross-cloud replication as the strongest ransomware defense, but also urge businesses to first assess what data is truly irreplaceable before investing in complex infrastructure.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Context window limits, not APIs, are the real bottleneck for MCP servers

A developer building an HTTPS-based MCP server to connect AskElephant's call-recording tool with claude.ai discovered that the core design challenge was not API integration but context window arithmetic. An analysis of 3,706 client call engagements showed that a single transcript could reach nearly 44,000 tokens, and a typical multi-meeting search query could naively generate over 325,000 tokens — far exceeding a 200,000-token context window. To solve this, the team split responsibilities so that a Worker function scans and filters transcripts locally, returning only relevant excerpts with speaker labels and timestamps, while Claude handles reasoning over the condensed output. This approach reduced a real 40-meeting search from a theoretical 325,000 tokens to roughly 19,200 tokens of targeted excerpts. The team also identified a compounding bug where three independently reasonable caps on meetings, hits, and context characters were never validated as a combined budget, allowing worst-case outputs to silently exceed the window limit.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer builds MCP server to query Google Analytics, Ads, and Search Console via AI

A developer has built Metrifyr, a remote Model Context Protocol (MCP) server that connects Google's marketing tools — including Analytics 4, Search Console, AdSense, Tag Manager, and Google Ads — to AI agents like Claude. Users can ask plain-language questions and receive answers drawn from across multiple Google products without manually navigating any dashboards. The tool goes beyond raw data retrieval, offering built-in analysis features such as content decay scanning, keyword cannibalization detection, and a cross-product view that joins Search Console and AdSense data to rank pages by actual ad revenue. To manage context load, the server only surfaces tools relevant to the Google accounts a user has connected, keeping each session lightweight despite the overall catalog exceeding 100 tools. Metrifyr is available on the Cursor marketplace and the official MCP Registry and requires no local installation.

0 comments Read more at DEV Community