Alibaba launches Qwen3.8-Max claiming top OSWorld score, but experts urge caution

·1 views

Alibaba released Qwen3.8-Max on August 3, a 2.4-trillion-parameter mixture-of-experts model with a 1-million-token context window, priced at $2/$6 per million tokens. The company claims an OSWorld-Verified score of 86.1, placing it ahead of competing models from OpenAI and Anthropic in agentic computer-use benchmarks. However, the model's license has not yet been disclosed, and open weights have only been promised for a future date, raising questions about how genuinely accessible the release will be. Analysts note that the 2.4T flagship is impractical to self-host, making a promised smaller 27B sibling the more consequential release for independent developers. All benchmark figures cited so far come from Alibaba's own materials, and independent verification on public leaderboards is still pending.

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)

Arrays Beat Linked Lists in Practice Due to CPU Cache Behavior

Computer science courses teach that linked lists outperform arrays for insertions, citing O(1) versus O(n) complexity, but real-world benchmarks tell a different story. A test iterating over 10 million integers showed a contiguous array completing in 0.53 seconds compared to 1.53 seconds for a linked list — nearly three times slower despite identical operation counts. The gap stems from how modern CPUs handle memory: a single 64-byte cache line fetch delivers eight array elements at once, while each linked list node requires a separate memory trip. This phenomenon, known as pointer chasing, forces the processor to wait for the current node's data before it can determine the next memory address, eliminating any chance of prefetching. As a result, cache efficiency — not algorithmic complexity — is often the dominant factor in real hardware performance.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Rust Powers High-Throughput Solana Swap Data Ingestion on a 4-CPU Pod

A technical deep-dive published on DEV Community outlines how to build a production-grade Solana transaction ingestion engine using Rust. The system tracks real-time swap activity across decentralized exchanges like Raydium, Orca, and Jupiter by maintaining persistent WebSocket connections to Helius RPC nodes instead of relying on traditional REST polling. To handle connection failures, the service uses an asynchronous reconnection loop with exponential backoff, automatically recovering from dropped sockets without manual intervention or data loss. Memory overhead is minimized by filtering out irrelevant transactions at the RPC layer and using zero-copy decoding via the Carbon crate, which converts raw byte arrays directly into typed Rust structures. The architecture is designed to sustain thousands of transactions per second on cost-effective infrastructure while decoupling ingestion from the database through a NATS message broker.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

EverAfter Platform Turns Wedding Invitations into Interactive AR and AI Experiences

A developer has built EverAfter, a digital wedding invitation platform that combines augmented reality, AI-generated themes, 3D models, and personalized media. The platform allows couples to create invitations featuring animated backgrounds, music, videos, and interactive 3D bride and groom models. Invitations can be shared via QR codes, replacing static digital cards with an immersive experience. The key challenge was integrating multiple technologies — AR, AI, and 3D content — while keeping the interface simple for end users. The finished platform is live at arweddingcard.com.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer ports decimal.js to Go in 72 hours, uncovers 2 bugs in original library

A solo developer ported decimal.js, a JavaScript arbitrary-precision arithmetic library, to Go during a 72-hour hackathon called Port Mortem 2026. The project involved replicating over 20 mathematical operations including trigonometry, transcendental functions, and nine rounding modes with zero dependencies. Rather than correcting errors, the developer prioritized behavioral parity, arguing that downstream users depend on the original library's exact output, including its quirks. To verify accuracy, a cross-validation harness was built that ran 1,518 test cases through both implementations, achieving byte-for-byte identical results. The process ultimately revealed three bugs in the Go port and two genuine upstream bugs in the original decimal.js library, discovered through differential fuzzing.

0 comments Read more at DEV Community