Alibaba's Qwen team has taken the top spot on the newly refreshed LiveCodeBench v6 coding leaderboard, evaluated on November 1, 2026, with its Qwen3.8-Flash-Next model scoring 91.90 percent and edging out sibling Qwen3.7-Max, Moonshot's Kimi K2.6, ByteDance's Seed 2.0 Pro and NVIDIA's Nemotron 3 Ultra. The result underscores a shift from raw scale toward architectural efficiency: a 125-billion-parameter mixture-of-experts model that fires only about 6 billion parameters per token, paired with a hybrid Gated DeltaNet and Sparse Attention design that Alibaba positions as a preview of the Qwen4 architecture.
A new leaderboard, a familiar winner
LiveCodeBench v6, updated on November 2, 2026, ranks 68 models on freshly collected competitive-programming problems. Qwen3.8-Flash-Next leads at 91.90 percent, with the production Qwen3.8 Flash close behind at the same score, and Qwen3.7-Max third at 91.60 percent. The breadth of the Qwen family at the top — five Qwen entries inside the top ten — shows how much of the leaderboard Alibaba now occupies.
Efficiency over brute force
Qwen3.8-Flash is a multimodal mixture-of-experts model with 125 billion total parameters but only about 6 billion active per token, cutting the memory and compute footprint versus dense peers. A 262,144-token native context window extends to roughly 1 million tokens through YaRN. Alibaba reports training cost at about one-ninth of Qwen3.7-Plus, a model three times its size in total parameters, and prices the API at 0.16 dollars per million input tokens and 0.47 dollars per million output tokens in international markets.
Why it matters
As frontier labs keep scaling parameter counts into the trillions, Qwen3.8-Flash argues that sparse activation and smarter attention can deliver flagship coding performance at a fraction of the cost. For developers, that means data-center-grade coding agents are increasingly runnable on modest hardware, and the per-token economics of agentic coding keep falling.
Key facts
LiveCodeBench v6 score: 91.90 percent, rank 1 of 68 models (evaluated November 1, 2026).
Architecture: 125 billion total parameters, about 6 billion active per token (mixture-of-experts).
Context: 262,144 tokens native, extendable to about 1,000,000 via YaRN.
API price: 0.16 dollars per million input tokens, 0.47 dollars per million output tokens.
Training cost: about one-ninth of Qwen3.7-Plus.
Sources: LiveCodeBench v6 leaderboard (llmboard.ai), Alibaba Qwen release notes, threerouter
Top comments (0)