Qwen3.8-Flash-Next vs DeepSeek V4 Flash 0731 284B

Analysis

Qwen3.8-Flash-Next and DeepSeek V4 Flash 0731 shipped four weeks apart, both MoE checkpoints built around a lightweight indexer that scores context in blocks to cut long-context cost, but they land in very different states: DeepSeek's is a finished, agentic-tuned flagship with a real quant ladder, Qwen's is an explicit preview of a future architecture with no confirmed community build yet.

The two indexer schemes work differently underneath the similar pitch. Qwen Sparse Attention (QSA) narrows what each query attends to for compute, but per Qwen's own technical report it still selects from the same full per-token cache a standard full-attention layer needs, so its KV cache grows exactly like an unmodified hybrid stack. DeepSeek's Compressed and Heavily Compressed Attention (CSA/HCA) genuinely folds multiple token positions into one stored entry, a real 10x reduction DeepSeek publishes directly: at the shared 262,144-token context, DeepSeek V4 Flash 0731's KV cache is 2.3 GB against Qwen3.8-Flash-Next's 6.4 GB, despite DeepSeek's context ceiling running four times further, to 1,048,576 tokens, where its cache is still only 9.2 GB. Weights run the opposite direction: DeepSeek is the much bigger checkpoint, 284B total parameters against 180B, activating more than twice as much per token, 13B against 6B, and its real download (Unsloth's natively-repacked quants, since the routed experts ship in MXFP4 rather than a conventional format) recommends UD-Q2_K_XL at 96.8 GB, close to Qwen3.8-Flash-Next's real 180B-parameter total at Q2_K, 68.6 GB. On the one benchmark field both report, the two land almost exactly together despite that 13B-vs-6B compute gap: GPQA Diamond 91.7 for Qwen against 91 for DeepSeek (Artificial Analysis's independently-run figure). Licensing differs in a way that matters practically: DeepSeek ships MIT, no conditions at all, while Qwen's Community 1.0 terms require a separate license for any Model-as-a-Service or 'AI Work Assistant' business. Modality splits too: Qwen accepts text, image, and video, DeepSeek V4 Flash 0731 is text only.

Bottom line: DeepSeek V4 Flash 0731 is the one to actually deploy: it matches Qwen3.8-Flash-Next's GPQA score while activating more than twice the parameters, ships a real quant ladder from Unsloth, and carries the least restrictive license of the two. Qwen3.8-Flash-Next is worth tracking for QSA and the eventual production 'Qwen3.8-Flash', and its own compute-per-token is genuinely lower, but as shipped today it has no confirmed community quant and needs a bigger real download than its 125B/6B headline suggests, because of the 51B n-gram table neither DeepSeek model has an equivalent of.