Firing 4.8% of the weights per token still leaves 125GB to keep resident

Qwen published an FP8 variant of the weights on Hugging Face and ModelScope ^[2]. FP8 is one byte per parameter, so the main model's 125B parameters ^[1] come to roughly 125GB resident before a single KV cache entry ^[2]. A 24GB consumer card holds 24GB, so the main weights alone need six of them ^[3].

The 51B N-gram table is the part that can live somewhere cheaper. Per Qwen's release blog, it is a lookup rather than a matrix multiply, so it can be offloaded to host memory and prefetched asynchronously ^[6]. The routed expert weights do not get that treatment. Which slice of the model fires is decided per token ^[1], so the fetch cannot begin before the router has chosen.

Against the dense sibling, the two costs move in opposite directions. Qwen3.8-27B fires all 27B of its parameters on every token and fits in roughly 17-19GB of VRAM quantized ^[12], so Flash-Next does about 22% of the per-token parameter work ^[4] while carrying roughly seven times the resident footprint ^[5].

The headline score is 62.5 on SWE-bench Pro, which the write-up labels vendor-reported ^[9]. For that to transfer you need the same harness, the same attempt budget, and issues that resemble SWE-bench Pro's; the post reports neither the harness nor an absolute training cost. The cost figure is vendor-reported on the same terms: Qwen reports training at roughly one-ninth of Qwen3.7-Plus alongside stronger coding and office-task results ^[10], and the one-ninth figure is Alibaba's training cost either way. The dev.to post says to "treat that as a vendor claim until independent replication" ^[13].

Context is 262,144 tokens natively and extensible to 1M with YaRN-style scaling ^[8]. The reason the long end is tractable is the hybrid attention: Gated DeltaNet layers compress history into a fixed-size state in linear time, and Qwen Sparse Attention uses a lightweight indexer to attend only to the micro-blocks of context that matter. Full attention, the thing being avoided here, grows quadratically with sequence length ^[5]. The post's pitch for the 1M configuration is that chunking, embedding and ranking collapse into one prompt against the raw corpus ^[11]. Anyone taking that trade moves spend out of a retrieval stack they operate and into context they pay for on every call.

On the single-card case the post is direct. It says Flash-Next is not for single-24GB-GPU local inference, because the weights footprint is far beyond consumer cards ^[7], and its recommendation is: "If you want a reliable local coding assistant on one consumer GPU, stay dense for now." ^[14]