Flint 40B (MoE) vs Qwen 3.8 27B

Analysis

Flint 40B and Qwen 3.8 27B answer the same question from opposite ends: how do you get a strong local coding agent onto one GPU? Qwen 3.8 27B is a dense model that fires all 27B of its parameters on every token (released 14 August 2026, Apache 2.0, text, image and video input). Flint 40B is a sparse mixture of experts that fires about 2.5B of its 39.8B (released 21 October 2026, Apache 2.0, text only). One is built for quality per gigabyte and the other for speed per gigabyte, and the numbers below show what each gives up.

Weights first. Qwen's Q4_K_M is 16.4 GB against 24.2 GB for Flint, because Flint stores all 39.8B parameters even though it computes with a small slice of them. That makes Qwen the easier fit on a 24 GB card at a short context: it fits an RTX 4090 at Q4_K_M with 8k of context (19.0 GB in total), while Flint needs a 32 GB card. Longer context narrows the gap. Qwen keeps 16 of its 64 layers as Gated Attention (4 KV heads of width 256) and gives the other 48 a fixed-size state; Flint caps 35 of its 49 layers at a 513-token window and grows only 14 (4 KV heads of width 128). At 262,144 tokens Qwen's cache is 17.2 GB against Flint's 7.6 GB, and the full-window totals are 37.7 GB for Qwen and 35.6 GB for Flint. The two totals cross at about 213,000 tokens: below that Qwen needs less memory, above it Flint does, and at the full window neither fits a 32 GB card. Speed is where the mixture of experts pays. This site's estimates put Flint at 194 tok/s on an RTX 5090 at Q4_K_M with 8k of context against 68.5 for Qwen, about 2.8 times faster, and at the full window, with both spilling to system RAM, 78.9 against 6.6. Recursive's launch post claims more than five times the speed of dense 25 to 40B models, a vendor figure measured on its own engine. Quality favors Qwen. On the one benchmark both report, Terminal-Bench 2.1, Qwen scores 73.0 against Flint's 54.9, each on its own vendor's harness (Qwen's card names Terminus, Recursive's names Terminus 2, and Recursive counts solutions found through web retrieval as failures). Qwen also reports GPQA Diamond 89.2, LiveCodeBench v6 90.3 and SWE-bench Pro 61.7; Flint's card gives SWE-bench Verified 74.2 and 89.7 on SWE-bench Pro v2, a different suite from the one Qwen reports, so those two figures do not compare. Qwen also reads images and video where Flint is text only, and Recursive's own post says Flint may need more guidance than the larger closed coding models. Tooling is level: both have an Ollama tag (qwen3.8:27b at 18 GB, flint:40b at 23 GB) and both have GGUF builds for llama.cpp.

Bottom line: Pick Qwen 3.8 27B if quality is the priority, you want image or video input, or you have a 24 GB card and a short context. Pick Flint 40B if decode speed matters most and you have a 32 GB card or a 48 to 64 GB Mac: this site's estimates put it at close to three times Qwen's speed at 8k tokens, and its cache stays smaller than Qwen's beyond roughly 213,000 tokens of context. Neither company's benchmark claims are independently verified and Flint was a day old when this page was written, so run both on your own repository.