Model: Recursive Flint 40B-A2.5B

Recursive released its open coding model Flint 40B-A2.5B on 21 October under Apache 2.0, with 39.8B total parameters and about 2.5B active per token, built for repository work such as tracing bugs and adding tests. Only 14 of the model's 49 layers keep a cache that grows with context, keeping long sessions cheap: bartowski's Q3_K_M GGUF file is 19.53 GB and needs about 20.6 GB of VRAM at 4K context, while Q4_K_M (25.35 GB) needs about 26.7 GB and so requires a 32 GB card. llama.cpp runs the model as `flint`, and Ollama v0.40.8 supports it on macOS only so far; no per-level quality loss has been published.

Recursive Flint 40B-A2.5B 40B-A2.5B Recursive Flint Recursive's open coding model, released on 21 October under Apache 2.0: 39.8B parameters, about 2.5B active per token. It is built for repository work such as tracing bugs and adding tests. Only 14 of its 49 layers keep a cache that grows with context, so long sessions stay cheap. On a 24 GB card the Q3 K M build bartowski's 19.5 GB file needs about 20.6 GB at 4K, which is comfortable, and stays only tight up to 128K. Q4 K M needs about 26.7 GB at 4K, so it wants a 32 GB card; ggml-org's Q4 K M file is about 2 GB smaller than bartowski's. llama.cpp runs it as flint ; Ollama v0.40.8 supports it on macOS only so far. Sizes come from real files; no per-level quality loss has been published. 262K Max Context 3 Quant Variants GGUF Q8 0 Best Quality — Perplexity rise best build Quantization Variants Per-quant VRAM, quality loss, and inference speed on RTX 4090 Measured = run on this site · Estimated = not run here: calculated, or a figure this site has not verified · Community = public reports | Format | Level | BPW | VRAM | PPL Loss | Speed | Source | Actions | |---|---|---|---|---|---|---|---| | GGUF | Q3 K M | 3.92 | 20.6 GB | — | — | Estimated | | | GGUF | Q4 K M | 5.09 | 26.7 GB | — | — | Estimated | | | GGUF | Q8 0 | 8.51 | 44.5 GB | — | — | Estimated | | Which file each size is for A file size is not a memory figure: the VRAM column is the calculator's total at 4K context — weights plus KV cache plus a 10% activation buffer. Units differ on purpose: file sizes are decimal GB, as Hugging Face lists them; the VRAM column is GiB 1024³ bytes with 2% added to the weights, so 7.12 GB of file counts as about 6.8 GiB there. A vision projector mmproj , where the repo ships one, is a separate download and is in neither number. - GGUF Q3 K M bartowski/RecursiveSI flint-GGUF https://huggingface.co/bartowski/RecursiveSI flint-GGUF · RecursiveSI flint-Q3 K M.gguf · 19.53 GB 19,529,151,712 bytes · listing read 2026-10-23 - GGUF Q4 K M bartowski/RecursiveSI flint-GGUF https://huggingface.co/bartowski/RecursiveSI flint-GGUF · RecursiveSI flint-Q4 K M.gguf · 25.35 GB 25,346,651,360 bytes · listing read 2026-10-23 - GGUF Q8 0 bartowski/RecursiveSI flint-GGUF https://huggingface.co/bartowski/RecursiveSI flint-GGUF · RecursiveSI flint-Q8 0.gguf · 42.35 GB 42,350,359,776 bytes · listing read 2026-10-23 Similar models Compare with Seed-OSS 36B https://quantized.uk/tools/compare/?a=flint-40b-a2-5b&b=seed-oss-36b 36BConsumer GPUPro GPU19.5 GBmin VRAM+1.3%perplexity rise https://quantized.uk/quant-hub/seed-oss-36b/ Seed-OSS 36B Instruct ByteDance Seed Dense 36B with a native 512K context. The Q4 K M file is 21.8 GB, so with cache and buffers it needs about 23.9 GB at 4K: tight on a 24 GB card, where Q3 K M 19.5 GB is the comfortable choice, or a 2×16GB split. The AWQ build is 21.1 GB, not the 4-bit rate the name suggests. Context is the real cost: 512K of KV cache is roughly 128GB on its own, so budget context first and weights second. 27BConsumer GPUPro GPU17.5 GBmin VRAM—perplexity rise https://quantized.uk/quant-hub/qwen3-8-27b/ Qwen3.8 27B Alibaba Qwen3.8 Hybrid attention: only 16 of its 64 layers keep a KV cache, which is why a 27B model holds a 262K context in about 16GB of cache rather than 64GB. Sizing it as a conventional stack overstates the cache fourfold — the estimates here count the 16 and match published measurements at 8K, 32K and 262K. Quality loss per level has not been published. 30B-A3BConsumer GPUPro GPU19.2 GBmin VRAM—perplexity rise https://quantized.uk/quant-hub/glm-4.7-flash/ GLM-4.7-Flash Zhipu GLM-4.7 Zhipu's lightweight member of GLM-4.7: 30B total, about 3B active per token, MIT licence. Its attention is MLA, which caches one compressed 576-value vector per token per layer instead of full keys and values — so at 32K context the cache is under 2 GB where a conventional 47-layer model would need several times that. llama.cpp runs it as a DeepSeek-2-style model. Sizes below use the calculator's generic Q4 K M/Q8 0 rates: no GGUF file size could be checked, and quality loss per level has not been published. 26B-A4BConsumer GPUPro GPU18.2 GBmin VRAM—perplexity rise https://quantized.uk/quant-hub/gemma-4-26b-a4b/ Gemma 4 26B-A4B IT Google Gemma 4 Google's Gemma 4 mixture-of-experts model: about 25B parameters in the text model, roughly 4B active per token, a 256K window and image input. Only 5 of its 30 layers keep a cache that grows with context, and those use fewer, wider heads, so the cache stays small — about 0.9 GB at 32K and 5.3 GB at the full 256K. At Q4 K M that is about 18.8 GB at 32K: comfortable on a 24 GB card, which can still load the whole window tight, about 23.6 GB . A 16 GB card cannot hold it, since the Q4 K M file alone is 17.0 GB. The vision encoder is a separate 1.2 GB file on top. Sizes come from bartowski's GGUF files, whose Q4 K M works out at 5.40 bits per weight, well above the generic 4-bit rate; Unsloth's UD-Q4 K M is a different file, 0.09 GB smaller. No per-level quality loss has been published. Running Recursive Flint 40B-A2.5B locally At Q4 K M and 4K of context, Recursive Flint 40B-A2.5B needs about 26.7 GB — 24.1 GB of weights, 0.19 GB of KV cache and a 2.4 GB activation buffer. The smallest cards in this index that clear that comfortably are 32 GB ones — 4 of them, from the RTX 5090 up — and 24 of the 78 cards here do. These are calculated figures, not measurements: the estimate stops counting a card as comfortable at 88% of its VRAM, which is roughly the room a desktop session needs. What longer context costs Going from 4K to 32K adds about 0.77 GB, taking the total to 27.5 GB. That is gentler than the parameter count suggests: only 14 of 49 layers keep a KV cache that grows with context, and 35 more cap theirs at a sliding window, so the cache scales at a fraction of the usual rate. The model's native window is 256K; holding all of it at Q4 K M would need about 34 GB. Which build to download This index only tracks Recursive Flint 40B-A2.5B in GGUF, across 3 levels Q3 K M, Q4 K M, Q8 0 . No per-level perplexity sweep has been published for this model, so the quality column is empty rather than estimated — within one model, more bits per weight is the only ordering the data supports. No throughput has been measured for this model on this site's hardware. GGUF runs on llama.cpp and Ollama across NVIDIA, AMD and Apple silicon; AWQ and GPTQ target vLLM on CUDA and ROCm; EXL2 is ExLlamaV2 — CUDA only, and now archived, so local use rather than serving. Common questions - How much VRAM does Recursive Flint 40B-A2.5B need? - About 26.7 GB at Q4 K M with 4K of context and batch 1, rising to roughly 27.5 GB at 32K. That figure is weights plus KV cache plus a 10% activation buffer, calculated from the model's architecture rather than measured on a card. - Will Recursive Flint 40B-A2.5B run on a 32GB GPU? - Yes — at Q4 K M and 4K context it needs about 26.7 GB, which leaves 5.3 GB spare on a RTX 5090, or on any of the other 32 GB cards here 4 in all . That is the smallest size in this index that clears it comfortably; 24 of 78 do. - Which quantization of Recursive Flint 40B-A2.5B should I use? - No published quality comparison exists for this model, so pick by footprint: Q4 K M is the level most people run, and a higher bits-per-weight level is more faithful. All 3 levels in the index are Q3 K M, Q4 K M, Q8 0. Note this is a mixture-of-experts model — all parameters must be resident even though only a fraction are active per token, so the memory cost follows the total, not the active count. Where this model fits Sized at Q4 K M with a 4K context window, smallest card first. Comfortable means the estimate uses at most 88% of the memory. - Runs comfortably on - RTX 509032GB · +5.3 https://quantized.uk/gpu/rtx-5090/ Arc Pro B70 32G32GB · +5.3 https://quantized.uk/gpu/arc-pro-b70-32g/ Radeon AI PRO R9700 32G32GB · +5.3 https://quantized.uk/gpu/radeon-ai-pro-r9700-32g/ Instinct MI100 32G32GB · +5.3 https://quantized.uk/gpu/instinct-mi100-32g/ A100 40G40GB · +13.3 https://quantized.uk/gpu/a100-40g/ - Tight but possible - Mac M3 Pro 36G36GB https://quantized.uk/gpu/mac-m3-pro-36g/ Mac M4 Max 36G36GB https://quantized.uk/gpu/mac-m4-max-36g/ - Ships in - GGUF https://quantized.uk/formats/gguf/ - A top pick for - Best local LLM for 24GB https://quantized.uk/best/24gb-vram/ - Try it yourself - Size it in the calculator https://quantized.uk/tools/vram-calc/?model=flint-40b-a2-5b&quant=Q4 K M Generate the run command https://quantized.uk/tools/cli-gen/?model=flint-40b-a2-5b&quant=Q4 K M How to actually run this Deployment guides for this model and this class of hardware. This page's figures change when the model or the runtime does.Last updated 2026-10-23 RSS → /feed.xml https://quantized.uk/feed.xml