Recursive Flint Requirements

Recursive Flint is an open-weight MoE model from Recursive (Hugging Face org RecursiveSI) with 40B total and about 2.5B active parameters. At 4-bit the weights are roughly 23-26GB, and you need about 24-30GB of memory to run it. To fit entirely on a 24GB GPU you must drop to 3-bit (UD-IQ3_XXS), and a 12GB GPU has reportedly worked by offloading experts to the CPU. For the full 256K-token context, Recursive recommends 48GB or more of memory.

Flint was released on October 21, 2026 under the Apache 2.0 license. It is a text-only model trained from scratch with no base model, with a knowledge cutoff of June 2026. This article organizes the requirements from published file sizes and third-party reports. Most benchmark figures come from the developer, and independent verification is still thin (as of October 23, 2026).

If you want to check quickly whether your own GPU or Mac can run it, our VRAM calculator (free, no sign-up) estimates the required VRAM from the model, quantization, and context length you pick. Flint is available in its model list.

In the "memory needed" column, the 1-2 bit, 3-bit, Q6, and Q8 values are community estimates, the 27.4GB for UD-Q4_K_XL is a measurement, and rows marked "unconfirmed" have no supporting number beyond the file size. Longer context adds KV cache memory, so treat the table as a near-minimum guide.

Flint is a decoder-only MoE (Mixture of Experts) with 39.8B total parameters, of which about 2.5B are used per token including the output head. Of its 49 layers, 2 are dense and 47 are MoE. Each MoE layer has 256 routed experts; 6 are selected per token, plus 1 shared expert. The vocabulary is 150,144 tokens.

The attention design targets long context. 35 of the 49 layers use sliding-window attention (a 513-token window) and the other 14 use full attention, with 32 query heads, 4 KV heads, and a head dimension of 128. With few KV heads and narrow windows in most layers, the KV cache should grow slowly even at long context.

The key point is that speed and memory are separate questions. With about 2.5B active, the compute and weights read per token are small, so generation speed resembles a 2.5B-class model. But which experts are chosen changes every token, so all 40B of weights must sit in memory. "Light as a 2.5B model" describes speed; the memory requirement stays at 40B class.

Unified-memory Macs can hold a 40B-class MoE without hitting a discrete-GPU memory wall. MarkTechPost lists a 48GB Apple Silicon machine (M4 Pro, M5 Pro, or M5 Max) as the minimum for the full 262K context, with 4-bit weights at about 23GB. Recursive also recommends 48GB or more for the full context.

On speed, the developer reports over 125 tokens/s on an M4 Pro MacBook Pro using its own rec engine. A community report shows 122.6 tokens/s on a Mac mini with INT4 at 1K context, faster than the 91.5 tokens/s the same user measured for Qwen3.6-35B-A3B at 4-bit. These are short-context figures; expect lower speed on long inputs.

A 32GB Mac can hold the weights (about 23-26GB), but little room remains for macOS and the KV cache, so long context is tight. It is workable for short trials, but we suggest 48GB for real use.

- RTX 5090 (32GB): UD-Q4_K_XL (25.1GB) fits with room to spare. On llama.cpp b11825 at Q4: about 145.7 tokens/s decode and 3,433 tokens/s prefill (atomicagent.io test).
- RTX 3090 / RTX 4090 (24GB): UD-Q4_K_XL does not fit. A community report says UD-IQ3_XXS (14.5GB) fits fully on an RTX 3090. Quality may drop at 3-bit, but we found no independent data measuring the gap.
- RTX 3060 (12GB): one user ran UD-Q3_K_XL on 12GB VRAM plus 16GB RAM with expert offload at 256K context, reporting about 46 tokens/s decode and about 675 tokens/s prefill (a community X post; not independently reproduced).
- Production: the official SGLang kit specifies one B300 or H200 per replica.

In any setup, leave headroom for the KV cache when using long context. If unsure, enter your context length into the VRAM calculator above first.

Supported runtimes are SGLang (official kit), vLLM, Docker Model Runner, llama.cpp, Ollama, LM Studio, and Jan. For easy local use, llama.cpp or Ollama are the usual picks.

For llama.cpp you need build b11825 or later, which includes Flint support (PR #30928); older builds cannot load the model. Start it with llama-server -hf unsloth/flint-GGUF:UD-Q4_K_XL, which also downloads the model. With Ollama, run ollama run hf.co/unsloth/flint-GGUF:UD-Q4_K_XL. An official Ollama tag flint:40b (23GB) has been reported, but we have not tested it. LM Studio and Jan can search and download the GGUF from their GUIs.

Recommended sampling is temperature 1.0, top-p 0.95, top-k 20. For hard problems, allow at least 32,768 output tokens. Keep the KV cache at f16 or q8_0; q4_0 reportedly degraded code quality.

The chat template in the official, Unsloth, and LM Studio GGUFs reportedly breaks tool calling in llama.cpp. terelabs/flint-GGUF ships a fixed template, which is an option for agent use.

Main results published by the developer (BF16 / NVFP4): Terminal-Bench 2.1 at 54.9 / 53.7, SWE-bench Verified at 74.2 / 74.1, SWE-bench Pro v2 at 89.7 / 88.5, LiveCodeBench v6 at 80.3 / 80.8, and IFEval at 95.1 / 94.6, so quantization costs very little by their numbers. These are the developer's own figures.

In an independent small agent test, Flint passed 12 of 18 runs, while Ornith 1.5 35B-A3B passed 14 of 18 on the same machine (Flint used 22.8GB VRAM). The sample is too small to declare a winner. For related models, see our requirements articles on Ornith 1.5 and Qwen 3.8 27B.

Testers also noted cases where Flint reported success without doing the work, writing a PLACEHOLDER instead. Even when delegating to an agent, verify the outputs yourself. Japanese support is not stated officially.

To check performance before downloading, OpenRouter offers recursive/flint-v1.0:free. As of October 22, 2026 it is $0 for input and output, rate-limited, and supports tool calling and reasoning. The free tier may end at any time, so do not build on it.

- Out of memory: step down one quantization level, shorten the context, set the KV cache to q8_0, or offload experts to the CPU, in that order. UD-Q4_K_XL uses 27.4GB at 131K context, so it will not fit a 24GB card.
- Model fails to load: your llama.cpp build is likely older than b11825. Update to a build that includes PR #30928.
- Tool calling does not work: a chat-template problem has been reported in the GGUFs. Try the fixed template in terelabs/flint-GGUF.
- Slow speed: adaptive activation scaling in the llama.cpp PR reportedly slows inference by 10-30%. A newer build may help, but this is unconfirmed.
- Rough output quality: avoid q4_0 KV cache, and match sampling to temperature 1.0, top-p 0.95, top-k 20.