flint-40b
Architecture mixture of experts49 LayersLicense apache-2.0Maximum context 256kVocabulary 150,144
At Q4_K_M it is a 23.61 GiB file, and it fits on 10 of the 18 machines we measure (those with the least memory: Mini-PC with an NPU and 32 GB LPDDR5X · Mac mini with M6 and 32 GB). Here the file is what fills the memory: even at its maximum context of 256k tokens, the context adds 7.03 GiB — and the same content in German takes ×1.55 the tokens of English.
Model size at Q4_K_M
Measured by Local AI Scope
Context memory at 8k tokens
Measured by Local AI Scope
What German costs versus English
Measured by Local AI Scope
Videos: flint-40b 1
The best walk-throughs we found. Press play and it opens right here.
The videos live on YouTube, and nothing loads from there until you press play (no-cookie mode). The thumbnail is served by us.
Sizes of the real published files
These are the sizes of the files actually published by the author, added up when the model ships split into parts. Not an estimate from the parameter count.
| Compression | Model size |
|---|---|
Q8_0 |
39.44 GiB |
Q6_K |
32.78 GiB |
Q5_K_M |
28.69 GiB |
Q5_K_S |
26.63 GiB |
Q4_K_M |
23.61 GiB |
Q4_K_S |
22.19 GiB |
IQ4_XS |
20.91 GiB |
Q3_K_M |
18.19 GiB |
Context memory
Computed from the attention pattern this model declares layer by layer, not from the generic formula. That is why the figure is often far smaller than other calculators tell you.
- 8k tokens → 0.25 GiB
- 32k tokens → 0.91 GiB
- 128k tokens → 3.53 GiB
What it costs in each language
Measured by running this model’s own tokenizer over the same content in four languages. Tokens are what you pay for in context, in memory and in your cloud bill.
| Kind of text | EN | ES | FR | DE |
|---|---|---|---|---|
| Code | — | ×1.27 | ×1.40 | ×1.59 |
| Contracts | — | ×1.33 | ×1.40 | ×1.67 |
| Business email | — | ×1.23 | ×1.32 | ×1.38 |
| Support tickets | — | ×1.39 | ×1.49 | ×1.52 |
English is the baseline: every figure is how many times more tokens the same content costs in that language. Average across the whole corpus: DE ×1.55.
Where it fits
Fits means the model, its context memory and the working headroom all fit at the stated context, leaving the system its share. 10 of the 18 machine profiles tested (Q4_K_M, 8k tokens).
If you are willing to compress it more, it does run where the table says no: at Q4_K_S (22.19 GiB), up to 8k tokens of context, on Workstation with an RTX 4090 24 GB; at Q3_K_M (18.19 GiB), up to 32k tokens of context, on Mac with M4 Pro and 24 GB. Below Q4_K_M the model loses a little quality, and at Q3_K_M you can tell; that is why the table above sticks to Q4_K_M.
Before you hit download: as of 23 October 2026, Recursive Flint is brand new and the programs that run it are still catching up, so check your version first. llama.cpp opens it from daily build b11803 (from 21 October 2026) onwards; its latest stable release, v0.7.0 (from 15 October 2026), does not. Ollama has run it on Linux and Windows as well as macOS since 0.40.9 (from 23 October 2026), the release that brought the llama.cpp built into Ollama up to build b11803. Heads-up: ollama run flint does not download the table’s Q4_K_M (23.61 GiB). On a Mac with Apple Silicon it fetches an MLX version; on Linux, Windows and Intel Macs, Recursive’s own NVFP4 file (21.29 GiB, meant mainly for NVIDIA cards). By our count the NVFP4 file fits a 24 GB graphics card up to 16k, but Recursive lists it as tested on four NVIDIA systems (B300, RTX PRO 6000 Blackwell, RTX Spark and DGX Spark) and recommends at least 48 GB of GPU memory; we have not checked whether it runs on an RTX 4090. Ollama also offers Recursive’s Q4_1 (flint:40b-a2b-q4_1, 24.21 GiB), meant mainly for Macs (Recursive asks for at least 48 GB of unified memory), and an MLX version (flint:40b-a2b-qat-int4, 21.03 GiB) that already ran on 0.40.8, only on macOS. LM Studio’s team has published a GGUF in its community program (lmstudio-community), whose Q4_K_M weighs 22.51 GiB; we have not checked which version of LM Studio’s engine you need, so update it before you try. mlx-lm does not open it in its latest version, 0.32.2. vLLM added it to its development branch on 22 October 2026, after its latest stable release, 0.32.0, which therefore does not open it yet. And the context: the table stops at 256k, which is what its config.json declares; Recursive says it supports up to 512k and recommends staying within 256k.