Model record
Flint 40B
Recursive · Apache 2.0 · released · verified
What this model is
Recursive reports 39.8B total parameters and approximately 2.5B active parameters per token, including the output head. The active body alone is about 2.2B parameters.
The architecture is mixture of experts. Memory must hold the full weight set, while active parameters affect the speed heuristic. Experts can be offloaded to system memory, but that trades speed for capacity.
Memory and quantization
Decimal GB from published GGUF file sizes. NVFP4 is Recursive's official file; the UD entries and Q8_0 are Unsloth dynamic GGUFs. These are text-only model files, not the 80 GB BF16 checkpoint.
| Quantization | Published files | 8K estimate | Smallest comfortable GPU | Smallest comfortable Mac | Context at smallest GPU / Mac | Speed estimate | Run commands |
|---|---|---|---|---|---|---|---|
| UD-IQ1_M | 11.35 GB | 13.09 GB | NVIDIA GeForce RTX 5060 Ti · 16 GB · 16 GB VRAM | Mac mini M6 · 24 GB · 24 GB unified memory | 109,776 / 262,144 tokens | Estimate · Fast relative class | lms get "Flint 40B"@ud-iq1_m |
| UD-IQ2_XXS | 11.86 GB | 13.59 GB | NVIDIA RTX PRO 4000 Blackwell · 24 GB · 24 GB VRAM | Mac mini M6 · 24 GB · 24 GB unified memory | 262,144 / 262,144 tokens | Estimate · Fast relative class | lms get "Flint 40B"@ud-iq2_xxs |
| UD-Q2_K_XL | 13.56 GB | 15.30 GB | NVIDIA RTX PRO 4000 Blackwell · 24 GB · 24 GB VRAM | Mac mini M6 · 24 GB · 24 GB unified memory | 262,144 / 262,144 tokens | Estimate · Fast relative class | lms get "Flint 40B"@ud-q2_k_xl |
| UD-IQ3_XXS | 14.54 GB | 16.27 GB | NVIDIA RTX PRO 4000 Blackwell · 24 GB · 24 GB VRAM | Mac mini M6 · 24 GB · 24 GB unified memory | 262,144 / 262,144 tokens | Estimate · Fast relative class | lms get "Flint 40B"@ud-iq3_xxs |
| UD-Q3_K_XL | 18.49 GB | 20.23 GB | NVIDIA GeForce RTX 5090 · 32 GB · 32 GB VRAM | Mac mini M6 · 32 GB · 32 GB unified memory | 262,144 / 262,144 tokens | Estimate · Fast relative class | lms get "Flint 40B"@ud-q3_k_xl |
| NVFP4 | 22.86 GB | 24.60 GB | NVIDIA GeForce RTX 5090 · 32 GB · 32 GB VRAM | Mac mini M6 · 32 GB · 32 GB unified memory | 262,144 / 262,144 tokens | Estimate · Fast relative class | ollama run flint:40b lms get "Flint 40B"@nvfp4 |
| UD-Q4_K_XL | 25.05 GB | 26.79 GB | NVIDIA RTX PRO 5000 Blackwell · 48 GB (72GB variant) · 48 GB VRAM | Mac mini M5 Pro · 48 GB · 48 GB unified memory | 262,144 / 262,144 tokens | Estimate · Fast relative class | lms get "Flint 40B"@ud-q4_k_xl |
| UD-Q5_K_XL | 29.92 GB | 31.65 GB | NVIDIA RTX PRO 5000 Blackwell · 48 GB (72GB variant) · 48 GB VRAM | Mac mini M5 Pro · 48 GB · 48 GB unified memory | 262,144 / 262,144 tokens | Estimate · Fast relative class | lms get "Flint 40B"@ud-q5_k_xl |
| UD-Q6_K_XL | 36.32 GB | 38.05 GB | NVIDIA RTX PRO 5000 Blackwell · 48 GB (72GB variant) · 48 GB VRAM | Mac mini M5 Pro · 48 GB · 48 GB unified memory | 262,144 / 262,144 tokens | Estimate · Fast relative class | lms get "Flint 40B"@ud-q6_k_xl |
| Q8_0 | 42.30 GB | 44.04 GB | NVIDIA RTX PRO 6000 Blackwell · 96 GB · 96 GB VRAM | Mac Studio M5 Max · 128 GB · 128 GB unified memory | 262,144 / 262,144 tokens | Estimate · Fast relative class | lms get "Flint 40B"@q8_0 |
Estimate formula: published weight files + 0.2349 GB per 8K tokens × context / 8K + 1.5 GB runtime overhead. Sizes are decimal GB, not GiB. At 8K context, the smallest listed build is about 13.1 GB before OS and application reserve.
Conservative FP16 estimate from the 14 full-attention layers. The other 35 layers use a 513-token sliding window and are excluded from the per-token linear estimate.
Ollama tags were checked against the linked Ollama library pages on 2026-11-02. LM Studio's lms get suffix selects a file when the search results expose that exact quantization; confirm the local key with lms ls.
KV method source: https://huggingface.co/RecursiveSI/flint/raw/main/config.json
Speed heuristic
Not benchmarked here. The 2.5B active path suggests a faster relative class than a dense 40B model, but expert routing and offload can dominate real throughput.
Active path: 2.5B parameters. This is a relative speed signal only; it is not a tokens/sec benchmark.
If the full weights do not fit in accelerator memory but do fit in system memory, the calculator marks the result as a tight CPU-offload fit. Expect substantially slower generation and verify the runtime's MoE offload support.
Runtimes
- Ollama
- llama.cpp
- SGLang
ollama run flint:40blms get "Flint 40B"@ud-iq1_m
lms load <model-key-from-lms-ls> --gpu max --context-length 8192License restrictions
Apache 2.0 permits commercial use and modification; retain notices and comply with the patent and termination terms.
This is a plain-language summary, not legal advice. Read the linked license before commercial deployment or redistribution.
Known issues
- Recursive's BF16 serving kit targets one B300 or H200 per replica. The lower-bit GGUF and Ollama builds trade quality and support for consumer hardware.
- The native context is 262K tokens; a long prompt can exhaust memory even after the weight file fits.
- Some community conversions and runtimes may lag the custom attention configuration. Verify the chosen runtime against the linked GGUF card.
If it does not fit, try these smaller records
These alternatives have a lower recommended memory target in the current catalog. Open each page to compare the quantization-specific fit before downloading.
Bonsai 2
Dense · starting at 8 GB
Gemma 4 12B Unified
Dense · starting at 10 GB
NVIDIA Nemotron 3.5 Bolt E4B
Dense · starting at 8 GB
Gemma 4 E4B
Dense · starting at 8 GB
FAQ
Model-specific questions
What is the smallest local file for Flint 40B?
The smallest published weight package in this record is UD-IQ1_M at 11.35 GB. The 8K runtime estimate adds 0.23 GB of KV cache and 1.5 GB overhead, for about 13.1 GB before operating-system reserve.
How much memory do the published quantizations need at 8K?
UD-IQ1_M about 13.1 GB at 8K; UD-IQ2_XXS about 13.6 GB at 8K; UD-Q2_K_XL about 15.3 GB at 8K; UD-IQ3_XXS about 16.3 GB at 8K; UD-Q3_K_XL about 20.2 GB at 8K; NVFP4 about 24.6 GB at 8K; UD-Q4_K_XL about 26.8 GB at 8K; UD-Q5_K_XL about 31.7 GB at 8K; UD-Q6_K_XL about 38.1 GB at 8K; Q8_0 about 44.0 GB at 8K. These are calculated estimates using the published weight sizes and the model's KV-cache assumption, not live memory measurements.
Which GPU and Mac configurations are comfortable for Flint 40B?
Against the current catalog, the smallest configurations that clear the 20% headroom target are: UD-IQ1_M: NVIDIA GeForce RTX 5060 Ti · 16 GB; UD-IQ2_XXS: NVIDIA RTX PRO 4000 Blackwell · 24 GB; UD-Q2_K_XL: NVIDIA RTX PRO 4000 Blackwell · 24 GB; UD-IQ3_XXS: NVIDIA RTX PRO 4000 Blackwell · 24 GB; UD-Q3_K_XL: NVIDIA GeForce RTX 5090 · 32 GB; NVFP4: NVIDIA GeForce RTX 5090 · 32 GB; UD-Q4_K_XL: NVIDIA RTX PRO 5000 Blackwell · 48 GB (72GB variant); UD-Q5_K_XL: NVIDIA RTX PRO 5000 Blackwell · 48 GB (72GB variant); UD-Q6_K_XL: NVIDIA RTX PRO 5000 Blackwell · 48 GB (72GB variant); Q8_0: NVIDIA RTX PRO 6000 Blackwell · 96 GB. A larger configuration may still be preferable for longer context or faster memory bandwidth.
What is the right Ollama or LM Studio command for Flint 40B?
Use the command attached to the specific quantization in the table. Ollama tags are taken from the checked Ollama library pages. For LM Studio, use lms get with the displayed quantization suffix, then load the resulting local key from lms ls with --gpu max and the desired context length.
What should I run instead if Flint 40B does not fit?
Smaller catalog alternatives include Bonsai 2, Gemma 4 12B Unified, NVIDIA Nemotron 3.5 Bolt E4B, Gemma 4 E4B. Compare their quantization tables and minimum memory before choosing; a smaller parameter count does not always mean a proportionally smaller download.
Sources
- https://huggingface.co/RecursiveSI/flint
- https://www.recursive.com/updates/introducing-recursive-flint
- https://huggingface.co/unsloth/flint-GGUF
- https://huggingface.co/RecursiveSI/flint-nvfp4-gguf
- https://ollama.com/library/flint
- https://huggingface.co/RecursiveSI/flint/raw/main/config.json
This page is an independent compatibility guide, not official documentation. Verify the primary source before downloading or deploying a model.