Falcon 40B — Hardware Requirements & GPU Compatibility
Chat
Falcon-40B is TII's 40-billion-parameter causal decoder-only base language model, pretrained from scratch on 1,000 billion tokens of the RefinedWeb dataset enhanced with curated corpora. At release TII described it as the best fully open model available, outperforming LLaMA, StableLM, and MPT on the Open LLM Leaderboard, with an inference-optimized architecture using FlashAttention and multi-query attention. It is a raw, pretrained model meant to be fine-tuned for chat or other downstream tasks rather than used directly; TII's own Falcon-40B-Instruct is the ready-to-use chat version built on top of it. Trained mainly on English, German, Spanish, and French, it needs a multi-GPU workstation in half precision, while a 4-bit quantization brings it within reach of a single high-memory GPU. License is Apache 2.0, permitting unrestricted commercial and research use. It was published in May 2023.
Usable on a 24 GB GPU
Runs on 24 GB (IQ3_S · ~34 tok/s · 16K)Macs from 24 GB (IQ2_M · ~14 tok/s)
Apache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Falcon 40B
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
ollama run hf.co/mradermacher/falcon-40b-i1-GGUF:Q4_K_MQ4_K_M · needs ~31.7 GB · 16K context
Download: falcon-40b.i1-Q4_K_M.gguf25.49 GBmradermacher
Ways to runOther ways
each opens step-by-stepRun in cloud
Your own private copy · Q4_K_M · 48 GB GPU
- The easiest way to try this model
- Running in a few minutes, easy setup
- Chat right here or connect any app
$5+ bonus for new users
Runs on A100 40GB (8 GB headroom) · Q4_K_M
- Generation speed
- ~37 tok/s
- generation speed
- Cost per 1M output tokens
- $4.62
- per 1M output tokens
How Much VRAM Does Falcon 40B Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 19.6 GB | 28.3 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 16K context · ~34 tok/s | |
| 20.1 GB | 28.9 GB | 3-bit small — 3-bit small quantization | Fits: 8K context · ~33 tok/s | |
| 22.4 GB | 31.2 GB | 3-bit medium — 3-bit medium quantization | Fits: 4K context · ~29 tok/s | |
| 23.0 GB | 31.8 GB | 4-bit · legacy — 4-bit legacy quantization | Fits: 4K context · ~29 tok/s | |
| 27.6 GB | 36.4 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Does not fit: needs 31.7 GB | |
| 32.8 GB | 41.5 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Does not fit: needs 36.9 GB | |
| 38.0 GB | 46.7 GB | 6-bit, very good — 6-bit quantization, very good quality | Does not fit: needs 42.1 GB | |
| 46.0 GB | 54.8 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Does not fit: needs 50.1 GB |
Selected Q4_K_M
4-bit medium quantization31.7GBat 16K context
- Weights
- 26.7 GB
- KV cache
- 4.7 GB
- Overhead
- 0.3 GB
Doesn't fit RTX 4090 (reference) · needs 7.7 GB more, IQ3_S fits
≥30 tok/s needs ≥1.27 TB/s · RTX 5090 ~42
Or .
Which Hardware Can Run Falcon 40B?
Q4_K_M · 27.6 GBFalcon 40B (Q4_K_M) requires 27.6 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 36+ GB is recommended. 1 GPU can run it, including NVIDIA GeForce RTX 5090. 97 devices with unified memory can run Falcon 40B, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Studio M5 Max (36 GB).
20 GB
Q4_K_M · best fit IQ2_M (19.6 GB)
1 GPU
RX 7900 XT~31Simulate output on RX 7900 XT32–40 GB
fits · 0.3 GB free · up to 16K ctx
1 GPU · 3 Mac classes · 1 device family
RTX 5090~42Simulate output on RTX 5090M5 Max36 GB~12Simulate output on Mac Studio M5 Max (36 GB)M532 GB~4Simulate output on MacBook Air 13" M5 (32 GB)Orin~5Simulate output on NVIDIA Jetson AGX Orin 32GB
By VRAM tier · typical GPUs and Macs at Q4_K_M
How good is Falcon 40B?
llmrun Score blends public benchmark boards into one number; 100 = the average open model.
Not enough data yet for an llmrun Score.
A Score needs results on enough independent boards. How the Composite Score is built
Benchmarks
The 6 boards that weigh most in the llmrun Score, of 10 with a result for this model; badges show who ran the test.
- MMLU
Epoch AI · Mar 2023Lab-reported
Lab-reported56.9%
#53/85 open
- BIG-Bench Hard
Epoch AI · Mar 2023Lab-reported
Lab-reported37.1%
#30/38 open
- GSM8K
Epoch AI · Mar 2023Lab-reported
Lab-reported25.0%
#49/62 open
- HellaSwag
Epoch AI · Mar 2023Lab-reported
Lab-reported85.3%
#7/47 open
- ARC Challenge
Epoch AI · Mar 2023Lab-reported
Lab-reported61.9%
#22/56 open
- WinoGrande
Epoch AI · Mar 2023Lab-reported
Lab-reported76.9%
#15/51 open
Alternatives to Falcon 40B
Falcon family · 13 models →Similar models
2| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| FlintRecursiveSI | — | 23.5 GB | ~205 | same size |
| Phi 3.5 MoE InstructMicrosoft | — | 23.9 GB | ~141 | same size |
Specifications
- Publisher
- TII UAE
- Family
- Falcon
- Parameters
- 41.8B
- Architecture
- FalconForCausalLM
- Vocabulary Size
- 65,024
- Release Date
- 2023-05-24
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does Falcon 40B need?
Falcon 40B requires 27.6 GB of VRAM at Q4_K_M, or 92.0 GB at BF16.
VRAM = Weights + KV Cache + Overhead
Weights = 41.8B × 4.8 bits ÷ 8 = 25.1 GB
KV Cache + Overhead ≈ 2.5 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run Falcon 40B?
Yes, at Q3_K_L (23.6 GB) or lower. Higher quantizations like IQ4_XS (24.7 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for Falcon 40B?
For Falcon 40B, Q4_K_M (27.6 GB) offers the best balance of quality and VRAM usage. Q5_K_S (31.6 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 12.7 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Falcon 40B on a Mac?
Yes — Mac Mini M5 Pro (24 GB) and 106 other Macs can run Falcon 40B, the smallest at IQ2_M. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 27.6 GB of usable unified memory (RAM minus macOS overhead), from Mac Studio M1 Max (32 GB).
Can I run Falcon 40B locally?
Yes — Falcon 40B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Falcon 40B?
At Q4_K_M, Falcon 40B can reach ~174 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = (bandwidth GB/s ÷ model GB) × efficiency
Example: NVIDIA B200 → 8000 ÷ 27.6 × 0.65 = ~188 tok/s
Estimated speed at Q4_K_M (27.6 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Falcon 40B?
At Q4_K_M, the download is about 25.10 GB. The full-precision BF16 version is 83.67 GB. The smallest option (IQ2_XXS) is 11.50 GB.
Which GPUs can run Falcon 40B?
1 consumer GPU can run Falcon 40B at Q4_K_M. Top options include NVIDIA GeForce RTX 5090.
Which devices can run Falcon 40B?
97 devices with unified memory can run Falcon 40B at Q4_K_M (27.6 GB), including ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB), GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.