readygpu

FrogNano 4B 2609

4.7B parameters · 256k context · Apache License 2.0 · released 2026-09-22

codingagenticreasoning

In short

FrogNano 4B 2609 is an open model from Microsoft with 4.7B parameters. At Q4_K_M, FrogNano 4B 2609 needs about 4 GB of memory at 8k context and 4 GB at 8k, counting the weights, the KV cache and runtime buffers. The cheapest hardware in Germany that runs FrogNano 4B 2609 fully in memory at 8k context is the NVIDIA GeForce RTX 3060 12GB, 309 € at Amazon, at about 45 tok/s (estimated).

Updated . Speeds are estimates; how they are worked out.

What to run it on in Germany

Fully in memory at 8k context, priced at retailers today.

All hardware that runs it

Everything we price, best value first. Two cards split the model by layers.

HardwareQuantFitsDecodePricetok/s per 1,000
NVIDIA GeForce RTX 3060 12GBQ8_0fits45 tok/s309 €146.3
Intel Arc B580Q8_0fits46 tok/s366 €126.0
NVIDIA Tesla P40Q8_0fits43 tok/s369 €117.6
Intel Arc A770 16GBQ8_0fits57 tok/s390 €145.4
NVIDIA GeForce RTX 3080 10GBQ8_0fits95 tok/s555 €171.9
AMD Radeon RX 9060 XT 16GBQ8_0fits36 tok/s592 €60.1
NVIDIA GeForce RTX 5060 Ti 16GBQ8_0fits56 tok/s701 €80.2
NVIDIA GeForce RTX 4060 Ti 16GBQ8_0fits36 tok/s734 €49.2
NVIDIA GeForce RTX 5070Q8_0fits84 tok/s798 €105.6
Mac mini M4 16GBQ8_0fits16 tok/s800 €19.8
AMD Radeon RX 7600 XTQ8_0fits32 tok/s813 €39.5
AMD Radeon RX 9070 XTQ8_0fits71 tok/s835 €85.3
AMD Radeon RX 7900 XTQ8_0fits89 tok/s899 €99.1
NVIDIA GeForce RTX 4070 SuperQ8_0fits63 tok/s950 €66.5
AMD Radeon RX 7800 XTQ8_0fits70 tok/s997 €69.7
Intel Arc Pro B60Q8_0fits46 tok/s1.074 €42.9
NVIDIA GeForce RTX 4080 SuperQ8_0fits92 tok/s1.100 €84.0
NVIDIA GeForce RTX 5070 TiQ8_0fits112 tok/s1.186 €94.7
NVIDIA GeForce RTX 4070 Ti SuperQ8_0fits84 tok/s1.247 €67.6
Mac mini M4 32GBQ8_0fits16 tok/s1.450 €10.9
NVIDIA GeForce RTX 5080Q8_0fits121 tok/s1.500 €80.3
AMD Radeon RX 7900 XTXQ8_0fits107 tok/s1.555 €68.7
NVIDIA GeForce RTX 3090Q8_0fits118 tok/s1.850 €63.5
AMD Radeon AI PRO R9700Q8_0fits71 tok/s2.390 €29.8
Mac mini M4 Pro 48GBQ8_0fits36 tok/s2.699 €13.3
NVIDIA GeForce RTX 3090 TiQ8_0fits127 tok/s2.790 €45.3
Mac mini M4 Pro 64GBQ8_0fits36 tok/s3.099 €11.6
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits29 tok/s3.700 €7.7
NVIDIA GeForce RTX 4090Q8_0fits127 tok/s3.955 €32.0
Mac Studio M4 Max 64GBQ8_0fits72 tok/s4.000 €18.0
Mac Studio M4 Max 128GBQ8_0fits72 tok/s5.800 €12.4
Mac Studio M3 Ultra 96GBQ8_0fits108 tok/s5.999 €18.0
NVIDIA GeForce RTX 5090Q8_0fits225 tok/s6.042 €37.2
NVIDIA DGX Spark 128GBQ8_0fits34 tok/s8.313 €4.1
ASUS Ascent GX10 128GBQ8_0fits34 tok/s8.540 €4.0
Mac Studio M3 Ultra 256GBQ8_0fits108 tok/s15.555 €6.9
NVIDIA RTX PRO 6000 BlackwellQ8_0fits225 tok/s18.527 €12.1
Mac Studio M3 Ultra 512GBQ8_0fits108 tok/s18.995 €5.7
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits29 tok/s––
Framework Desktop Ryzen AI Max+ 395 64GBQ8_0fits29 tok/s––
Beelink GTR9 Pro Ryzen AI Max+ 395 128GBQ8_0fits29 tok/s––

Memory needed

Weights, KV cache and runtime buffers at each context length (KV cache at fp16).

QuantFile@8k@32k@64k@128k
Q4_K_Mgguf2.6 GB3.6 GB4.5 GB5.7 GB8.1 GB
Q5_K_Mgguf3.1 GB4.2 GB5.1 GB6.3 GB8.6 GB
Q6_Kgguf3.5 GB4.5 GB5.4 GB6.6 GB9.0 GB
Q8_0gguf4.3 GB5.3 GB6.2 GB7.4 GB9.8 GB

Run it

Copy a command for the runtime you use.

  • Q4_K_M
    • Ollamaollama run hf.co/bartowski/FrogNano-4B-2609-GGUF:Q4_K_M
    • LM Studiolms get bartowski/FrogNano-4B-2609-GGUF@q4_k_m
    • llama.cppllama-server -hf bartowski/FrogNano-4B-2609-GGUF:Q4_K_M
  • Q5_K_M
    • Ollamaollama run hf.co/bartowski/FrogNano-4B-2609-GGUF:Q5_K_M
    • LM Studiolms get bartowski/FrogNano-4B-2609-GGUF@q5_k_m
    • llama.cppllama-server -hf bartowski/FrogNano-4B-2609-GGUF:Q5_K_M
  • Q6_K
    • Ollamaollama run hf.co/bartowski/FrogNano-4B-2609-GGUF:Q6_K
    • LM Studiolms get bartowski/FrogNano-4B-2609-GGUF@q6_k
    • llama.cppllama-server -hf bartowski/FrogNano-4B-2609-GGUF:Q6_K
  • Q8_0
    • Ollamaollama run hf.co/bartowski/FrogNano-4B-2609-GGUF:Q8_0
    • LM Studiolms get bartowski/FrogNano-4B-2609-GGUF@q8_0
    • llama.cppllama-server -hf bartowski/FrogNano-4B-2609-GGUF:Q8_0

Notes

Card metadata says license: mit but the body states Apache License 2.0; verify before commercial use. Specialized for Python-heavy repo tasks with the Leaf harness; needs a Qwen3.5-capable runtime (SGLang used), and image/video are unsupported.

Questions

How much VRAM does FrogNano 4B 2609 need?
At Q4_K_M, FrogNano 4B 2609 needs about 4 GB of memory at 8k context and 4 GB at 8k, counting the weights, the KV cache and runtime buffers.
What is the best GPU for running FrogNano 4B 2609 locally?
The cheapest hardware in Germany that runs FrogNano 4B 2609 fully in memory at 8k context is the NVIDIA GeForce RTX 3060 12GB, 309 € at Amazon, at about 45 tok/s (estimated). The best value is the NVIDIA GeForce RTX 3080 10GB (about 95 tok/s for 555 €). The fastest we price is the NVIDIA GeForce RTX 5090, at about 225 tok/s.
Can I run FrogNano 4B 2609 on a 16 GB GPU?
Yes: at Q8_0 it fits fully in 16 GB at 8k context. On the NVIDIA GeForce RTX 5080 it runs at about 121 tok/s (estimated).
Can I run FrogNano 4B 2609 on a 24 GB GPU?
Yes: at Q8_0 it fits fully in 24 GB at 8k context. On the NVIDIA GeForce RTX 3090 Ti it runs at about 127 tok/s (estimated).
How do I run FrogNano 4B 2609 locally?
With Ollama: ollama run hf.co/bartowski/FrogNano-4B-2609-GGUF:Q4_K_M (the Q4_K_M build). LM Studio and llama.cpp commands are on this page.
What licence is FrogNano 4B 2609 released under?
Apache License 2.0, which allows commercial use.