# EdgeLLM Bench

> Open leaderboard of real, sourced local-LLM speed results (tokens per second) on edge boards: Raspberry Pi 5/4/CM4, NVIDIA Jetson Orin Nano and RK3588 boards (Orange Pi 5, Radxa ROCK 5B+, Turing Pi RK1), paired with India retail prices. 111 results, 10 boards, 70 models. Every number links to the public post it came from.

Gen t/s is token generation speed (usually llama-bench tg128). Prompt t/s is prompt processing (usually pp512). Results from different runtimes, quants, threads and cooling are not strictly comparable. India prices are the cheapest listing seen on the date shown, not live prices. Model size buckets: small is 1-3B params, medium is 7-8B, big is 14B and up.

## Boards

- [Jetson AGX Orin 64GB](https://edgellm-bench.vercel.app/boards/jetson-agx-orin-64gb): 1 result; no size buckets yet; no India listing yet; fastest result 25.8 t/s with Qwen3.8-Flash-Next Gyro-M TQ2_0 (2-bit trellis experts) (Strata Orin fork bb98d09 (CUDA SM87, MTP speculative decoding, MAXN)), source https://github.com/kim-haneol/Strata-Orin/blob/main/docs/OPTIMIZATION.md
- [Jetson Orin Nano (original)](https://edgellm-bench.vercel.app/boards/jetson-orin-nano-original): 7 results; small models median 24.7 t/s over 3 tests; medium models median 14.0 t/s over 3 tests; no India listing yet; fastest result 41.0 t/s with SmolLM2 INT4 (MLC (jetson-containers)), source https://www.jetson-ai-lab.com/archive/benchmarks.html
- [Jetson Orin Nano Super](https://edgellm-bench.vercel.app/boards/jetson-orin-nano-super): 27 results; small models median 38.1 t/s over 13 tests; medium models median 19.1 t/s over 3 tests; cheapest India price seen Rs 67,980 at Amazon.in (seller Gamingicons) (8GB dev kit, checked 2026-10-17); fastest result 159.5 t/s with SmolLM2 135M Q4_K_M (llama.cpp b9292 (MAXN power mode)), source https://yuvrajsingh.io/posts/jetson-nano-super-benchmark-non-reasoning/
- [Orange Pi 5 (RK3588S)](https://edgellm-bench.vercel.app/boards/orange-pi-5-rk3588s): 5 results; small models median 6.9 t/s over 1 test; cheapest India price seen Rs 29,000 at IndiaMART (Avinya, Pune) (16GB kit, checked 2026-10-17); fastest result 55.2 t/s with Granite 4.0 350M Q4_0 (rk-llama.cpp fork (CPU)), source https://github.com/milhcbt/rk-llama.cpp
- [Orange Pi 5 Plus (RK3588)](https://edgellm-bench.vercel.app/boards/orange-pi-5-plus-rk3588): 6 results; small models median 6.2 t/s over 6 tests; cheapest India price seen Rs 84,864 at Amazon.in (seller Waterberry) (16GB, checked 2026-10-17); fastest result 25.9 t/s with Qwen2.5 1.5B W8A8 (rknn-llm (NPU)), source https://jishuzhan.net/article/2099289570509246466
- [Radxa ROCK 5B+ (RK3588)](https://edgellm-bench.vercel.app/boards/radxa-rock-5b-rk3588): 4 results; small models median 17.0 t/s over 1 test; big models median 6.8 t/s over 2 tests; no India listing yet; fastest result 17.0 t/s with Gemma 3 1B IT Q8_0 (rk-llama.cpp fork (NPU)), source https://github.com/Mojo24x7/rk-llama.cpp
- [Raspberry Pi 4](https://edgellm-bench.vercel.app/boards/raspberry-pi-4): 3 results; small models median 1.1 t/s over 3 tests; cheapest India price seen Rs 18,497 at Robu.in (8GB, checked 2026-10-16); fastest result 1.6 t/s with Gemma-derived ~1B (Ollama), source https://medium.com/@thomasnahon/llms-on-a-budget-testing-serving-frameworks-on-the-raspberry-pi-4-5fc56623840e
- [Raspberry Pi 5](https://edgellm-bench.vercel.app/boards/raspberry-pi-5): 52 results; small models median 5.4 t/s over 19 tests; medium models median 1.9 t/s over 14 tests; big models median 1.3 t/s over 17 tests; cheapest India price seen Rs 18,643 at Tech Depot (8GB, checked 2026-10-17); fastest result 33.8 t/s with Qwen2.5 0.5B Instruct Q4_0 (llama.cpp (CPU)), source https://github.com/ggml-org/llama.cpp/discussions/30069
- [Raspberry Pi CM4](https://edgellm-bench.vercel.app/boards/raspberry-pi-cm4): 5 results; small models median 2.1 t/s over 3 tests; medium models median 0.8 t/s over 2 tests; cheapest India price seen Rs 6,185 at Hubtronics (4GB/8GB eMMC (CM4004008), checked 2026-10-17); fastest result 5.1 t/s with TinyLlama 1.1B Chat (Ollama), source https://gist.github.com/hammad93/eab28862024e480a2046d51bc39bd462/7e502ea4019400a75f4fc5c369992b273f2a5398
- [Turing Pi RK1 (RK3588)](https://edgellm-bench.vercel.app/boards/turing-pi-rk1-rk3588): 1 result; small models median 14.0 t/s over 1 test; no India listing yet; fastest result 14.0 t/s with LFM2.5 2.6B Q4_K_M (llama.cpp CPU (4 threads)), source https://turingpi.com/lfm2-5-2-6b-rk3588-llama-cpp-turing-pi-rk1/

## Buying guides and comparisons

- [Best board for running LLMs in India under ₹10,000 (2026, real benchmarks)](https://edgellm-bench.vercel.app/guides/best-llm-board-india-under-10k): Which edge AI board under ₹10,000 runs local LLMs best? Real tokens/s data and India prices.
- [Best board for running LLMs in India under ₹20,000 (2026, real benchmarks)](https://edgellm-bench.vercel.app/guides/best-llm-board-india-under-20k): Which edge AI board under ₹20,000 runs local LLMs best? Real tokens/s data and India prices.
- [Best board for running LLMs in India under ₹35,000 (2026, real benchmarks)](https://edgellm-bench.vercel.app/guides/best-llm-board-india-under-35k): Which edge AI board under ₹35,000 runs local LLMs best? Real tokens/s data and India prices.
- [Best board for running LLMs in India under ₹75,000 (2026, real benchmarks)](https://edgellm-bench.vercel.app/guides/best-llm-board-india-under-75k): Which edge AI board under ₹75,000 runs local LLMs best? Real tokens/s data and India prices.
- [Jetson Orin Nano Super vs Orange Pi 5 (RK3588S) for LLMs: real benchmarks + India prices](https://edgellm-bench.vercel.app/guides/jetson-orin-nano-super-vs-orange-pi-5-rk3588s): Compare Jetson Orin Nano Super, Orange Pi 5 (RK3588S) for running local LLMs: tokens/s by model size and INR prices.
- [Raspberry Pi 5 vs Jetson Orin Nano Super vs Orange Pi 5 (RK3588S) for LLMs: real benchmarks + India prices](https://edgellm-bench.vercel.app/guides/raspberry-pi-5-vs-jetson-orin-nano-super-vs-orange-pi-5): Compare Raspberry Pi 5, Jetson Orin Nano Super, Orange Pi 5 (RK3588S) for running local LLMs: tokens/s by model size and INR prices.
- [Raspberry Pi 5 vs Jetson Orin Nano Super for LLMs: real benchmarks + India prices](https://edgellm-bench.vercel.app/guides/raspberry-pi-5-vs-jetson-orin-nano-super): Compare Raspberry Pi 5, Jetson Orin Nano Super for running local LLMs: tokens/s by model size and INR prices.
- [Raspberry Pi 5 vs Orange Pi 5 (RK3588S) for LLMs: real benchmarks + India prices](https://edgellm-bench.vercel.app/guides/raspberry-pi-5-vs-orange-pi-5-rk3588s): Compare Raspberry Pi 5, Orange Pi 5 (RK3588S) for running local LLMs: tokens/s by model size and INR prices.

## Key pages

- [Home and full leaderboard](https://edgellm-bench.vercel.app/): sortable table of every result with filters
- [Boards](https://edgellm-bench.vercel.app/boards) and [Models](https://edgellm-bench.vercel.app/models): one page per board and per model
- [Wanted tests](https://edgellm-bench.vercel.app/wanted): board and model-size combinations with no results yet
- [Submit a result](https://edgellm-bench.vercel.app/submit): send a benchmark with a public link to its raw output
- [About and methodology](https://edgellm-bench.vercel.app/about)
- [GPU Rent Index](https://edgellm-bench.vercel.app/gpu): hourly cloud GPU rental prices (H100, A100, L40S, L4, RTX 4090) in INR and USD with source and date
- [Advertise](https://edgellm-bench.vercel.app/advertise)

## Data

- [results.json](https://edgellm-bench.vercel.app/results.json): every result with board, RAM, model, params, quant, runtime, prompt and gen t/s, date, source URL and credit
- [results.csv](https://edgellm-bench.vercel.app/results.csv): same data as CSV
- [prices.json](https://edgellm-bench.vercel.app/prices.json): India retailer listings with price, stock and date checked
- [gpu_prices.json](https://edgellm-bench.vercel.app/gpu_prices.json): cloud GPU hourly prices with provider, currency, FX rate, source URL and date checked

## Optional

- [Can a board run BitNet b1.58 2B?](https://edgellm-bench.vercel.app/models/bitnet-b1-58-2b)
- [Can a board run Bonsai 8B?](https://edgellm-bench.vercel.app/models/bonsai-8b)
- [Can a board run Dolphin 2.6 Mistral 7B DPO laser?](https://edgellm-bench.vercel.app/models/dolphin-2-6-mistral-7b-dpo-laser)
- [Can a board run GLM-4.7-Flash 30B-A3B?](https://edgellm-bench.vercel.app/models/glm-4-7-flash-30b-a3b)
- [Can a board run Gemma 2 2B?](https://edgellm-bench.vercel.app/models/gemma-2-2b)
- [Can a board run Gemma 2 9B?](https://edgellm-bench.vercel.app/models/gemma-2-9b)
- [Can a board run Gemma 2B Instruct?](https://edgellm-bench.vercel.app/models/gemma-2b-instruct)
- [Can a board run Gemma 3 12B IT?](https://edgellm-bench.vercel.app/models/gemma-3-12b-it)
- [Can a board run Gemma 3 1B?](https://edgellm-bench.vercel.app/models/gemma-3-1b)
- [Can a board run Gemma 3 1B IT?](https://edgellm-bench.vercel.app/models/gemma-3-1b-it)
- [Can a board run Gemma 3 1B IT Heretic?](https://edgellm-bench.vercel.app/models/gemma-3-1b-it-heretic)
- [Can a board run Gemma 4 26B-A4B IT?](https://edgellm-bench.vercel.app/models/gemma-4-26b-a4b-it)
- [Can a board run Gemma 4 E2B IT?](https://edgellm-bench.vercel.app/models/gemma-4-e2b-it)
- [Can a board run Gemma 4 E2B IT Heretic?](https://edgellm-bench.vercel.app/models/gemma-4-e2b-it-heretic)
- [Can a board run Gemma 4 E4B IT?](https://edgellm-bench.vercel.app/models/gemma-4-e4b-it)
- [Can a board run Gemma 4 E4B IT Heretic?](https://edgellm-bench.vercel.app/models/gemma-4-e4b-it-heretic)
- [Can a board run Gemma-derived ~1B?](https://edgellm-bench.vercel.app/models/gemma-derived-1b)
- [Can a board run Granite 4.0 350M?](https://edgellm-bench.vercel.app/models/granite-4-0-350m)
- [Can a board run Hermes 3 8B?](https://edgellm-bench.vercel.app/models/hermes-3-8b)
- [Can a board run LFM2.5 1.2B?](https://edgellm-bench.vercel.app/models/lfm2-5-1-2b)
- [Can a board run LFM2.5 2.6B?](https://edgellm-bench.vercel.app/models/lfm2-5-2-6b)
- [Can a board run LFM2.5 350M?](https://edgellm-bench.vercel.app/models/lfm2-5-350m)
- [Can a board run Llama 3 8B?](https://edgellm-bench.vercel.app/models/llama-3-8b)
- [Can a board run Llama 3 8B Instruct?](https://edgellm-bench.vercel.app/models/llama-3-8b-instruct)
- [Can a board run Llama 3.1 8B?](https://edgellm-bench.vercel.app/models/llama-3-1-8b)
- [Can a board run Llama 3.1 8B IT?](https://edgellm-bench.vercel.app/models/llama-3-1-8b-it)
- [Can a board run Llama 3.1 8B IT Heretic?](https://edgellm-bench.vercel.app/models/llama-3-1-8b-it-heretic)
- [Can a board run Llama 3.2 1B?](https://edgellm-bench.vercel.app/models/llama-3-2-1b)
- [Can a board run Llama 3.2 1B IT?](https://edgellm-bench.vercel.app/models/llama-3-2-1b-it)
- [Can a board run Llama 3.2 3B?](https://edgellm-bench.vercel.app/models/llama-3-2-3b)
- [Can a board run Llama 3.2 3B IT?](https://edgellm-bench.vercel.app/models/llama-3-2-3b-it)
- [Can a board run Magistral Small?](https://edgellm-bench.vercel.app/models/magistral-small)
- [Can a board run Ministral 3 14B IT?](https://edgellm-bench.vercel.app/models/ministral-3-14b-it)
- [Can a board run Ministral 3 3B IT?](https://edgellm-bench.vercel.app/models/ministral-3-3b-it)
- [Can a board run Ministral 3 8B IT?](https://edgellm-bench.vercel.app/models/ministral-3-8b-it)
- [Can a board run Mistral 3 14B?](https://edgellm-bench.vercel.app/models/mistral-3-14b)
- [Can a board run OLMo 2 1B IT?](https://edgellm-bench.vercel.app/models/olmo-2-1b-it)
- [Can a board run Phi 3.5 3B?](https://edgellm-bench.vercel.app/models/phi-3-5-3b)
- [Can a board run Phi-2 (dolphin-2.6)?](https://edgellm-bench.vercel.app/models/phi-2-dolphin-2-6)
- [Can a board run Phi-3?](https://edgellm-bench.vercel.app/models/phi-3)
- [Can a board run Phi-3 3.8B?](https://edgellm-bench.vercel.app/models/phi-3-3-8b)
- [Can a board run Phi-4 Mini Instruct?](https://edgellm-bench.vercel.app/models/phi-4-mini-instruct)
- [Can a board run Phi-4 Mini Reasoning?](https://edgellm-bench.vercel.app/models/phi-4-mini-reasoning)
- [Can a board run Phi-4 Reasoning Plus?](https://edgellm-bench.vercel.app/models/phi-4-reasoning-plus)
- [Can a board run Qwen2.5 0.5B?](https://edgellm-bench.vercel.app/models/qwen2-5-0-5b)
- [Can a board run Qwen2.5 0.5B Instruct?](https://edgellm-bench.vercel.app/models/qwen2-5-0-5b-instruct)
- [Can a board run Qwen2.5 1.5B?](https://edgellm-bench.vercel.app/models/qwen2-5-1-5b)
- [Can a board run Qwen2.5 3B?](https://edgellm-bench.vercel.app/models/qwen2-5-3b)
- [Can a board run Qwen2.5 7B?](https://edgellm-bench.vercel.app/models/qwen2-5-7b)
- [Can a board run Qwen3 0.6B?](https://edgellm-bench.vercel.app/models/qwen3-0-6b)
- [Can a board run Qwen3 30B-A3B?](https://edgellm-bench.vercel.app/models/qwen3-30b-a3b)
- [Can a board run Qwen3 4B Instruct 2507?](https://edgellm-bench.vercel.app/models/qwen3-4b-instruct-2507)
- [Can a board run Qwen3-Coder 30B-A3B?](https://edgellm-bench.vercel.app/models/qwen3-coder-30b-a3b)
- [Can a board run Qwen3.5 0.8B?](https://edgellm-bench.vercel.app/models/qwen3-5-0-8b)
- [Can a board run Qwen3.5 122B-A10B?](https://edgellm-bench.vercel.app/models/qwen3-5-122b-a10b)
- [Can a board run Qwen3.5 27B?](https://edgellm-bench.vercel.app/models/qwen3-5-27b)
- [Can a board run Qwen3.5 2B?](https://edgellm-bench.vercel.app/models/qwen3-5-2b)
- [Can a board run Qwen3.5 35B-A3B?](https://edgellm-bench.vercel.app/models/qwen3-5-35b-a3b)
- [Can a board run Qwen3.5 4B?](https://edgellm-bench.vercel.app/models/qwen3-5-4b)
- [Can a board run Qwen3.5 9B?](https://edgellm-bench.vercel.app/models/qwen3-5-9b)
- [Can a board run Qwen3.6 35B-A3B?](https://edgellm-bench.vercel.app/models/qwen3-6-35b-a3b)
- [Can a board run Qwen3.8-Flash-Next Gyro-M?](https://edgellm-bench.vercel.app/models/qwen3-8-flash-next-gyro-m)
- [Can a board run SmolLM2?](https://edgellm-bench.vercel.app/models/smollm2)
- [Can a board run SmolLM2 1.7B IT?](https://edgellm-bench.vercel.app/models/smollm2-1-7b-it)
- [Can a board run SmolLM2 135M?](https://edgellm-bench.vercel.app/models/smollm2-135m)
- [Can a board run SmolLM2 360M?](https://edgellm-bench.vercel.app/models/smollm2-360m)
- [Can a board run SmolLM3 3B?](https://edgellm-bench.vercel.app/models/smollm3-3b)
- [Can a board run TinyLlama 1.1B Chat?](https://edgellm-bench.vercel.app/models/tinyllama-1-1b-chat)
- [Can a board run gpt-oss 120B?](https://edgellm-bench.vercel.app/models/gpt-oss-120b)
- [Can a board run gpt-oss 20B?](https://edgellm-bench.vercel.app/models/gpt-oss-20b)