P5 of 21 ▲14 · 75 pts · +1178 this week
online nowclaimedUnited States1 machineearliest result held 2026-11-05
Points
75
2648 behind P1
Signed runs
11
8 in 30 days
Builds measured
11
Best results held
8
3 pts each
Firsts
4
10 pts each
Rooms
0
Carbon per M tokens
991 g
median · modelled · United States grid average
Gram ·4e94b2 is P5 of 21 on the 50GRAMx provider standings with 75 points, from 11 signed runs on 1 machine in United States. Their fastest signed result is MiniCPM-V 4.7 1B at 15 tokens per second. They hold the network's best result on 8 builds and were the first machine anywhere to measure 4. Every number is signed by the machine that measured it.
Fastest models on their machines
Output tokens per second, signed · higher is better
MiniCPM-V 4.7 1B15 tok/s8-bit by openbmb
Llama 3.2 1B11 tok/s
Huihui-Qwen3.6-35B-A3B-abliterated-MTP-GGUF5 tok/s2-bit by huihui-ai
Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved5 tok/s4-bit by llmfan46
gpt-oss 20B5 tok/s4-bit by unslothA million tokens here
Metered on their machine · OpenRouter price beside it where one is sold
MiniCPM-V 4.7 1B$2.09no API sells itLlama 3.2 1B$2.82OpenRouter $0.20Huihui-Qwen3.6-35B-A3B-abliterated-MTP-GGUF$8.15OpenRouter $0.90gpt-oss 20B$8.63OpenRouter $0.13Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved$9.07no API sells itWhat they measured, by family
Signed runs per model family
Every part as the gram benchmarked and signed it.
CPU only
linux · x86_64
CPU
Intel Xeon @ 2.2 GHz
2 cores / 4 threads · 55 MB L3 · 263 hashes/s all cores
Memory
31.3 GB
0.1 GB/s · 308.2 ns latency
Disk
overlayfs
1112 MB/s read · 26 MB/s write · 35870 IOPS
Where
United States
Against every machine on the network
Linux · no GPU · 4 CPUs · 31 GB RAM. Every part is benchmarked and signed by the machine itself.
11 signed runs across 11 builds; 8 in the last 30 days.
MiniCPM-V 4.7 1B (8-bit by openbmb) at 15 tokens per second, signed.
1 point per signed run, 3 per build where the provider holds the network’s fastest result on that kind of machine, 10 per model they were first to measure.
Yes. Each run publishes the exact bytes the machine signed and its signature; “verify this yourself” checks it in your browser against the machine’s own key.
The seasons table, narrowed to this provider’s signed runs: speed and cost on their machines, next to what OpenRouter charges, the checks each model passed, and its licence.
| parameters | tokens/sec | proved | per M tokens | per M tokens | Verdict | ||||
|---|---|---|---|---|---|---|---|---|---|
Text & chat · by openai8 builds · 5 variants | 20B | 10.8 GB | 4.64 CPUs, no GPU· Kaggle | 128k | 8 + 18 lang | $8.63metered | $0.1311 providers | Anyone | API 66× cheaper |
Text & chat · by openbmb2 builds | 1B | 0.6 GB | 154 CPUs, no GPU·Kaggle | 256k | 6 + 3 lang | $2.09metered | not sold | Anyone | only on this network |
Text & chat · by meta-llama10 builds | 1.2B | 0.8 GB | 114 CPUs, no GPU·Kaggle | 32k | 5 + 8 lang | $2.82metered | $0.201 provider | Anyone | API 14× cheaper |
 Image generation · by stabilityai1 build | 900M | 2.4 GB | ran, not timed | — | 1 | — | not sold | Not checked | only on this network |
Speech to text · by openai1 build | 810M | 1.5 GB | ran, not timed | — | 1 + 10 lang | — | not sold | Anyone | only on this network |
 Translation · by facebook1 build | 600M | 2.3 GB | ran, not timed | — | 0 + 12 lang | — | not sold | Own machines only | only on this network |
