Sangama → Results

Results

Test write-ups, newest first. Raw data and full reports are in the repository.

DateTitle
30 Sep 2026Qwen3.5-397B across three US states, with guesses
One request runs at 28 to 49 tokens per second, against 12 without guesses. Guessing with several requests at once showed a fault and is not yet trusted.
30 Sep 2026Qwen3.5-397B across three US states: relay or direct, TCP or QUIC
Direct connections make one request 17% faster; QUIC matches TCP in speed and halves the wait for the first token.
30 Sep 2026Qwen3.5-397B: guesses under load
Guessing ahead is correct with shared steps after a fix, but lowers the total under load: 174 tokens per second against 233 for 16 requests.
30 Sep 2026Qwen3.5-397B: several requests per step
Each GPU works on several waiting requests at once: total throughput from 119 to 403 tokens per second.
30 Sep 2026Qwen3.5-397B in three large stages
Three 96 GB GPUs in one datacenter: 48 tokens per second for one request, up to 107 with the model's draft layer.
30 Sep 2026Qwen3.5-397B: many requests at once, and drafted tokens
96 requests at once on 20 GPUs, up to 90 tokens per second in total; single requests up to 2.2× faster with the model's draft layer.
30 Sep 2026Qwen3.5-397B on 20 GPUs, through a relay
A 397-billion-parameter model on 20 GPUs with 16 GB each; from 0.6 to 5.5 tokens per second.
29 Sep 2026A small model across CPUs, GPUs and the Internet
Identical output on every engine and device; 3.1 tokens per second between a home Mac and a US GPU.

Sangama is open source under the MIT License. Contact: diljit@diljitpr.net · Discord