Benchmark v1

Real agents, an open model, old hardware. Every number labelled.

These are measured results with the hardware, model, image and method on the page. They are single runs, so they are evidence and not certified bars. Certified results come with our paper.

What was measured, and how much to trust it.

Two things were measured. First, real agents calling a real open-weight model through the one outbound checkpoint. Second, how many calls a single VM can hold open at once, for five VM sizes.

Each result below is a single run. Our own rule is 3 runs, report the median, and treat more than 10% spread between runs as a failure. Single runs do not meet that rule, so they are labelled as evidence. The same harness will run three times on the final image, and those certified numbers will be published with our paper.

The memory figures are the runtime's total runtime memory as the harness reported it, not operating-system RSS.

Real agents, one open model, one checkpoint.

Real open model · 2 vCPU / 4 GB only

Every model call goes through the single outbound checkpoint. Two separate runs are shown here, and the numbers are not added together. Real-model runs on every VM size, 3 runs each, are in progress for the paper.

Measured, single run, 2026-09-23, pre-release image cf922365, 2013 Intel i7-4770 under KVM, open-weight model llama3.2:1b on CPU.

896calls in the three concurrency runs (128 + 256 + 512 at 4, 8 and 16 agents). Checkpoint tally: allow 896, deny 0. 0 delivery failures.

768calls in the per-task run (24 tasks, 32 runs each, 8 at a time). Checkpoint tally: allow 768, deny 0. 0 failed.

3 of 3controls behaved: a host not on the list was refused, an emptied trust store failed closed with no downgrade, and a scan of the node's own output for the provisioned credential came back clean.

5 of 24tasks were answered correctly on all 32 runs by the 1B model. Others were not. The tables below show every task.

Correctness by tier: this measures the model.

The 24 tasks fall into four tiers. A one-billion-parameter model gets some tiers right and others wrong. That is a property of the model, and this table measures the model, not VOLENT.

Measured, single run, 2026-09-23, pre-release image cf922365, 2013 Intel i7-4770 under KVM, open-weight model llama3.2:1b on CPU.

Tier	Correct	Share	p50 ms	p95 ms
recall	163/192	85 %	1,353	2,282
arithmetic	67/192	35 %	1,124	3,500
reasoning	49/192	26 %	4,160	12,296
structured	167/192	87 %	4,352	9,598

All 24 tasks.

Each task ran 32 times. Correct is runs answered correctly out of runs scored. Times are milliseconds per call on CPU inference.

Measured, single run, 2026-09-23, pre-release image cf922365, 2013 Intel i7-4770 under KVM, open-weight model llama3.2:1b on CPU.

#	Tier	Task	Correct	p50 ms	p95 ms
1	recall	What is the capital city of Japan?	32/32	1,585	2,287
2	recall	Which planet is closest to the Sun?	23/32	1,642	2,181
3	recall	What is the chemical symbol for water?	24/32	1,648	2,400
4	recall	Which ocean is the largest?	28/32	1,084	1,817
5	recall	How many days are in a week?	27/32	1,051	1,954
6	recall	What is the freezing point of water in Celsius?	29/32	1,424	2,664
7	arithmetic	Compute 17 plus 25.	3/32	1,116	1,555
8	arithmetic	Compute 8 times 7.	12/32	1,062	1,934
9	arithmetic	Compute 144 divided by 12.	5/32	1,259	1,859
10	arithmetic	A box holds 6 eggs. How many eggs are in 9 boxes?	16/32	2,524	10,920
11	arithmetic	Compute 100 minus 37.	11/32	937	1,491
12	arithmetic	What is one quarter of 80?	20/32	1,172	1,914
13	reasoning	Anna is taller than Ben. Ben is taller than Carl. Who is the shortest?	14/32	2,518	8,779
14	reasoning	A train leaves at 09:00 and the journey takes 3 hours. At what time does it arrive?	13/32	9,496	12,832
15	reasoning	All roses are flowers. This object is a rose. Is it a flower?	12/32	1,173	2,082
16	reasoning	Sam had 12 apples, gave away 5, then bought 3 more. How many does Sam have now?	5/32	9,218	13,150
17	reasoning	If today is Monday, what day will it be in 3 days?	0/32	1,087	1,802
18	reasoning	A shirt costs 20 dollars and is discounted by 25 percent. What is the new price in dollars?	5/32	8,556	13,090
19	structured	Reply with exactly this JSON and nothing else: {"status": "ok", "count": 3}	31/32	4,968	9,598
20	structured	List exactly three primary colours, separated by commas, in this order: red, then yellow, then b	32/32	4,461	9,797
21	structured	Extract the year from this sentence and reply with JSON of the form {"year": NNNN}: The treaty w	32/32	7,535	11,858
22	structured	Reply with a JSON object whose only key is "capital" and whose value is the capital of France.	32/32	4,840	6,605
23	structured	Summarise in one short sentence, beginning with the word Water: Water boils at 100 degrees Celsi	32/32	4,016	6,184
24	structured	Reply with a comma-separated list of the first four positive even numbers, digits only.	8/32	2,806	7,885

More agents at once.

Agents ran against the same open model on a 4 GB, 2 vCPU guest. Throughput flattened as agents were added, while memory stayed at 78 MB. The model server was the ceiling, not the runtime.

Measured, single run, 2026-09-23, pre-release image cf922365, 2013 Intel i7-4770 under KVM, open-weight model llama3.2:1b on CPU.

Agents	Calls	Completed	p50 ms	p95 ms	Throughput	Total runtime memory
4	128	128	1,203	7,521	1.40/s	78 MB
8	256	256	1,674	11,118	2.11/s	78 MB
16	512	512	5,353	14,592	2.23/s	78 MB

How many calls one VM can hold open.

The model is replaced by a 5 ms stand-in, so the test measures the runtime alone. Five VM sizes, four load levels, 20 runs. 0 refused in all 20 runs. 150,000 calls in flight fit on 1 vCPU and 1 GB at 212 MB. More processors did not help, because the work is wait-bound: each call spends its time waiting.

Capacity figures use a 5 ms mock model, so they measure VOLENT, not a model. Certified three-run results on the final image will follow with our paper.

Measured, single run, 2026-09-23, pre-release image cf922365, 2013 Intel i7-4770 under KVM, open-weight model llama3.2:1b on CPU.

VM size	1,500 in flight	10,000 in flight	50,000 in flight	150,000 in flight
1 vCPU / 1 GB	0 refused · 79 MB	0 refused · 86 MB	0 refused · 123 MB	0 refused · 212 MB
2 vCPU / 2 GB	0 refused · 81 MB	0 refused · 92 MB	0 refused · 140 MB	0 refused · 217 MB
2 vCPU / 4 GB	0 refused · 79 MB	0 refused · 92 MB	0 refused · 111 MB	0 refused · 217 MB
4 vCPU / 8 GB	0 refused · 82 MB	0 refused · 90 MB	0 refused · 112 MB	0 refused · 336 MB
8 vCPU / 16 GB	0 refused · 85 MB	0 refused · 95 MB	0 refused · 144 MB	0 refused · 222 MB

Each cell is calls refused, then total runtime memory. The highlighted cell, 4 vCPU and 8 GB at 150,000 in flight, used 336 MB where its neighbours used 212 to 222 MB. It is unexplained, and we left it in.

Same tasks, five VM sizes.

For each VM size and each load: how many tasks were asked for, how many finished in parallel, how long it took, and how much runtime memory it used. Every task finished, and none were refused.

Capacity test with a 5 ms mock model — measures VOLENT's own scheduling, not model speed · single run · 2026-09-23 · pre-release image cf922365 · 2013 Intel i7-4770

VM size	Tasks asked for	Completed in parallel	Time taken	Tasks per second	Runtime memory
1 vCPU / 1 GB	1,500	1,500	69 ms	21,739 /s	79 MB
1 vCPU / 1 GB	10,000	10,000	1.3 s	7,570 /s	86 MB
1 vCPU / 1 GB	50,000	50,000	33.0 s	1,517 /s	123 MB
1 vCPU / 1 GB	150,000	150,000	347.5 s	432 /s	212 MB
2 vCPU / 2 GB	1,500	1,500	69 ms	21,739 /s	81 MB
2 vCPU / 2 GB	10,000	10,000	1.4 s	7,013 /s	92 MB
2 vCPU / 2 GB	50,000	50,000	41.0 s	1,219 /s	140 MB
2 vCPU / 2 GB	150,000	150,000	363.0 s	413 /s	217 MB
2 vCPU / 4 GB	1,500	1,500	95 ms	15,790 /s	79 MB
2 vCPU / 4 GB	10,000	10,000	1.5 s	6,859 /s	92 MB
2 vCPU / 4 GB	50,000	50,000	41.9 s	1,192 /s	111 MB
2 vCPU / 4 GB	150,000	150,000	360.4 s	416 /s	217 MB
4 vCPU / 8 GB	1,500	1,500	89 ms	16,854 /s	82 MB
4 vCPU / 8 GB	10,000	10,000	1.7 s	5,963 /s	90 MB
4 vCPU / 8 GB	50,000	50,000	41.0 s	1,219 /s	112 MB
4 vCPU / 8 GB	150,000	150,000	344.3 s	436 /s	336 MB
8 vCPU / 16 GB	1,500	1,500	109 ms	13,762 /s	85 MB
8 vCPU / 16 GB	10,000	10,000	1.6 s	6,219 /s	95 MB
8 vCPU / 16 GB	50,000	50,000	42.4 s	1,180 /s	144 MB
8 vCPU / 16 GB	150,000	150,000	339.3 s	442 /s	222 MB

Each row is one single run. Time is wall-clock time from the first task to the last. More processors did not make the largest loads faster, because each task spends its time waiting.

Completed, and how long it took

1 vCPU / 1 GB

1,500 tasks69 ms
10,000 tasks1.3 s
50,000 tasks33.0 s
150,000 tasks347.5 s

2 vCPU / 2 GB

1,500 tasks69 ms
10,000 tasks1.4 s
50,000 tasks41.0 s
150,000 tasks363.0 s

2 vCPU / 4 GB

1,500 tasks95 ms
10,000 tasks1.5 s
50,000 tasks41.9 s
150,000 tasks360.4 s

4 vCPU / 8 GB

1,500 tasks89 ms
10,000 tasks1.7 s
50,000 tasks41.0 s
150,000 tasks344.3 s

8 vCPU / 16 GB

1,500 tasks109 ms
10,000 tasks1.6 s
50,000 tasks42.4 s
150,000 tasks339.3 s

Every bar is full because every task asked for was completed. The label is the time it took.

The clean baseline, from the tagged image.

These came from a later, cleaner pass on the image built from the phase-7 tag. Where a test ran three times, the median is shown.

Measured 2026-10-14 on the image built from tag phase-7, median of 3 runs where run three times, 2013 Intel i7-4770 under KVM.

What	Measured	Detail
Cold start, power-on to agents running	3.486 s	spread 3.0%, limit 5 s
Total runtime memory, 10,000 idle agents	40.5 MB	about 4 KB each
Control API p99 at 10,000 agents	50 µs	limit 100 ms
Concurrent calls in flight	1,500	0 refused, 2 vCPU, mock model server
Forced crashes at five points	0 lost	0 duplicates; 100 agents × 3 steps, 400/400 completed
Sandboxed tool runs	1,000/1,000	0 crashed
Root payload	53 MB	282/282 third-party crates digest-verified

100,000 idle agents in one computer.

Measured 2026-10-17 on the image built from tag phase-7, median of 3 runs, 4 GB / 2 vCPU guest, 2013 Intel i7-4770 under KVM.

Total runtime memory 446 MB for 100,000 idle agents, about 4.5 KB each, with the three runs within 0.15% of each other. In every run, 99 of 100 control requests were answered in under 0.1 ms. A 250,000-agent run did not finish within two hours, so we publish no number for it. Raw files: bench/archive/2026-10-17-acer-idle-100k.

150,000 agents, confirmed on a second machine.

Measured 2026-10-18 on the image built from tag phase-7, median of 3 runs, 4 GB / 2 vCPU guest, Dell PowerEdge T430 (Xeon E5-2620 v3) under KVM.

The same image, run on a different computer. 150,000 agents waiting for work used 729.7 MB in total, about 4.9 KB each, with the three runs within 0.07% of each other. One run on the original test machine used 730 MB. In every run, 99 of 100 control requests were answered in under 0.15 ms. Raw files: bench/archive/2026-10-18-dell-idle-150k.

Method.

Everything needed to judge these numbers, and to check them against the raw files.

Hardware. One Intel i7-4770 from 2013, 8 cores, 31 GB, running guests under KVM. Capacity guests came in five shapes: 1 vCPU with 1 GB, 2 with 2 GB, 2 with 4 GB, 4 with 8 GB and 8 with 16 GB. The real-model runs used 2 vCPU with 4 GB.
Image. A pre-release build, digest cf922365, taken before the final phase closed. It is not the frozen release image.
Model. llama3.2:1b through Ollama on the host, CPU inference, over TLS and HTTP/2, reached through the checkpoint. The capacity runs use a stand-in model with a fixed 5 ms delay.
Corpus. 24 tasks in four tiers: recall, arithmetic, reasoning and structured output. Corpus digest 9d19af46e999. 32 runs per task, 8 at a time.
Rule. Our standard is 3 runs, median reported, and more than 10% spread counts as a failure. These are single runs and say so.
Raw data. The raw files are archived byte for byte, with sha256 checksums for each. The archive tarball's sha256 starts d9bb2569. Every cell in the tables above is checked against those files before this page is published.
To rerun. Boot the image on a machine like this one, point it at a local open-weight model, run the 24-task corpus 32 times at 8 at once, then run the five VM sizes at the four load levels. The same harness will run three times on the final image.

How to compare

Snapshot and warm-pool products quote start-up times on modern servers. This is a cold start of a whole sealed system on 2013 hardware. We do not quote any competitor's numbers here.