Analytics ältere Läufe

Was die Läufe gekostet haben, gelesen aus den aufgezeichneten Rundenergebnissen. Wallzeit und Prefix-Cache sind keine Leistungsmaße — sie messen Parallellast und Serverwärme mit.

Durchsatz

ZeileRtok/sTTFT sPrompt maxCacheReasoningTool-AnteilAufrufegleichzeitig
v3-hermes-qwen-ninfer338.20.09555870%68%18%72
v3-hermes-qwen-ninfer121.40.07313350%73%9%162
v3-hermes-qwen-ninfer212.20.10532010%82%16%252
v3-hermes-qwen-ninfer311.10.21969810%87%16%1022
v3-hermes-qwen-ninfer422.70.10488290%86%0%113
v3-hermes-qwen-ninfer59.00.251137860%83%20%822
v3-hermes-qwen-ninfer64.30.641904130%84%26%1462
v3-hermes-qwen-ninfer74.70.861016570%86%16%852

gelb: ≥ 4 gleichzeitig offene Streams beim Aufruf — tok/s und TTFT stehen dann unter Last.

Zeitstrahl

E-20260908T203536 E-20260908T203536 R1 cc-fable-5-1 cc-fable-5-1 R1 cc-fable-5-1 R2 cc-fable-5-1 R3 cc-fable-5-1 R4 cc-fable-5-1 R5 cc-fable-5-1 R6 cc-fable-5-1 R7 cc-opus-5 cc-opus-5 R1 cc-opus-5 R2 cc-opus-5 R3 cc-opus-5 R4 cc-opus-5 R5 cc-opus-5 R6 cc-opus-5 R7 cc-sonnet-5 cc-sonnet-5 R1 cc-sonnet-5 R2 cc-sonnet-5 R3 cc-sonnet-5 R4 cc-sonnet-5 R5 cc-sonnet-5 R6 cc-sonnet-5 R7 hermes-claude-fable-5-1 hermes-claude-fable-5-1 R1 hermes-claude-fable-5-1 R2 hermes-claude-fable-5-1 R3 hermes-claude-fable-5-1 R4 hermes-claude-fable-5-1 R5 hermes-claude-fable-5-1 R6 hermes-claude-fable-5-1 R7 hermes-claude-haiku-4-5 hermes-claude-haiku-4-5 R1 hermes-claude-haiku-4-5 R2 hermes-claude-haiku-4-5 R3 hermes-claude-haiku-4-5 R4 hermes-claude-haiku-4-5 R5 hermes-claude-haiku-4-5 R6 hermes-claude-haiku-4-5 R7 hermes-claude-opus-5 hermes-claude-opus-5 R1 hermes-claude-opus-5 R2 hermes-claude-opus-5 R3 hermes-claude-opus-5 R4 hermes-claude-opus-5 R5 hermes-claude-opus-5 R6 hermes-claude-opus-5 R7 hermes-claude-sonnet-5 hermes-claude-sonnet-5 R1 hermes-claude-sonnet-5 R2 hermes-claude-sonnet-5 R3 hermes-claude-sonnet-5 R4 hermes-claude-sonnet-5 R5 hermes-claude-sonnet-5 R6 hermes-claude-sonnet-5 R7 hermes-qwen hermes-qwen R1 hermes-qwen R2 hermes-qwen R3 hermes-qwen R4 hermes-qwen R5 hermes-qwen R6 hermes-qwen R7 hermes-qwen R1 hermes-qwen R2 hermes-qwen-prune-low hermes-qwen-prune-low R1 hermes-qwen-prune-low R2 hermes-qwen-prune28k hermes-qwen-prune28k R1 hermes-qwen-prune28k R2 hermes-qwen-prune28k R3 hermes-qwen-prune28k R4 hermes-qwen-prune28k R5 hermes-qwen-prune28k R6 v3-hermes-qwen-ninfer v3-hermes-qwen-ninfer R1 v3-hermes-qwen-ninfer R2 v3-hermes-qwen-ninfer R3 v3-hermes-qwen-ninfer R1 v3-hermes-qwen-ninfer R2 v3-hermes-qwen-ninfer R3 v3-hermes-qwen-ninfer R4 v3-hermes-qwen-ninfer R5 v3-hermes-qwen-ninfer R6 v3-hermes-qwen-ninfer R7 v3-pi-qwen-ninfer v3-pi-qwen-ninfer R1 v3-pi-qwen-ninfer R2 v3-pi-qwen-ninfer R3 v3-pi-qwen-ninfer R1 v3-pi-qwen-ninfer R2 v3-pi-qwen-ninfer R3 v3-pi-qwen-ninfer R4 v3-pi-qwen-ninfer R5 v3-pi-qwen-ninfer R6 v3-pi-qwen-ninfer R7 v3-pi-qwen-ninfer R1 v3-pi-qwen-ninfer R2 v3-pi-qwen-ninfer R3 v3-pi-qwen-ninfer R4 v3-pi-qwen-ninfer R5 gleichzeitig laufende Runden, max 7

Gesamt

17Läufe
92Runden
44.3Stunden Agentenzeit
46.3 MPrompt-Token
3.02 Merzeugte Token

E-20260908T203536

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
14/6—/2416371,371,15583,1532,2470.91ok—

20260917T020302

hermes · Qwen3.8-27B-ninfer · prune aus

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
16/5—/2419local runtime without vLLM counters (ninfer :8100)okfdcb2b6b
keine Proxy-Daten
26/5—/2438local runtime without vLLM counters (ninfer :8100)okc28b3be0
keine Proxy-Daten
36/5—/2425local runtime without vLLM counters (set ARENA_VLLM_METRICS okc28b3be0 unverändert
400400400

20260916T205515

pi · Qwen3.8-27B-ninfer

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
10/5—/246local runtime without vLLM counters (ninfer :8100)oke3b0c442
keine Proxy-Daten
20/5—/246local runtime without vLLM counters (ninfer :8100)oke3b0c442 unverändert
keine Proxy-Daten
30/5—/2410local runtime without vLLM counters (ninfer :8100)oke3b0c442 unverändert
keine Proxy-Daten

20260916T201230

pi · Qwen3.8-27B-ninfer

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
10/5—/240local runtime without vLLM counters (ninfer :8100)rc=1e3b0c442
keine Proxy-Daten
20/5—/240local runtime without vLLM counters (ninfer :8100)rc=1e3b0c442 unverändert
keine Proxy-Daten
30/5—/240local runtime without vLLM counters (ninfer :8100)rc=1e3b0c442 unverändert
keine Proxy-Daten
40/5—/240local runtime without vLLM counters (ninfer :8100)rc=1e3b0c442 unverändert
keine Proxy-Daten
50/5—/240local runtime without vLLM counters (ninfer :8100)rc=1e3b0c442 unverändert
keine Proxy-Daten
60/5—/240local runtime without vLLM counters (ninfer :8100)rc=1e3b0c442 unverändert
keine Proxy-Daten
70/5—/240local runtime without vLLM counters (ninfer :8100)rc=1e3b0c442 unverändert
keine Proxy-Daten

20260916T200540

pi · Qwen3.8-27B-ninfer

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
10/5—/240TimeoutError: timed outrc=1e3b0c442
keine Proxy-Daten
20/5—/240TimeoutError: timed outrc=1e3b0c442 unverändert
keine Proxy-Daten
30/5—/240TimeoutError: timed outrc=1e3b0c442 unverändert
keine Proxy-Daten
40/5—/240TimeoutError: timed outrc=1e3b0c442 unverändert
keine Proxy-Daten
50/5—/240TimeoutError: timed outrc=1e3b0c442 unverändert
keine Proxy-Daten

20260917T050549

hermes · Qwen3.8-27B-ninfer · ctx 235.520 · prune aus

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
16/5—/2416local runtime without vLLM counters (set ARENA_VLLM_METRICS ok059d33ea
400
26/5—/2434local runtime without vLLM counters (set ARENA_VLLM_METRICS okea1219e8
400
36/5—/24168local runtime without vLLM counters (set ARENA_VLLM_METRICS ok77c69d14
400
46/5—/2422local runtime without vLLM counters (set ARENA_VLLM_METRICS okf2b27ea2
400
56/5—/2487local runtime without vLLM counters (set ARENA_VLLM_METRICS ok598c7092
400
66/5—/24363local runtime without vLLM counters (set ARENA_VLLM_METRICS ok04b6676c
400
76/5—/2491local runtime without vLLM counters (set ARENA_VLLM_METRICS ok3ccde97d
400

J-20260909T163414--hermes-qwen-prune-low

hermes · Qwen3.8-27B-Instruct · ctx 65.536 · prune 28000

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
16/610/2447531,832,743187,0433,5290.725ok202c6658
20/60/24901587,457,572470,5212,9770.733harness-failure19b6433a

I-20260909T085125--hermes-qwen-prune28k

hermes · Qwen3.8-27B-Instruct · ctx 65.536 · prune 28000

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
16/610/2431341,302,87477,5242,2800.931ok—
26/610/2429TimeoutError: timed outok—
36/614/2416332113,668,303796,7922,4820.808ok—
46/617/2487195,98524,7303,5320.21ok—
56/617/2444461,763,555187,6734,0790.571harness-failure—
66/617/241310337,52446,8814,6880.401ok—

H-20260909T003403--hermes-qwen

hermes · Qwen3.8-27B-Instruct · ctx 65.536 · prune aus

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
16/68/2439421,518,313125,5332,9880.293ok—
26/611/2492803,602,422205,9152,5730.268harness-failure—
36/612/2421TimeoutError: timed outharness-failure—
42/610/2451974,418,149241,6052,4900.73harness-failure—
52/610/2425TimeoutError: timed outharness-failure—
64/610/241613412,45659,6214,5860.336harness-failure—
73/69/246483,05824,8986,2240.0ok—

H-20260909T003403--hermes-claude-sonnet-5

hermes · claude-sonnet-5

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
12/69/2414backend=remote: no local model, no vLLM countersok—
22/610/247backend=remote: no local model, no vLLM countersok—
32/610/2425backend=remote: no local model, no vLLM countersrc=-15—
42/614/2414backend=remote: no local model, no vLLM countersok—
52/616/2418backend=remote: no local model, no vLLM countersok—
62/618/2415backend=remote: no local model, no vLLM countersok—
72/620/2432backend=remote: no local model, no vLLM countersok—

H-20260909T003403--hermes-claude-opus-5

hermes · claude-opus-5

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
16/612/2429backend=remote: no local model, no vLLM countersok—
26/612/2412backend=remote: no local model, no vLLM countersok—
36/612/2415backend=remote: no local model, no vLLM countersrc=-15—
46/617/2416backend=remote: no local model, no vLLM countersok—
56/620/2413backend=remote: no local model, no vLLM countersok—
66/624/2415backend=remote: no local model, no vLLM countersok—
76/624/2451backend=remote: no local model, no vLLM countersok—

H-20260909T003403--hermes-claude-haiku-4-5

hermes · claude-haiku-4-5-20251001

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
12/67/2410backend=remote: no local model, no vLLM countersok—
26/611/2412backend=remote: no local model, no vLLM countersok—
36/611/2414backend=remote: no local model, no vLLM countersok—
46/613/2414backend=remote: no local model, no vLLM countersok—
56/612/2429backend=remote: no local model, no vLLM countersok—
64/613/244backend=remote: no local model, no vLLM countersok—
71/60/244backend=remote: no local model, no vLLM countersok—

H-20260909T003403--hermes-claude-fable-5-1

hermes · claude-fable-5-1

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
12/69/2410backend=remote: no local model, no vLLM countersok—
26/611/2418backend=remote: no local model, no vLLM countersok—
36/612/2412backend=remote: no local model, no vLLM countersok—
46/616/2413backend=remote: no local model, no vLLM countersok—
56/616/2417backend=remote: no local model, no vLLM countersok—
66/620/2419backend=remote: no local model, no vLLM countersok—
76/620/2415backend=remote: no local model, no vLLM countersok—

H-20260909T003403--cc-sonnet-5

claude-code · claude-sonnet-5

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
12/68/2410backend=remote: no local model, no vLLM countersok—
22/69/247backend=remote: no local model, no vLLM countersok—
32/610/2481backend=remote: no local model, no vLLM countersok—
42/612/2423backend=remote: no local model, no vLLM countersok—
52/616/2462backend=remote: no local model, no vLLM countersok—
62/619/247backend=remote: no local model, no vLLM countersok—
72/619/248backend=remote: no local model, no vLLM countersok—

H-20260909T003403--cc-opus-5

claude-code · claude-opus-5

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
16/610/2435backend=remote: no local model, no vLLM countersok—
26/610/2412backend=remote: no local model, no vLLM countersok—
36/611/2410backend=remote: no local model, no vLLM countersok—
46/615/2413backend=remote: no local model, no vLLM countersok—
56/619/2420backend=remote: no local model, no vLLM countersok—
66/623/2435backend=remote: no local model, no vLLM countersok—
76/622/2427backend=remote: no local model, no vLLM countersok—

H-20260909T003403--cc-fable-5-1

claude-code · claude-fable-5-1

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
16/610/2424backend=remote: no local model, no vLLM countersok9fc80fa0
26/611/2414backend=remote: no local model, no vLLM countersok3625c546
36/612/2425backend=remote: no local model, no vLLM countersok1f326f22
46/616/2414backend=remote: no local model, no vLLM countersokcaf97952
56/618/2428backend=remote: no local model, no vLLM countersokd1d5cc94
66/623/2424backend=remote: no local model, no vLLM countersokddbe2634
76/623/2427backend=remote: no local model, no vLLM countersok8bd1e7f9

G-20260908T220038

hermes · Qwen3.8-27B-Instruct

Rgatecapmin*reqprompt-tokgen-tokø gen/reqcache*exitBuild
12/69/2431692,654,095111,2711,6120.891ok—
20/60/24801275,691,292377,8162,9740.776rc=8—

Traces

ein Span je Runde, Projekt antifa-survivors-<zeile>

Span-Attribute je Runde: arena_run_id, arena_config, arena_harness, arena_model, arena_round, arena_exit_code, arena_wall_s, arena_ladder_rung, arena_capability_passed sowie ein arena_<feld> je Modell-Metrik.

* Wallzeit und Prefix-Cache tragen Einschränkungen: die eine misst die gleichzeitige Last mit (2 bis 15 parallele Runden), die andere die Serverwärme und das Alter der Session. Beide stehen hier als Kontext, nicht als Ergebnis.