Qwen3 4B

model page · concurrency sweep 1–24 agents

Qwen3 4B

server model id: qwen3-4b
thinkingKV Q4_0ctx 32k4B params

The worst latency profile of any model tested — 11.8s TTFT solo, 40–56s under load — and ~95% of its output is hidden reasoning at high concurrency. Peak: 340 tok/s.

Key findings

  • The worst latency profile of any model tested: TTFT is 11.8s solo and 40–56s under load, with max values up to 60.0s — at high concurrency agents often wait the *entire* 60s window before producing a single token. The 4b params + Q4_0 KV + thinking prefill makes every request start extremely slowly.
  • Very slow absolute speed: only 60 tok/s solo (47 visible), and combined all-token throughput peaks at just 340 tok/s at 24 agents (5.7x) — near the bottom of the field.
  • Reasoning dominates almost everything at high concurrency: visible content collapses to 8–17 tok/s at 18–24 agents (~95% of output is hidden reasoning_content).
  • Errors: 0 across all 24 runs — but as a latency-sensitive or content-heavy workload, this model is a poor fit.

Sweep — total concurrency 1–24

agentswall (s)contentreasontotalcomb tok/scomb allper-agentper-agent allscale %TTFT meanTTFT maxoktimeouterr
160.02,8337733,606476058.860.3100%11.8s11.8s010
260.02,9702,9615,931499944.049.6165%27.0s41.0s020
360.06,0163,0779,09310015150.954.4252%20.6s23.6s030
460.05,9414,56310,5049917545.249.2292%27.4s41.3s040
560.05,9805,63011,61010019340.945.0322%30.9s38.5s050
660.06,1366,58412,72010221231.541.4353%27.5s29.8s060
760.06,0427,55213,59410122633.938.3377%35.0s53.5s070
860.04,7808,89213,6728022829.834.6380%40.1s54.6s080
960.03,97611,04715,0236625025.934.1417%43.2s56.6s090
1060.03,96711,27115,2386625425.432.9423%44.6s55.1s0100
1160.04,10612,21016,3166827225.031.6453%45.1s59.5s0110
1260.04,09213,22217,3146828817.831.3480%41.1s54.3s0120
1360.03,83913,84617,6856429521.530.4492%46.4s58.4s0130
1460.04,72413,27117,9957930021.829.4500%44.7s55.6s0140
1560.03,38214,96718,3495630616.228.4510%46.1s53.8s0150
1660.03,26916,19419,4635432418.527.4540%49.0s59.6s0160
1760.02,56315,52318,0864330110.724.6502%46.1s55.6s0170
1860.01,37317,30918,682233119.623.6518%52.0s59.1s0180
1960.02,28616,81719,1033831811.722.9530%49.8s58.8s0190
2060.01,70017,37619,0762831810.022.6530%51.6s59.1s0200
2160.01,37117,36318,734233129.422.1520%52.4s59.9s0210
2260.01,04918,46919,518173257.921.4542%54.0s59.0s0220
2360.050918,86919,37883234.421.1538%55.7s60.0s0230
24 peak60.099319,39620,389173405.020.6567%51.6s57.2s0240

* burst artifact — agents queued ~11–16s then generated in a short burst; the per-agent timer inflates the number. Trust combined throughput and TTFT columns. Peak row = highest combined (all-token) throughput.

Charts

Whole sweep at a glance — combined throughput, scaling efficiency, per-agent throughput and outcomes across every concurrency level.
Whole sweep at a glance — combined throughput, scaling efficiency, per-agent throughput and outcomes across every concurrency level.
Combined throughput (all tokens) vs concurrent agents. Annotated peak with the concurrency level where it occurred.
Combined throughput (all tokens) vs concurrent agents. Annotated peak with the concurrency level where it occurred.
Per-agent throughput erosion as slots contend. Red stars mark burst-timing artifacts (queued ≥ 10s then generated in a burst — inflated values, not real throughput).
Per-agent throughput erosion as slots contend. Red stars mark burst-timing artifacts (queued ≥ 10s then generated in a burst — inflated values, not real throughput).
Scaling efficiency relative to the single-agent baseline (100%).
Scaling efficiency relative to the single-agent baseline (100%).
Time to first token — mean (solid) and max (dashed). Log scale. Gaps mean no visible content ever arrived within the 60s window.
Time to first token — mean (solid) and max (dashed). Log scale. Gaps mean no visible content ever arrived within the 60s window.
Total tokens per 60s run — all tokens vs visible content only.
Total tokens per 60s run — all tokens vs visible content only.
How each run of N agents finished: completed, timed out (60s budget), or errored.
How each run of N agents finished: completed, timed out (60s budget), or errored.
The concurrency trade-off on one chart — combined throughput rises while per-agent speed falls.
The concurrency trade-off on one chart — combined throughput rises while per-agent speed falls.
Visible content vs hidden reasoning tokens per run, plus the reasoning share (%).
Visible content vs hidden reasoning tokens per run, plus the reasoning share (%).

Raw data & downloads