11,667 tokens of shared state over 21 artifacts. Every artifact each agent
was declared to need still arrived, and the routed run scored 0.95 on graded answers
against 0.81 for the agent handed everything.
Agent workflowsToken cost
Research fan-out wide parallel gather
five parallel gatherers → synthesizer → critic → editor
Without Tileward364,306 · $0.364/run
With Tileward65,513 · $0.066/run
5.6×
Fewer tokens
$298.79
Saved per 1,000 runs
8
Agents · 4 rounds
45,712 tokens of shared state over 32 artifacts. Every required artifact
still arrived. Answers were not graded on this flow.
Conversation memoryToken costQuality
Long conversation 142 turns, asked back
Whole transcript resent3,122 per question
With Tileward103–127 per question
25–30×
Fewer tokens
20 of 20
Answered, either way
0 of 20
Answered with no history
12 facts mentioned once each, 20 questions, 8 needing two or three facts at
once. Qwen3.5-9B, exact substring match. With decomposition turned off it answers 16 of 20,
so the question-splitting is doing that work rather than the retrieval.
Model compressionQuality
Compressed model against the publisher’s own build
Publisher’s official FP835.7 GB · 82.60% MMLU
Tileward 35B-A3B24.5 GB · 82.58% MMLU
Community GPTQ Int423.3 GB · 83.30% MMLU
Community AWQ 4-bit24.3 GB · 83.17% MMLU
1.46×
Smaller on disk
0.02
Points apart on MMLU
±0.31
Error bar on each
Field
Tileward
GPTQ
AWQ
FP8
Humanities
77.77
78.36
79.02
78.47
Social sciences
89.76
90.06
89.70
89.37
STEM
80.40
81.70
80.49
79.58
Other
84.97
85.71
85.71
85.23
Level with FP8, behind both community 4-bit builds.
The FP8 gap is inside the error bar, which is what a substitute should look like. GPTQ leads by
0.72 and AWQ by 0.59, at 1.4 to 1.7 times the combined error, so neither is settled.
Five-shot, 14,042 questions, lm-evaluation-harness, no sampling. Same base model, one machine.
Bars are weights on disk. Serving VRAM is 25.8 GB against FP8's 40.4 GB.
MMLU is multiple choice: no signal on generation, instruction following or long context.
Matching FP8 is consistent with the builds behaving alike, not proof. Per-token agreement not run.
No evals carry that tag yet.
Agent
flows measured on gpt-oss-120b, tokens counted by the serving model rather than estimated.
Costs are that model’s published rate of $1.00 per million tokens, on uncached serving:
a provider that discounts a repeated prefix narrows the gap, because the run without Tileward
sends every agent the same block.
Bring your own workload.
Point the harness at something you actually run.
Talk to an engineer.