Kimi K3 (FW)
Last updated: September 15, 2026
Kimi K3 (FW) completed 37 of 55 private-markets finance tasks (67.3%) at $0.31 per task and 118 seconds each, ranking #8 of 17 models tested — strongest on metric lookups, weakest on error detection & repair.
Tested on 55 tasks across seven kinds of financial work — reading documents, looking up figures, building statements, finding and fixing errors, assembling dashboards, drawing charts, and multi-step requests. It passed 12 of 12 on metric lookups and 3 of 7 on error detection & repair. Run-to-run noise is about ±1.8 tasks, so gaps of two tasks or fewer are ties.
Overall Rank
#8
of 17 engines tested
Pass Rate
67.3%
37 of 55 tasks passed
Judge Score
39.6
out of 55.0 available
Suite Split
33/49
analysis suite · 4/6 ingest harness
Cost per Task
$0.31
6th cheapest of 17
Cost, Full Pass
$17
all 55 tasks at list prices
Cost per Passed Task
$0.46
4th best value of 17
Latency
118s
per task · 1h 48m for a full pass
Category Breakdown
Kimi K3 (FW) against the average of every engine tested, category by category. The gap between the two bars is the only thing worth reading here — a category where the whole field struggles is a property of the tasks, not of the engine.
| Category | Passed | Pass Rate | Field Avg | Rank | Score | Cost/Task | Latency | Tool Calls |
|---|---|---|---|---|---|---|---|---|
| Document Ingestion | 4/6 | 66.7% | 42.2% | #2 of 17 | 4.70 | $0.93 | 272s | 48.5 |
| Metric Lookups | 12/12 | 100.0% | 88.7% | #1 of 17 | 12.00 | $0.06 | 57s | 8.1 |
| Financial Statements | 3/7 | 42.9% | 43.7% | #8 of 17 | 3.61 | $0.14 | 121s | 19.6 |
| Error Detection & Repair | 3/7 | 42.9% | 60.5% | #14 of 17 | 3.05 | $0.31 | 142s | 34.3 |
| Dashboards & Widgets | 7/8 | 87.5% | 84.6% | #8 of 17 | 6.93 | $0.43 | 131s | 32.9 |
| Charts | 2/4 | 50.0% | 66.2% | #13 of 17 | 2.05 | $0.37 | 117s | 32.0 |
| Multi-Step Requests | 3/6 | 50.0% | 54.9% | #9 of 17 | 4.07 | $0.32 | 90s | 13.8 |
Every column is measured within that category. The ingest harness is a separate suite of longer, heavier workflows — several times the tool calls of an analysis task — so its cost and latency are not comparable with the six rows below it.
Where it fell short
The 16 tasks Kimi K3 (FW) did not pass, of 50 graded individually. A benchmark that publishes only totals has to be taken on trust; these are the specific misses behind this model's score.
| Task | Kind of work | What it tests | Result |
|---|---|---|---|
| TBU-01 | Document Ingestion | 556-account trial balance, signed, zero rows skipped | Failed |
| RCN-03 | Document Ingestion | 87 stores × 9 metrics, dimensioned, totals excluded | Failed · 0.70 |
| CTA-03 | Financial Statements | — | Failed · 0.17 |
| CTA-04 | Financial Statements | — | Failed · 0.34 |
| CTA-05 | Financial Statements | — | Failed · 0.10 |
| SBM-01 | Financial Statements | — | Failed |
| BUG-02 | Error Detection & Repair | — | Failed |
| DISC-01 | Error Detection & Repair | — | Failed |
| DISC-02 | Error Detection & Repair | — | Failed · 0.05 |
| DISC-03 | Error Detection & Repair | — | Failed |
| BOARD-01 | Dashboards & Widgets | — | Failed |
| CHART-02 | Charts | — | Failed · 0.05 |
| CHART-04 | Charts | — | Failed |
| CMP-01 | Multi-Step Requests | — | Failed · 0.38 |
| CMP-03 | Multi-Step Requests | — | Failed · 0.45 |
| MT-03 | Multi-Step Requests | — | Failed · 0.25 |
A fraction means the task was run more than once and the runs disagreed. A number after a failure is the partial credit awarded. Five of the 55 tasks in the headline total are not broken out individually in the published results.
Against the Field
Every engine on the same axis, with Kimi K3 (FW) picked out. Grey bars are the rest of the field.
Overall Pass Rate
#8 of 17 · higher is better
Cost — One Full Pass (55 tasks)
#6 of 17 · lower is better
Cost per Passed Task
#4 of 17 · lower is better
Latency — Average per Task
#12 of 17 · lower is better
Tool Calls — Average per Task
#7 of 17 · lower is leaner
Cost vs Pass Rate
Up and to the left is better
Henon Tasks vs Tools benchmark — Ingest v5.1 + analysis suite v4, 17 executor models, 935 graded runs. Last updated September 15, 2026. Full results at henon.ai/benchmarks