Are Chinese AI models really viable alternatives for investors?
•By Dave Wang
The newest Chinese open source model Kimi K3 just dropped last weekend.
Within hours, it ranked first on benchmarks, ahead of Claude Fable 5.
X did what X does. The frontier-model race was apparently over.
I research AI academically and deploy it in real investment workflows.
Benchmarks matter.... But statisticians have an old joke: torture the data long enough and it will confess to anything :)
For any new model release, the real test is if it passes the "smell test"
So I ran my own buy-side benchmark test.
The 60 second answer
Where Kimi looks strongest:
Cost advantage (~80% cost savings)
Long-document and large-codebase review
High-volume first-pass research
Python prototypes and research scaffolding
Charts, dashboards, and polished research artifacts
Structured extraction and mechanical workbook cleanup
Acting as a second opinion on another model's work
Where I would not trust Kimi alone:
Investment-grade financial modeling
Point-in-time financial-data alignment
Statistical and causal claims
Production backtests
Accounting, valuation, or covenant signoff
Portfolio construction and trade decisions
Sensitive fund data without a security review
The attractive Kimi tasks share three traits: the input is large or messy, the deliverable is easy to inspect, and mistakes are relatively cheap to catch.
The dangerous tasks hinge on subtle assumptions or conclusions aka the decision / inference work.
My Buy-Side Benchmark Test
Most model benchmarks can be rigged, and even more importantly the tests are coding optimized rather than testing for investment tasks.
One of my favorite real-world tests is to give a model a loosely defined market question and see whether it can turn that ambiguity into defensible research.
For this test, I asked Kimi K3 and Claude Fable the same question: does crypto lead or lag changes in equity-market risk appetite?
I gave them only rough guidance and asked each model to design the methodology, source the data, write the code, run the backtests, and produce an institutional research report.
This is a much harder test than answering a finance question.
I like this finance benchmark as it combines research design, data collection, quantitative analysis, judgment, and communication.
More importantly, it forces the model to decide what evidence should count.
Kimi's first result was impressive at the surface.
It produced a polished 46-page report with a broad methodology, functioning code, charts, and a backtest.
If I had judged the models by their finished PDFs, Kimi would have looked surprisingly close to Fable.
But (!)
Polished research can still be wrong.
I spent more than an hour working through both reports in Codex, independently rerunning the backtests and separating the problems into hard errors and soft errors.
What Kimi did well:
Turned a vague investment question into a complete research project with very little direction
Generated sensible hypotheses, code, charts, and a useful starting framework
Produced something polished enough that an analyst could continue working from it
Used ~50% fewer tokens than Fable on this assignment
The hard errors uncovered during review:
A Bitcoin date-alignment problem in part of the dataset materially influenced the headline conclusion
The weekend signal used information that would not have been fully available at the assumed Monday trading point, introducing look-ahead bias
The apparent backtest improvement weakened once the timing and testing assumptions were corrected
The softer, more dangerous problems:
Kimi treated fragile evidence as if it were a durable market relationship
It did not challenge its own methodology aggressively enough when the result looked economically interesting
The confidence and polish of the report were stronger than the underlying evidence justified
It moved too quickly from statistical association to something that sounded potentially tradable
The verdict: Kimi K3 is "worse" than Fable from my buyside benchmark
Kimi can do an impressive amount of research production, but the hardest part of investment work is not generating more analysis.
It is knowing when the analysis should not be trusted.
Fable was materially stronger on data discipline, robustness, and inference.
Honestly, I would still use Claude Code or Codex for most serious investment work.
Investing is less of a box-optimization problem and more of a player-versus-player problem.
You are not merely trying to complete a well-defined task.
You are competing against other smart investors who have access to the same filings, data, models, and increasingly the same AI tools.
That final increment of reasoning quality matters.
Kimi K3 shines on cost savings primarily
But ... If cost is becoming a bottleneck, or you specifically want to incorporate lower-cost Chinese or open-weight models, these are the jobs I would consider delegating to Kimi:
Reasonable Kimi-first tasks:
First-pass synthesis across large document collections
Extracting, tagging, and organizing information for later review
Python scaffolding, utility scripts
Visualization
Formatting memos, presentations
High-volume work (e.g., scraping / downloading)
Tasks where Kimi can help, but needs an independent challenger:
Idea and variant-perception generation
Portfolio scenario exploration
Filing and transcript comparisons
The Bottom Line
IMO the Chinese models are clearly closing the gap fast but they're still not ready yet to replace top tier models for investing work.
Whenever you see buzz around new model benchmarks on X, just know none of that counts until the models pass the smell test.
One other thing I've noticed: A lot of tech influencers / commentators throw around casual hyperboles but I know they've barely used the model .... This benchmark test I ran here took me several hours!
Personal
One of the biggest lessons I have learned from AI is that better answers usually start with better data.
So I recently bought a Google Fitbit to begin building a useful dataset on myself: sleep, activity, heart rate, recovery, and stress signals.
Over time, I want to use AI to find correlations I might otherwise miss.
Does certain work consistently disrupt my sleep?
Do particular meetings leave my stress elevated?
Are there responsibilities I should delegate more often?
Are some people more draining than I realize?
So far it's been solid - apparently I've been drinking too much water before bed which causes me to get poor sleep?
Anyways .... a little experiment I've been running on myself.
