kimi-k2-0905-preview vs llama-3.3-70b-versatile

Capability for the money

Not the same tier: kimi-k2-0905-preview leads llama-3.3-70b-versatile by 46.8 points on tool use (τ²-bench), for 1.7× the price. They are level on following instructions (IFBench), so the premium only pays off on tool use work.

1. Following instructions

   extraction, classification, sticking to a format

   IFBench
   kimi-k2-0905-preview41.7%

   llama-3.3-70b-versatile47.1%

   llama-3.3-70b-versatile +5.4

2. Coding

   writing and fixing code

   LiveCodeBench
   kimi-k2-0905-preview61.0%

   llama-3.3-70b-versatile28.8%

   kimi-k2-0905-preview +32.2

3. Tool use

   multi-step agent work against real tools

   τ²-bench
   kimi-k2-0905-preview73.4%

   llama-3.3-70b-versatile26.6%

   kimi-k2-0905-preview +46.8

4. Hard reasoning

   graduate-level science and multi-step analysis

   GPQA Diamond
   kimi-k2-0905-preview76.7%

   llama-3.3-70b-versatile49.8%

   kimi-k2-0905-preview +26.9

Averaged over the benchmarks both models sat, llama-3.3-70b-versatile returns 59.5 points of pass rate per dollar of blended price against 58.8 — the ratio that decides whether the dearer model earns its rate.

Each figure is a pass rate: the share of that test set the model answered correctly, measured by a third party rather than self-reported by the lab. Rows are different tests of different difficulty, so compare a model against the other model, not one row against the next. And a benchmark is not your workload — it narrows the shortlist, it does not pick for you.