ModelsKimi K2 Thinking Guide: Long Tool Chains and Native INT4
Kimi K2 Thinking added long-horizon reasoning to K2 and claimed 200 to 300 sequential tool calls. What it does well, what its INT4 weights change, and what has replaced it.
Most agents do not fail on the hard step. They fail on step forty, when the model has lost track of what it was doing and starts repeating itself or wandering off task. Kimi K2 Thinking was Moonshot's answer to that specific failure, and its headline claim was about endurance rather than raw intelligence.
It launched on 6 November 2025 alongside a faster turbo version. Since then Moonshot has shipped K2.5, K2.6 and K2.7 Code, so the question now is less whether K2 Thinking is good and more whether it still has a job.
What K2 Thinking adds to K2
The architecture is the familiar K2 shape: a mixture-of-experts model with 1T total parameters and 32B active, 384 experts with eight selected per token, and a 256K context window. What changes is the behaviour. Where Kimi K2 answers immediately, K2 Thinking reasons before it acts and interleaves that reasoning with tool calls.
The API returns the reasoning in a separate reasoning_content field alongside the answer. That matters for your client code: a parser that only reads the main content field will work, but will silently discard the trace you need when debugging a bad run.
Moonshot recommends temperature 1.0. Reusing a lower temperature tuned for a non-reasoning model is a common mistake when migrating prompts.
200 to 300 tool calls, and what that claim means
The model card claims K2 Thinking can sustain 200 to 300 consecutive tool calls, against earlier models it describes as degrading after 30 to 50 steps. That is the claim to care about if you run long agent loops.
It is also the claim to read carefully. A model being able to make three hundred tool calls coherently is not the same as your task needing three hundred, or your harness surviving that many. Each call adds tokens to the context, so a long chain inside 256K depends heavily on how much each tool returns. Context compaction strategies covers keeping a long loop inside the window.
The supporting benchmarks lean agentic. Moonshot reports 44.9 on Humanity's Last Exam with tools, 23.9 without them, 60.2 on BrowseComp and 71.3 on SWE-bench Verified. The gap between the with-tools and without-tools HLE figures is the interesting part: much of the score comes from searching and computing, not from recall.
Native INT4 and the self-hosting arithmetic
The less discussed feature is quantisation. Moonshot applied quantisation-aware training during post-training, with INT4 weight-only quantisation on the expert layers, and reports every benchmark under INT4. The model card claims roughly a two-times generation speedup.
That changes how you should read the numbers. With most open models the published score is for full precision, and the quantised build you can afford to run is somewhat worse. Here the published score is already for the quantised build, so there is no hidden discount when you self-host it.
The memory arithmetic follows directly. A trillion parameters at one byte each is about a terabyte; at four bits it is roughly half that, before the KV cache. That is still a multi-GPU cluster, but a smaller one. Quantization explained covers why training with quantisation in the loop loses less than quantising afterwards.
Paying for the thinking
On OpenRouter, K2 Thinking is listed at $0.60 input and $2.50 output per million tokens at the time of writing, with cached input at $0.15. The output price is where reasoning models get expensive, because the reasoning trace is billed as output.
On a long agent run that adds up quickly. A model that thinks for a few thousand tokens before every one of two hundred tool calls produces a large output bill even if each visible action is short. The token cost of reasoning models works through the arithmetic. Measure output tokens per completed task before you commit to it as a default.
Where it sits now
K2 Thinking was the first reasoning model in the K2 line, and it has been superseded on most fronts:
- K2.5 keeps thinking mode, adds native vision and an instant mode, and reports 76.8 on SWE-bench Verified. See the K2.5 guide.
- K2.6 is Moonshot's current general model in the same 1T/32B shape.
- K2.7 Code is the coding specialist, with thinking always on.
The case for staying on K2 Thinking is narrow: a text-only agent pipeline that already passes your regression suite, where a migration would cost more engineering time than it saves. It ships under the same Modified MIT licence as the rest of the K2 family, so licence terms are not a reason to move either way.
How to test it fairly
Long-horizon claims are hard to verify from short benchmarks, so test on the shape of work it was built for. Take ten tasks from your own backlog that need at least thirty tool calls, run them through K2 Thinking and K2.6 on the same harness, and record completion rate, number of steps and total output tokens. If K2 Thinking finishes the same tasks for fewer tokens, keep it. If not, move to the newer model. Evaluating agent reliability covers what to measure beyond a single pass rate.
Common questions
When was Kimi K2 Thinking released?
Moonshot launched Kimi K2 Thinking and a turbo version on 6 November 2025, according to its platform changelog. It uses the same 1T total, 32B active mixture-of-experts architecture as Kimi K2.
Why does Kimi K2 Thinking use INT4?
Moonshot used quantisation-aware training to ship native INT4 weights on the expert layers, claiming roughly double the generation speed. Its published benchmarks were measured under INT4, so self-hosted quality should match them.
Is Kimi K2 Thinking better than Kimi K2.5?
Mostly no. K2.5 keeps a thinking mode, adds native vision, and reports higher scores on Moonshot's coding benchmarks. K2 Thinking is worth keeping mainly where an existing text-only pipeline already works well.
