Kimi K2 Thinking

Reason, act, observe, revise

In a reasoning loop with tools, the model can propose a calculation, ask a program to run it, read the result and decide what to do next. The result becomes part of the conversation. That differs from writing out every step without ever checking anything against the outside world.

Moonshot’s announcement presents K2 Thinking as an agent that interleaves reasoning with tool calls, and reports tests with 200 to 300 tool calls in sequence. Those results come from specific task setups; they do not promise that any task will run successfully for that many steps. [1]

Suppose a proposed solution to an equation fails when substituted back in. A tool reveals the mismatch, and the useful next step is to recheck the derivation or try another candidate. Writing a longer explanation of the failed candidate adds cost without adding evidence.

Four-bit weights for long generation

Every generated token requires reading the weights of the experts it uses, and in long generation that reading takes much of the time. Storing weights in fewer bits reduces both the memory they occupy and the data that must be moved. K2 Thinking stores the weights of its expert layers as 4-bit integers (INT4). This is weight-only quantization: activations and the attention cache keep their own formats. The model was trained with quantization-aware training during post-training, and the model card says its benchmark results were measured with the INT4 checkpoint. That evidence covers this trained recipe; it does not show that converting some other model to four bits after training would work as well. [2]

Quantization-aware training lets the model experience the rounding while it is still learning. Think of practicing measurements with the ruler you will actually use. If your method depends on millimeter marks and the real ruler shows only centimeters, the method fails; practicing with the coarse ruler lets you adapt. The analogy says nothing about the exact rounding rule or how training passes gradients through it.

In the figure, 20 billion weights take 40 GB at 16 bits and 10 GB at 4 bits. That ratio is exact for this made-up uniform array. It does not size the K2 Thinking checkpoint, which quantizes only some weights and also stores scale factors.

One early error changes everything after it

Each generated token becomes input for the next, so a small change in one token can redirect everything that follows: a different search query, a different file opened, a wrong assumption accepted. That makes a long reasoning model harder to evaluate than a model judged on a few isolated outputs.

Take two versions of a model, one in full precision and one in four bits. Both answer nine easy questions correctly. On a tenth, one writes “plus” where the other writes “minus” in an intermediate step. A test catches the difference only if it includes that harder problem and runs it enough times to reveal unstable behavior. A short test set can miss this kind of sensitivity.

A fair precision comparison therefore keeps the prompt, tools, context and generation budget the same, checks complete outcomes, and repeats each run several times, because generation involves random sampling. The model card’s speed claims apply to its own low-latency setup; the storage arithmetic above cannot produce them. [2]

Measuring the whole task

A task’s total time includes generating the reasoning, waiting for tools, and recovering from mistakes. A quick first answer can turn out expensive if it leads to repeated failed actions. A long initial plan can be wasteful when a single lookup would settle the question. The budget that matters is the completed task, not the first token.

K2 Thinking keeps the original K2’s sparse expert architecture and its MLA attention. What it adds is training for long reasoning and for low-bit inference, not a new attention design. [2]

To inspect such a system, record which claims were checked, which tool calls changed the plan and when the final answer became justified. A long transcript is valuable when its extra steps reduce uncertainty. Its length alone does not show whether the model found better evidence or simply spent more computation.

Simplifications and limits

• The storage figure uses a made-up array of 20 billion weights; it is not a sizing calculator for Kimi deployments.
• Moonshot’s reported speeds and tool-call counts do not establish a general speed-up or guarantee reliability.

Follow the idea

Kimi K2Kimi K2.5

Sources and further reading

Primary papers, model cards, and release posts. Facts describe the versions named here; the interactive calculations are teaching models. Research cutoff: October 3, 2026.

1. Kimi K2 Thinking announcementInterleaved reasoning and tool use; reported long-horizon experiments.
2. Kimi-K2-Thinking model cardArchitecture, native INT4 MoE weight-only QAT, evaluation precision and deployment qualifications.
3. Moonshot research chronologyOfficial November 6, 2025 announcement date.