benchr
Moonshot released Kimi K3 on July 16, 2026 and later published the weights under the Kimi K3 License. benchr checked the release chronology, hosted API guide, price page, and official model card. The sources document one model but two very different procurement paths: buy tokens from Moonshot or operate the weights yourself.
Choose the control plane before comparing scores
“Open weight” is not a deployment plan. K3's first decision is who owns the inference stack, on-call burden, scaling, and license review. Pick a route with the evidence you can collect now, then compare model quality inside that route.
| Route | Good fit | Release gate | Reason to wait |
|---|---|---|---|
| Kimi hosted API | You need the model quickly, expect reusable prompt prefixes, and can keep data within the service policy. | Measure cache-hit share, output volume, tool retries, rate limits, and total billed cost on your own traces. | A temporary web-search limitation blocks a required workflow, or hosted governance is unacceptable. |
| Official open weights | You need infrastructure control and already have distributed-model operations, monitoring, and incident ownership. | Prove the chosen runtime, precision, memory plan, throughput, recovery, and license process with the exact artifact. | The business case depends only on “no token fee” or on an unverified hardware estimate. |
| Keep the current model | Your present system meets acceptance targets and K3 adds no task-specific advantage. | Document the missing capability and the cost of the current bottleneck before reopening migration. | None. Avoiding change is valid when the evidence does not pay for it. |
2.8T total is not the number on your invoice
The official card lists 2.8 trillion total parameters, 104 billion active parameters, 896 experts, 16 selected per token, and two shared experts. Sparse activation means K3 is not equivalent to running a dense 2.8T model on every token. It also does not mean you can size a production cluster from 104B alone.
A self-hosting study must account for stored weights, the chosen precision, expert placement, interconnect, context memory, replicas, batching, cold starts, monitoring, and spare capacity. Record accepted-output throughput instead of raw token speed alone. If those inputs are unknown, use the open-model reference to frame the license and deployment questions, but keep the hosted API as the cost baseline.
A million-token contract needs a smaller default
The API ID is kimi-k3. Moonshot documents a 1,048,576-token context window, a default maximum completion of 131,072, and an allowed maximum completion of 1,048,576. K3 always reasons. You can choose low, high, or max reasoning effort, with max as the default.
Those limits are capabilities, not recommended request sizes. Set a smaller workload-specific history and completion ceiling. Test what happens when a trace is compacted, when a tool result is large, and when the agent repeats a failed action. Preserve the complete assistant message in later tool-call turns, as Moonshot's guide requires, instead of retaining only the final text.
Cache behavior decides the hosted economics
Moonshot's global rates are $3 per million cache-miss input tokens, $0.30 per million cache-hit input tokens, and $15 per million output tokens. The rates are flat across context length. That tenfold input difference rewards a stable shared prefix, but it does not discount changing user content or generated output.
Separate cache misses, cache hits, reasoning output, final output, and retries in your logs. A single blended average hides the event that breaks the budget. Model the two input paths and an output-heavy case in the cost calculator before committing volume.
Tool support has one immediate blocker
K3 supports tool calls, tool-choice constraints, dynamic tool loading, JSON mode, strict structured output, partial mode, automatic context caching, and Batch. Moonshot also says its web-search tool is being updated and is not recommended for near-term production use. Treat that statement as a release blocker if provider-hosted search is mandatory.
You can still evaluate custom retrieval or another approved search layer, but score that whole system rather than crediting K3 for a tool it did not run. Require strict schemas on machine-consumed steps, return a result for every tool-call ID, and stop when the same call fails repeatedly.
Run one evaluation with two cost tracks
- Fix the cases. Use representative long-code, document, vision, structured-output, and tool-recovery tasks with written acceptance rules.
- Run the hosted route first. Pin
kimi-k3, reasoning effort, prompt, tools, and completion limit. Capture cache state, token use, retries, and reviewer acceptance. - Qualify self-hosting separately. Use the official weights and license, record the exact runtime and precision, then measure capacity, failures, and operations time.
- Compare complete cost. Hosted cost includes tokens and tool work. Self-hosted cost includes infrastructure, engineering, observability, idle capacity, and recovery.
- Canary only the winning route. Keep a fallback model and a hard stop for schema, tool, cost, or quality failures.
The Labs workspace provides a fixed-case and rubric structure, but it contains no K3 outputs or benchr score. If K3 replaces an earlier Moonshot model, compare the same cases against the existing Kimi K2.6 review rather than comparing provider headline numbers from different harnesses.
| Field | Provider record | Operational decision |
|---|---|---|
| Model and license | kimi-k3; Kimi K3 License; official weights published | Legal review plus hosted or self-hosted ownership |
| Architecture | 2.8T total / 104B active / 896 experts | Runtime qualification instead of hardware guesses |
| Context / allowed output | 1,048,576 / up to 1,048,576 tokens; 131,072 default completion | Smaller task limits, compaction, and failure stops |
| Hosted pricing | $3 miss / $0.30 hit input; $15 output per 1M | Cache-aware trace analysis and output budget |
| Hosted web search | Temporarily not recommended while Moonshot updates it | Block release or qualify another retrieval path |