Kimi K3 is not important simply because Moonshot AI attached the largest number yet to an open-weight model. It matters because the company has combined 2.8 trillion total parameters, unusually sparse expert activation, native vision, a one-million-token context window and an aggressive API price into a model that independent evaluators place near the proprietary frontier.
The careful version is more interesting than the hype. Kimi K3 is available through Kimi's web products and API now, but the full model weights are not available at the time of publication. Moonshot says they will be released by 27 July 2026. Its launch benchmarks are substantial, yet several use different agent harnesses, and the detailed technical report is still pending. Independent testing has produced impressive knowledge-work results while also exposing high task cost and long completion time.
That leaves a clear conclusion: K3 is already a credible model to test by API, and it may become a strategically important open-weight release. It is not yet a model that most organisations can independently reproduce, audit or deploy from downloaded weights.
Kimi K3: the verified facts, the independent evidence and the open questions
This matrix separates what Moonshot has published, what an independent evaluator has measured and what remains unresolved.
| Question | What is verified now | Evidence status | Practical consequence |
|---|---|---|---|
| How large is Kimi K3? | Moonshot states 2.8 trillion total parameters. | Official specification; technical report pending. | Parameter count signals system scale, not automatic quality. |
| How many experts are active? | 16 of 896 experts are activated in its sparse MoE design. | Official architecture description. | Each token uses a small part of the total expert pool, reducing active compute relative to a dense 2.8T model. |
| What context does it support? | 1,048,576 tokens through the Kimi API. | Official product and pricing documentation. | Large repositories and document sets can fit, but long-context accuracy and cost still need testing. |
| Is it multimodal? | It accepts native visual input and is designed for vision-in-the-loop work. | Official capability statement and demonstrations. | It can inspect screenshots while iterating on code, interfaces and 3D work. |
| Is it open weight today? | No full weight package was publicly available at publication time. Moonshot promises release by 27 July. | Official dated commitment. | API access and future self-hosting must be treated as different products for now. |
| How strong is it independently? | Artificial Analysis gives K3 an Intelligence Index score of 57 and an AA-Briefcase Elo of 1,543. | Independent evaluation, published 21 July. | K3 is in the frontier conversation, especially for agentic knowledge work. |
| What does it cost by API? | $3 per million uncached input tokens, $0.30 cached input and $15 output. | Official Kimi pricing. | It undercuts GPT-5.6 Sol list price, but agent behaviour can dominate total task cost. |
| Can it run on a workstation? | Moonshot recommends supernodes with 64 or more accelerators. | Official deployment recommendation. | This is data-centre infrastructure, not a normal local-LLM download. |
| Is every benchmark directly comparable? | No. Moonshot discloses different harnesses for some competing models. | Official benchmark footnotes. | Buyers should run the same workload, harness and acceptance criteria across candidates. |
What 2.8 trillion parameters actually means
The headline number describes the total parameter pool, not the amount of computation applied to every token. Kimi K3 is a Mixture-of-Experts model. Moonshot says its Stable LatentMoE design routes a token through 16 of 896 experts. That is roughly 1.8% of the expert pool.
Sparse activation is the reason a multi-trillion-parameter model can be served at all. It lets the system store a very large collection of specialised weights while activating only a subset for a particular token. But sparse does not mean small. The entire weight set still has to be stored and distributed, routing has to remain balanced, and expert traffic must move quickly between accelerators.
Moonshot describes several mechanisms intended to make that scale stable. Kimi Delta Attention, or KDA, is the model's attention foundation. Attention Residuals allow the network to retrieve representations across depth rather than simply accumulate them in one path. Quantile Balancing allocates experts from router-score quantiles, while Per-Head Muon applies optimisation independently to attention heads. The company also describes MXFP4 weights and MXFP8 activations through quantisation-aware training.
These are meaningful engineering claims, but they should not be converted into a universal efficiency verdict before the technical report and weights arrive. The most defensible statement is that Moonshot has designed K3 around sparse scale from training through serving. The design is more consequential than the parameter count alone.
Why the one-million-token context matters—and why it can mislead
A one-million-token window is large enough to ingest substantial repositories, research collections or transaction histories in a single request. Kimi positions K3 for long-horizon coding and end-to-end knowledge work, where losing context between turns is expensive.
Capacity, however, is not the same as dependable recall. A model may accept one million tokens without using every distant fact equally well. Teams should measure retrieval accuracy at multiple positions, instruction retention, tool-state continuity and the cost of repeatedly sending or caching that context. Kimi's 90% cached-input discount makes repeated context cheaper, but it does not remove the operational burden of a very large prompt.
The right question is not “does it have one million tokens?” It is “at what context length does this particular workflow stop improving?” A 70,000-token repository slice that has been deliberately selected may produce a better result than an undifferentiated million-token dump.
What the benchmark evidence says
Moonshot's launch page presents K3 as a strong coding and agent model. It reports competitive kernel optimisation, a GPU compiler built from scratch, visual game-development work and a 48-hour chip-design experiment. The company also says K3 performed strongly on DeepSWE, Terminal-Bench 2.1, Program Bench, FrontierSWE and several internal knowledge-work evaluations.
The footnotes are essential. K3 is sometimes evaluated with Kimi Code, while GPT-5.6 Sol may use Codex and Anthropic models may use Claude Code or another harness. A harness can affect planning, tool use, context management and retry behaviour. Results from different harnesses can still be informative, but they do not isolate the base model.
Artificial Analysis's independent AA-Briefcase evaluation adds a valuable second view. It placed K3 second only to Claude Fable 5 on its agentic knowledge-work benchmark, with an Elo of 1,543. K3 achieved a 51% rubric pass rate and an analytical-quality Elo of 1,754. Its presentation-quality Elo, however, trailed GPT-5.6 Sol.
The efficiency numbers are sobering. Artificial Analysis measured an average cost of $10.57 and 56.4 minutes per AA-Briefcase task. K3 averaged 83 turns and 120,000 output tokens. A low per-token price did not make the completed task cheap because the model used many turns and generated a great deal of output.
That is the most useful buyer lesson in the launch. Compare cost per accepted result, not price per million tokens.
Is Kimi K3 really open?
“Open” currently describes a commitment and a direction, not a complete package available for inspection today. Moonshot's official Kimi K3 launch post says full model weights will be released by 27 July, alongside further technical detail. Until that happens, the public cannot verify the shipped files, licence, quantisation options, exact runtime requirements or whether independent deployments reproduce API behaviour.
Open weight and open source are also not synonyms. A weight licence may permit broad use while withholding training data, training code or parts of the data recipe. Alternatively, it may impose use restrictions that matter to commercial operators. The licence text—not the launch adjective—will decide what organisations may legally do.
This does not make the release meaningless. API access is real, the architecture description is detailed, and the dated weight commitment is unusually specific. It does mean that “download Kimi K3 now” is the wrong description before 27 July.
Availability, price and the demand shock
K3 is available through Kimi.com, Kimi Work, Kimi Code and the Kimi API. The company lists API rates of $3 per million uncached input tokens, $0.30 for cached input and $15 for output. K3 uses maximum thinking effort by default at launch, with lower effort modes planned.
Demand has been more than theoretical. The Associated Press reported that Moonshot temporarily halted new subscriptions after demand exceeded capacity. That is evidence of attention and infrastructure pressure, not proof of long-term market share. It also reinforces why buyers should test latency and availability rather than assume the published token rate guarantees predictable throughput.
Compared with earlier Kimi releases, the jump is substantial. Our guide to Moonshot Kimi K2.7 Code covers a one-trillion-parameter coding model with a smaller context and different cost profile. The earlier Kimi K2.5 model guide shows how quickly Moonshot has expanded from multimodal agents to a much larger long-horizon system.
What Kimi K3 changes for model buyers
First, proprietary frontier models can no longer assume that open-weight challengers will compete only on price. K3's independent knowledge-work result puts capability, not just accessibility, into the decision.
Second, API economics and ownership economics are diverging. The API is competitively priced, while self-hosting the full model is likely to require an expensive multi-node system. “Open” may improve control and portability without making infrastructure inexpensive.
Third, evaluation has to become workload-specific. K3 can look excellent on a long, agentic task and still be the wrong choice for a low-latency support workflow. Gemini 3.6 Flash may finish routine agent loops more quickly; GPT-5.6 Sol may offer stronger presentation or tool integration; a smaller open model may be dramatically cheaper to own.
Fourth, the release raises the value of caching and orchestration. With long contexts and high output use, prompt reuse, tool policies, stopping criteria and verifier design can affect the bill more than the nominal input rate.
A practical Kimi K3 test plan
Before moving production work, use a controlled sequence:
Select 20 to 50 real tasks with known acceptance criteria.
Fix the tools, system prompt, context and retry budget across models.
Record completed-task rate, human correction time, p50 and p95 latency, input, cached input and output tokens.
Separate API failures from model-quality failures.
Test long context at several lengths rather than starting at one million tokens.
Review sensitive-data terms and regional availability before sending production material.
Re-run the deployment assessment after the weights, licence and technical report arrive.
Do not begin with the hardest possible demo. Begin with the workload that costs the organisation the most time today.
The bottom line
Kimi K3 is a credible frontier-model event because three things coincide: unprecedented open-weight scale, independent evidence of strong agentic knowledge work and an API price that invites comparison with closed leaders. Its limitations are equally concrete: full weights are still pending, self-hosting points to 64-plus accelerators, benchmark harnesses differ and independent tasks can be slow and output-heavy.
The right response is neither to dismiss the model as a parameter-count stunt nor to declare the proprietary frontier finished. Test K3 now where API access is appropriate. Treat 27 July as a separate verification point for the weights, licence and deployment reality. That is when the most consequential part of the “open” claim can actually be examined.
