AI Blog

Kimi K3 Is Open Weights. That Is Not the Same as Cheap, Local, or Unrestricted

Moonshot released 2.8 trillion parameters as a free download on 27 July — and priced its own API above the model it replaced, while no single GPU on the market can hold the weights. Open weights buy agent builders exactly one thing that closed APIs cannot, and it is not cost.

By Agentic AI Wiki 14 min read

Moonshot put 2.8 trillion parameters on the internet as a free download and, in the same fortnight, priced its own hosted API at roughly three times what the previous generation cost. Both facts are true, neither is a contradiction, and holding them together is the only way to work out whether an open-weight frontier model belongs anywhere in your agent stack.

At a glance

Kimi K3 reached the API on 16 July 2026; the weights went up at midnight UTC on 27 July. The specification is unusual in ways that matter more for agents than for chat.

PropertyKimi K3What it implies for an agent
Parameters2.8T total, sparse MoELargest open-weight release to date. All of it has to be resident in memory.
Routing16 of 896 experts per tokenUnder 2% fires per token, so compute per token stays modest — but see the memory line above.
Context1,000,000 tokensLong agent transcripts stop needing aggressive truncation. The KV cache bill arrives instead.
AttentionKimi Delta Attention, hybrid linearMoonshot claims ~6.3× decode speedup at million-token contexts — the constraint that usually kills long-horizon loops.
WeightsMXFP4, ≈1.4 TB downloadQuantisation-aware from fine-tuning onward, so this is the intended precision, not a lossy afterthought.
Hosted price$3 / $0.30 / $15 per M tokensCache-miss input, cache-hit input, output. Output dominates agent bills.
Kimi K3 hosted API pricing against the model it replaces Horizontal bar chart of published list prices in US dollars per million tokens. Kimi K3 charges fifteen dollars for output against roughly four for K2.6, and three dollars for cache-miss input against roughly ninety-five cents. Only the cache-hit input price, thirty cents, is cheaper than anything in the previous generation. USD PER MILLION TOKENS — PUBLISHED LIST PRICE $3.75 $7.50 $11.25 $15.00 Kimi K3 — output what agent loops actually burn $15.00 Kimi K2.6 — output the previous generation $4.00 Kimi K3 — input cache miss $3.00 Kimi K2.6 — input the previous generation $0.95 Kimi K3 — input cache hit, 90% off $0.30
Published list prices. The open-weight release did not make the hosted model cheaper — it made it dearer than the model it replaces.

On capability the picture is consistent across independent scorers. Artificial Analysis placed K3 fourth of 189 on its Intelligence Index v4.1 with a score of 57, against 44 for K2.6. Arena ranked it first on frontend code at 1,679 in blind developer testing. The Vals Index put it second at 74.70%. On AA-Briefcase, a private long-horizon agentic benchmark, it scored 1,527 — ahead of GPT-5.6 Sol Max at 1,495 and behind Fable 5 Max at 1,587. Moonshot's own framing is unusually candid: K3 still trails Claude Fable 5 and GPT-5.6 Sol overall, while beating everything else the company tested it against.

That is the headline, and it is genuinely a first — near-frontier agentic performance you can download. The rest of this post is about the three separate things people hear in the phrase "open weights", only one of which K3 actually delivers.

What actually shipped

Kimi K3 architecture and its memory footprint Tokens from a context window of up to one million pass through Kimi Delta Attention, a hybrid linear attention used in selected layers, then a router that fires sixteen of eight hundred and ninety-six experts per token — under two percent of the pool. Sparse activation keeps compute per token modest, but every one of the 2.8 trillion parameters must still be resident, which is why the MXFP4 weights alone come to roughly 1.4 terabytes. PER-TOKEN PATH Context up to 1,000,000 tokens native vision folded into pretraining, not bolted on Kimi Delta Attention hybrid linear attention replacing the quadratic form in selected layers vendor claim: 6.3× decode at 1M Sparse MoE router picks 16 experts out of 896 for this token latent-space routing over a pool that never fully activates THE EXPERT POOL — WIDTH IS PROPORTIONAL 16 experts fire under 2% of the pool 896 experts resident, 880 idle for this token WHAT HAS TO BE IN MEMORY ANYWAY 2.8T parameters sparse activation cuts the compute per token, not the memory the model occupies MXFP4 weights ≈1.4 TB to download, against ≈5.6 TB at FP16 — quantisation aware from fine-tuning onward No single GPU fits not an H100, H200 or B200; Moonshot points production users at 64+ accelerators
Sparse activation cuts the compute per token. It does not cut the memory the model occupies.

The architectural bet is sparsity taken further than anyone has shipped openly before. A pool of 896 experts with 16 firing per token means the model reasons with a fraction of its parameters at any moment, which is how a 2.8-trillion-parameter system stays economical to serve at all. Reported figures for the effective active parameter count differ between write-ups; the routing ratio is the number consistently disclosed, and it is the one that matters for the argument below.

Kimi Delta Attention is the part agent builders should look at hardest. Replacing quadratic attention with a hybrid linear form in selected layers targets exactly the failure that makes long-horizon agents expensive: an agent re-sends its whole transcript every step, so decode cost at step 400 is nothing like decode cost at step 4. A claimed 6.3× decode speedup at million-token contexts is a vendor number and should be treated as one, but it is aimed at the right bottleneck — see agent cost control for why that curve, not the per-token price, usually decides what an agent costs.

Moonshot also ships an orchestration layer it calls Agent Swarm, claiming coordination of up to 300 sub-agents and over 4,000 tool calls per task at around 4.5× the throughput of running them in sequence. Treat the multiplier as marketing and the shape as informative: this model was trained and packaged for multi-agent work, not for chat with an agent mode bolted on.

"Open weights" is three claims, and only one holds

What each deployment shape actually gives an agent builder A matrix comparing a closed frontier API, Kimi K3 through its hosted API, Kimi K3 self-hosted, and a small local model across four properties. Only self-hosting a frontier open-weight model gives both frontier capability and freedom from another company's usage policy, and it is the one option with a hard hardware floor. DEPLOYMENT SHAPE × WHAT YOU ACTUALLY GET FRONTIER CAPABILITY LOW COST PER TOKEN RUNS ON MODEST HARDWARE NO EXTERNAL USAGE POLICY Closed frontier API someone else runs it Yes No No hardware at all No Kimi K3, hosted API open weights, rented Near-frontier Dearer than K2.6 No hardware at all No Kimi K3, self-hosted the weights, your cluster Near-frontier Only at high volume 64+ accelerators Yes — the whole point Small local model runs on one machine No Marginal cost ~0 One GPU, or a laptop Yes The row that has no cheap column is the one that made news. The column that only self-hosting fills is the one worth paying for. Fully Conditionally Not at all
No row fills every column. Choosing well means knowing which column you actually came for.

When a frontier model is released as open weights, three different promises get heard at once, and they come apart immediately on inspection.

It does not mean cheap

The clearest evidence is Moonshot's own pricing. K3 lists at $3 per million cache-miss input tokens and $15 per million output tokens; K2.6 sat near $0.95 and $4. The open-weight model is roughly three to four times dearer than the closed-ish predecessor it succeeds. Frontier capability costs what it costs to serve, and publishing the weights does not change the arithmetic of serving them. The one genuine discount is structural rather than competitive: cache-hit input at $0.30 is a 90% reduction, which for an agent re-sending a mostly-stable transcript is the line item worth engineering around. Prompt caching is doing more for your bill here than the license is.

It does not mean local

At MXFP4 the weights are roughly 1.4 TB — against about 5.6 TB at FP16, so the quantisation is doing real work, and because it is quantisation-aware from the fine-tuning stage onward it is the intended precision rather than a lossy compression of some truer model. It is still 1.4 TB. No single H100, H200 or B200 holds it. Distributed inference is not an optimisation here, it is the only mode. Moonshot points production users at supernode configurations of 64 or more accelerators; community self-hosting guides land around eight H100s as a floor and sixteen across two NVLink-connected nodes to run at full speed with the full context window.

"Free to download" and "runs on your hardware" are different sentences. If what you wanted was a model on a machine you own, small and local models remain the answer, and K3 is not in that category — it is a datacentre model whose weights happen to be public.

It does mean nobody else's policy applies

This is the claim that survives, and it is worth more than it sounds. Run K3 on your own cluster and there is no rate limit imposed from outside, no deprecation notice retiring the model under your production traffic, no prompts leaving your boundary, and no refusal behaviour you did not choose. For regulated deployments that is the difference between a workable architecture and an unworkable one — see data residency and sovereignty.

The case for policy independence just got concrete

Until recently the argument for self-hosting a frontier open-weight model was mostly hypothetical for teams outside regulated industries. This month supplied a specific, load-bearing example.

When Hugging Face reconstructed the intrusion described in our write-up of the ExploitGym incident, its responders wanted to point language models at more than 17,000 recorded attacker actions. Commercial frontier models refused: to a safety classifier, close analysis of offensive tradecraft looks much like producing it. The forensics were completed on a self-hosted open-weight model instead, turning what is normally days of work into hours.

Generalise past security. Any workload sitting near a refusal boundary — abuse investigation, content moderation policy work, medical or legal drafting, adversarial red-teaming of your own agent — is a workload where a vendor's judgment about your request can stop your production system on a Tuesday, with no bug to fix and no appeal. That risk does not show up in a benchmark or a price comparison. It shows up once, at the worst moment. A self-hosted frontier model is insurance against it, and until this release, insurance at that capability level was not purchasable.

One caveat worth stating plainly: the same property that makes an unrestricted self-hosted model useful to defenders makes it useful to attackers, and that symmetry is real. It is an argument for knowing which of your workloads genuinely need it, not for pretending the trade-off is free.

When to pick which

Three ways to run an open-weight frontier model, and what each is for Renting Kimi K3 through a hosted API costs the least to start and buys none of the independence open weights promise. Self-hosting demands a cluster of 64 or more accelerators and only pays back at sustained volume, but it is the only shape where no external usage policy applies. A small local model gives up frontier capability entirely in exchange for running anywhere. PICK THE SHAPE THAT MATCHES WHY YOU WANTED OPEN WEIGHTS Rent it HOSTED API Zero hardware, same-day start, $3 in and $15 out per million. You are back under someone else's terms of service — the openness bought you a second vendor, not freedom. Own it SELF-HOSTED CLUSTER 64+ accelerators for production; ≈1.4 TB of weights to place. Pays back only at sustained volume — but no refusal policy, no rate limit and no data leaving your boundary. Shrink it SMALL LOCAL MODEL One GPU or a laptop; marginal cost per call rounds to zero. Gives up frontier reasoning, so route to it for the narrow, high-volume steps and escalate the hard ones elsewhere. Most agent stacks end up using two of these three, not one.
Most agent stacks end up using two of these three.
SituationRent K3 hostedSelf-host K3Stay on a closed API
Evaluating whether it fits at allYes — start here alwaysNoKeep as the baseline to beat
Bursty or low volumeYesNo — idle accelerators dominateFine
Sustained high volumeCompare against amortised cluster costWhere it starts to pay backCompare on total, not per-token
Data cannot leave your boundaryNoYes — the reason to do itNo
Workload sits near a refusal boundaryNoYesNo
Frontier reasoning is the deciding factorNear-frontier, not topNear-frontier, not topStill ahead per Moonshot's own comparison

The practical shape for most teams is not a choice but a split. Rent K3 to find out whether its agentic strengths hold on your traffic; keep a closed frontier model for the hardest reasoning steps; run something small and local for the high-volume narrow steps where capability is not the constraint. That is model routing, and an open-weight frontier model makes it a more interesting problem, not a solved one.

Two things to check before committing. First, the licence: pre-release coverage did not settle the exact terms, so read the model card rather than assuming "open weights" means what it meant for the last release you looked at. Second, run your own evaluation. Every number in this post is someone else's benchmark, and reading benchmarks well mostly means remembering that a leaderboard position is evidence about a leaderboard.

FAQ

Can I run Kimi K3 on my own machine?

No, not in any ordinary sense. The MXFP4 weights are around 1.4 TB and no single H100, H200 or B200 can hold them. Community guides put the practical floor near eight H100-class GPUs, with Moonshot pointing production deployments at 64 or more accelerators. It is a datacentre model with public weights.

Is Kimi K3 cheaper than closed frontier models?

Not automatically, and it is not even cheaper than its own predecessor — K3 lists at $3 and $15 per million input and output tokens against roughly $0.95 and $4 for K2.6. Whether it beats a given closed model depends on that model's pricing and on how many tokens each takes to finish your task, which is a per-workload measurement rather than a table lookup.

What does open weights actually buy me, then?

Independence from another company's operational decisions: no external rate limits, no deprecation of the model under your traffic, no data leaving your boundary, and no refusal policy you did not choose. If none of those constrain you today, the honest answer is that hosted access to a closed model is simpler and probably better.

Is it good enough to replace my current agent model?

On agentic benchmarks it is close to the top and ahead of GPT-5.6 Sol Max on at least one long-horizon suite, while Moonshot itself says K3 trails Claude Fable 5 and GPT-5.6 Sol overall. "Close to the frontier and downloadable" is the accurate summary. Whether close is enough is a question about your task's error tolerance, and the only way to answer it is to run your own evals.

Does the million-token context mean I can stop managing context?

No. A larger window changes what you are able to do, not what you should do — attention quality degrades across very long contexts, and every token you keep is re-sent and re-paid on every subsequent step. The window is a ceiling, not a strategy.

Further reading

On this wiki:

Sources: