Moonshot AI's Kimi K3 launched on July 16, 2026, at 2.8 trillion total parameters - the largest open-weight model ever shipped - and Artificial Analysis gave it a score of 57 on its Intelligence Index, ranking it fourth among 189 models. That puts it, by a public independent measure, ahead of Claude Opus 4.8. The reaction ran immediately to "open weights have arrived." That claim deserves a closer look before you reorganize your model stack around it.
What Kimi K3 actually is
Kimi K3 is a 2.8 trillion parameter Mixture-of-Experts multimodal model from Moonshot AI, released July 16, 2026, with a 1 million token context window, always-on reasoning with a tunable reasoning_effort control, and input support for text, images, and video.
It uses Stable LatentMoE with 896 experts and activates 16 of them
- meaning that despite the headline parameter count, the active compute per token is closer to a mid-sized dense model than the 2.8T figure implies.
This is typical MoE design, but it means "2.8 trillion parameter model" is not comparable to a hypothetical 2.8T dense model. If someone tells you K3 has "2.8 trillion parameters of intelligence," they are confusing total capacity with per-token compute.
The architecture does have genuine novelty. K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), two architectural updates designed to improve how information flows across sequence length and model depth. It scales up MoE sparsity, activating 16 out of 896 experts with a Stable LatentMoE framework. Together with refined training and data recipes, these structural changes yield an approximate 2.5× improvement in overall scaling efficiency compared to Kimi K2.
Where the benchmark results hold up - and where they don't
Coding is the strongest case. SWE Marathon, the longest-running software engineering benchmark covering tasks that require sustained multi-step reasoning over extended codebases, puts K3 at 42.0, ahead of Opus 4.8 (40.0), GPT-5.6 (39.0), and Fable 5 (35.0) by meaningful margins. For teams running agentic coding pipelines, this is the most relevant result in the table.
On AA-Briefcase, a private agentic benchmark from Artificial Analysis designed to test long-horizon knowledge work, K3 climbed to second place with a score of 1,527 - beating GPT-5.6 Sol Max (1,495) and trailing only Fable 5 Max (1,587).
The honest read is more mixed elsewhere:
- K3 wins the endurance benchmarks - Marathon and Program Bench - but loses on deep repo comprehension (DeepSWE).
- Artificial Analysis independently places Kimi K3 near the frontier, but also measures high output-token use, slower-than-median generation, and premium pricing.
- Moonshot's 77.8 ProgramBench figure is raw hidden-test pass rate, not the much stricter "fully resolved" task rate.
- Third-party trackers like BenchLM are still declining to give K3 a formal composite ranking pending their own verification.
The pricing structure compounds the ambiguity. K3 always reasons, and the only launch setting is max. A $15 output rate becomes expensive when a task generates a long reasoning trace, retries failed tool calls, or drags its history through many turns. The headline number does not reflect what a real agentic workload costs. A teammate routing dozens of tasks through K3 per day will hit this faster than any benchmark session does. Tools like Beagle, which draft responses and surface context in-thread, are particularly sensitive to per-output pricing at scale.
| Benchmark | K3 | Claude Fable 5 | GPT-5.6 Sol | Claude Opus 4.8 |
|---|---|---|---|---|
| AA Intelligence Index | 57 | higher | higher | lower |
| SWE Marathon | 42.0 | 35.0 | 39.0 | 40.0 |
| AA-Briefcase Elo | 1,527 (2nd) | 1,587 (1st) | 1,495 | below K3 |
| Frontend Code Arena | #1 (1,679) | 2nd | 3rd | - |
| Output price / M tokens | $15 | higher | higher | lower |
Sources: Artificial Analysis Intelligence Index, a composite benchmark evaluating models across reasoning, knowledge, mathematics, and coding; Moonshot launch materials; API pricing at $0.30 per million cache-hit input tokens, $3 per million on cache misses, and $15 per million output tokens.
The bigger story: open-weight cyber capability is a timed window
Buried in the same week as the K3 launch, the UK's AI Security Institute published its first public measurement of the open-weight cyber capability gap. This is the finding most model coverage skipped.
The gap between the cyber capabilities of open and closed models matters because it provides a preparation time: a window for cyber defenders with access to the most capable closed systems to take action before today's frontier cyber capabilities might become available without the same safeguards. In April 2026, two closed models, Mythos Preview and GPT-5.5, demonstrated some of the largest jumps in AI cyber capability AISI has observed since testing began.
AISI's evaluations found that GLM-5.2 (June 2026) was the most cyber-capable open-weight model at time of testing. On narrow tasks, GLM-5.2 performs comparably to the most cyber-capable models released four months before it (Opus 4.6 and GPT-5.3-Codex), and this holds across all four difficulty levels. DeepSeek V4-Pro is comparable to Opus 4.5, released five months before it. Both gaps are narrower than internal evaluations AISI conducted in 2025, when open-weight models lagged the frontier by six to ten months.
The same evaluation covers one detail the summary numbers obscure: AISI reports its evaluations were "largely unimpeded," with DeepSeek V4-Pro occasionally refusing reverse-engineering tasks - refusals circumvented by simply attempting them again.
AISI intends to evaluate Moonshot's Kimi K3 on the same basis once its weights are released at the end of July - a test of whether the largest open model yet published continues the narrowing trend or resets it.
This is the non-obvious second-order consequence of weeks like this one: the "open-weight frontier gap" is not just a capability story. It is a security clock. As teams evaluate whether to self-host open-weight models, the AISI finding is worth reading alongside the benchmark tables.
What a team should actually do this week
K3 is not yet deployable unless you are going through Moonshot's API or OpenRouter. Its weights are not public until July 27, so treat it as pending rather than deployable today.
If your workload is long-horizon agentic coding - pipelines that run for many steps, touch multiple files, or retry on failure - K3's SWE Marathon and AA-Briefcase results are worth watching closely. K3 wins the endurance benchmarks but loses on deep repo comprehension. If your workload involves long agentic loops, K3's profile fits. If you need single-pass deep file analysis, Fable 5 and GPT-5.6 still lead.
If you are evaluating open-weight self-hosting, the hardware constraint is real. Running a model of this size requires substantial GPU infrastructure. Inference at 2.8 trillion parameters is not something that runs on a single server rack.
Moonshot recommends a deployment spanning at least 64 accelerators. For most teams, that points toward API access rather than self-hosting - which means the "open-weight" benefits of privacy and control do not yet apply to K3 for most organizations.
A more practical path for teams who need open-weight today: GLM-5.2, Z.ai's 744-billion-parameter MoE, scores 91.2% on GPQA Diamond and 62.1% on SWE-bench Pro, at a fraction of frontier API pricing. It is available now, independently verified, and runs on hardware that does not require a multi-rack deployment.
Beagle in action#ai-tools, 10:02amOpen-weight frontier model gap: common questions
What is the open-weight frontier model gap?
The open-weight frontier gap is the performance distance between the best freely downloadable models and the best closed, API-only models. AISI's July 2026 measurement puts open-weight cyber capability at four to seven months behind the closed frontier, down from six to ten months through most of 2025. On general benchmarks, the gap varies by task type - it is narrowest on coding and widest on safety-critical and reasoning-heavy tasks.
Is Kimi K3 the best open-weight model available now?
As of July 19, 2026, GLM-5.2 tops most open-source LLM leaderboards for agentic coding and reasoning. Kimi K3 is a close second on raw capability, but its weights are not public until July 27, so treat it as pending rather than deployable today.
Can I run Kimi K3 locally?
Not yet, and not easily even after July 27. K3 has 2.8 trillion total parameters, a 1M-token context window, 896 experts, and a recommended deployment spanning at least 64 accelerators. Most teams will access it through the Moonshot API or OpenRouter rather than self-hosting.
Does the open-weight vs closed model choice matter for security?
Yes, and the AISI report makes this concrete. AISI has framed the open-closed capability gap as something concrete: preparation time. Cyber defenders with access to the most capable closed systems get a window to act before equivalent capability becomes downloadable without the same controls. That window is now measured, and it is shrinking.
How should teams evaluate Kimi K3 benchmarks?
Look at which benchmark matches your workload, not the composite rank. No single coding benchmark tells the full story: DeepSWE tests repo-level understanding, Terminal-Bench tests command-line tool use, SWE Marathon tests sustained multi-step reasoning, and Program Bench tests general program construction. Run a narrow eval on your own task distribution before committing.