Moonshot AI has launched Kimi K3, a 2.8-trillion-parameter model with native multimodal capabilities and a 1-million-token context window. The company positions it as the world’s first open 3T-class model. K3 is available today on kimi.com, Kimi Work, Kimi Code, and via the Kimi API. Full model weights are scheduled for release by July 27, 2026.
What’s New
According to Moonshot’s official technical blog, Kimi K3 introduces two architectural updates that improve how information flows across sequence length and model depth:
- Kimi Delta Attention (KDA): An efficient attention foundation designed for long-context scaling. Moonshot says KDA helps serve million-token contexts at a competitive token price when paired with prefill caching.
- Attention Residuals (AttnRes): Selectively retrieves representations across depth rather than accumulating them uniformly.
- Stable LatentMoE: Kimi K3 uses Stable LatentMoE, activating 16 of 896 experts. Combined with KDA, Attention Residuals, and updated training recipes, Moonshot reports roughly 2.5× better scaling efficiency versus Kimi K2.
The model applies quantization-aware training from the supervised fine-tuning stage onward, using MXFP4 weights and MXFP8 activations. Moonshot recommends deploying K3 on supernode configurations with 64 or more accelerators. Its serving stack, built on Mooncake’s disaggregated inference architecture, achieves above 90% cache hit rate on coding workloads.
API pricing (official): $0.30 per million tokens for cache-hit input, $3.00/MTok for cache-miss input, and $15.00/MTok for output. At launch, K3 uses max reasoning effort by default; lower- and high-effort modes are planned for later updates.
Moonshot notes that overall performance still trails the strongest proprietary systems it cites — Claude Fable 5 and GPT 5.6 Sol — while claiming frontier-level results against other tested models, especially on long-horizon coding and multimodal agent workflows.
Why It Matters
Open-model releases have mostly clustered well below the trillion-parameter class. Kimi K3 is Moonshot’s bid to push that ceiling: the company says that for nine of the past twelve months, Kimi models have set the upper bound of open-model sizes. A 2.8T MoE with planned weight release matters less as a vanity count and more as a signal that self-hostable frontier-scale options keep expanding — if the July 27 weight drop ships as promised.
Two practical storylines stand out from Moonshot’s case studies:
- Self-improving infrastructure: Late in development, an early K3 build handled most of the team’s own GPU kernel optimization work.
- MiniTriton: In Moonshot’s evaluation, K3 produced a compact Triton-like GPU compiler (MiniTriton) with its own IR and PTX pipeline, matching or beating Triton on some roofline benchmarks and sustaining nanoGPT training.
Because multimodal training starts from the first step — not as a bolted-on vision adapter — K3 can act as a “vision-in-the-loop” agent: write code, take screenshots of the result, and iterate. That is the capability Moonshot showcases for game development, frontend work, and CAD-style workflows.
Our Take
Treat “open-source” marketing carefully. What Moonshot is promising is an open-weight release of a 3T-class model, with training code and full reproducibility still proprietary. That is still a big deal for the open ecosystem — but the real test is July 27, when outsiders can verify parameter counts, serving recipes, and benchmark claims.
Hardware reality will gate adoption. A recommended footprint of 64+ accelerators means this is not a “download and run on one box” model. For most teams, the near-term product is the API and Moonshot’s apps, with self-hosting reserved for well-funded labs and enterprises.
If the weight release lands cleanly and community stacks (including Moonshot’s promised vLLM KDA/prefill-cache work) catch up, K3 resets the open-model size bar. Until then, the responsible headline is not “fully open today,” but “largest open 3T-class launch with a dated weight commitment.”
FAQ
When will Kimi K3 weights be released?
Moonshot says full model weights will be released by July 27, 2026.
How many parameters does Kimi K3 have?
2.8 trillion total parameters. It is marketed as the first open 3T-class model.
How does Stable LatentMoE work in K3?
K3 activates 16 of 896 experts per token. Moonshot attributes roughly 2.5× better scaling efficiency versus Kimi K2 to the combination of Stable LatentMoE, KDA, AttnRes, and updated training recipes — not to MoE routing alone.
What hardware is needed to self-host Kimi K3?
Moonshot recommends supernode setups with 64 or more accelerators, reflecting the model’s scale and communication needs.
How much does the Kimi K3 API cost?
Input: $0.30/MTok (cache hit) or $3.00/MTok (cache miss). Output: $15.00/MTok. Official coding workloads report >90% cache hit rates on Mooncake.
Is Kimi K3 better than GPT 5.6 Sol or Claude Fable 5?
Moonshot itself says K3 still trails those proprietary leaders overall, while claiming strong results versus other open models and competitive performance on coding and multimodal agent tasks.
Is Kimi K3 fully open source?
Note in the strict OSI sense today. The API and apps are live; weights are scheduled for July 27. Training code and datasets remain proprietary. Prefer “open-weight / open 3T-class” wording until the weight license is public.
