Moonshot Launches Kimi K3 With 2.8 Trillion Parameters and 1M Context

Moonshot AI said Thursday that it had launched Kimi K3, a model built with 2.8 trillion total parameters. K3 is available now through Moonshot’s products and API. Full weights remain unavailable; Moonshot has scheduled their release for July 27.

Moonshot activates 16 of 896 experts

K3 combines native vision with a one-million-token context window. Moonshot says its Stable LatentMoE design effectively activates 16 of 896 experts. Kimi Delta Attention processes long sequences more efficiently, according to the company, while Attention Residuals retrieves representations from different depths. Moonshot puts the resulting gain in overall scaling efficiency at about 2.5 times that of Kimi K2. Outside researchers had no released K3 weights to test the measure on launch day.

For inference, Moonshot recommends supernode configurations with at least 64 accelerators. According to the launch post, the company began quantization-aware training at the supervised fine-tuning stage, using MXFP4 weights and MXFP8 activations for broader hardware compatibility. Kimi Delta Attention complicates conventional prefix caching, Moonshot notes, and the company plans to release a corresponding vLLM implementation with the model.

Moonshot’s benchmark table

Moonshot’s vendor-run table presents K3 as competitive with proprietary rivals. It reports a score of 88.3 on Terminal Bench 2.1 and 81.2 on FrontierSWE. Different agent harnesses limit direct model-to-model comparisons: the notes list KimiCode, Claude Code and Codex, depending on the benchmark. They add that some Claude Fable 5 results may include fallback to Claude Opus 4.8. The launch material did not provide a mature independent K3 benchmark using one controlled setup.

Moonshot’s vendor-run benchmark comparison. Source: Moonshot AI, “Kimi K3: Open Frontier Intelligence”.

Moonshot warns that K3 may become highly unstable when agent software fails to return the model’s full prior reasoning. It also cautions that ambiguous tasks can prompt unexpected decisions and recommends explicit behavioral constraints for bounded applications.

Moonshot’s internal knowledge-work evaluation. Source: Moonshot AI, “Kimi K3: Open Frontier Intelligence”.

Moonshot’s launch post acknowledges that K3’s overall user experience trails Claude Fable 5 and GPT 5.6 Sol, while presenting benchmark scores that put K3 near those systems on several coding and agent tests.

TestingCatalog and HCPTangHY test K3

On one universe-simulation prompt before the release, TestingCatalog compared Claude Fable 5 with an anonymized Arena checkpoint believed to be K3. Fable 5 finished faster and produced sturdier interface components, according to the publication. The likely K3 checkpoint made the more elaborate visual build, including a first-person camera view around a selected planet. TestingCatalog also reported long runtimes on difficult agent tasks. The anonymized, pre-release comparison does not establish the final model’s performance.

After testing the released build, the account HCPTangHY posted observations on Linux.do describing few severe bugs and strong coding ability. The post judged K3 less polished than frontier systems and said reasoning runs above 1,000 seconds were common. HCPTangHY also reported that K3 performed at least as well as Opus 4.6 in the account’s bot environment. The findings came from one developer’s setup and are not a representative performance measurement.

Moonshot sets API prices

Moonshot prices cache-hit input at $0.30 per million tokens. Cache-miss input costs $3 per million tokens, and output costs $15 per million. The company is serving K3 through Kimi.com, Kimi Work, Kimi Code and the Kimi API, with maximum thinking effort enabled by default at launch. Low- and high-effort modes will arrive in later updates, Moonshot says. The company also claims an API cache-hit rate above 90% in coding workloads, a workload-specific figure that has not been independently verified.

Moonshot’s Kimi K3 showcase graphic. Source: Moonshot AI, “Kimi K3: Open Frontier Intelligence”.

The full weights are due July 27. Moonshot plans a technical report with additional details on K3’s architecture, training and evaluations, but has not set a publication date. Until the weights arrive, outside researchers cannot examine the model directly or run it under independently controlled setups.

Thinking Machines’ Inkling Takes U.S. Open-Model Lead With 41 Score

Artificial Analysis independently scored Inkling at 41 on its Intelligence Index, putting Thinking Machines Lab’s first production model ahead of every other U.S. open-weight release the firm tracks.

The Implicator

GLM-5.2 Edges Kimi K2.7 Code in Early Coding Tests

The two strongest Chinese open models to launch this month, Z.ai's GLM-5.2 and Moonshot's Kimi K2.7 Code, have now been run side by side by five independent reviewers, and the early verdict gives GLM-

The Implicator

Alibaba Ships Qwen3.6-27B, an Open-Weight Coding Model That Beats Its 397B MoE

Alibaba on Wednesday released Qwen3.6-27B, a dense 27-billion-parameter open-weight model under Apache 2.0 that tops its own 397B-parameter predecessor on every major agentic coding benchmark. The mod

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai