Moonshot AI released the full version of Kimi K2.6 last week. This is not a "code preview" beta but a complete production-grade model. The earlier Code Preview had already drawn attention, but the technical specifications and benchmark results of the full version turned out to be more interesting than many expected.
HLE-Full 54.0: First place among open-source models is real
HLE (Humanity's Last Exam) is one of the most prestigious AI benchmarks this year, featuring 2,500 PhD-level questions spanning over 100 academic fields. It is not the kind of test that can be faked with a few few-shot examples.
Kimi K2.6 results:
| Model | HLE-Full (with tools) |
|---|---|
| Kimi K2.6 | 54.0 |
| Claude Opus 4.6 | 53.0 |
| GPT-5.4 | 52.1 |
The gap is not large, but the direction is clear: an open-source model has outperformed two closed-source flagships on a top-tier academic reasoning benchmark. SWE-Bench Pro 58.6, BrowseComp 83.2, CharXiv (with Python) 86.7, Math Visual Reasoning 93.2.
Moonshot AI used "far more execution and imagination" in its release notes to describe the improvements in this version.
1 trillion parameters, but only 32 billion activated
This is the core logic of the MoE architecture. Kimi K2.6 is a 1-trillion-parameter Mixture-of-Experts model that activates only 32B parameters per inference. 384 experts are managed in groups, and each response calls 8 routing experts plus 1 shared expert. The benefits are two-way:
- Training side: Larger total parameters mean more knowledge learned
- Inference side: Fewer activated parameters keep computational costs manageable
The attention mechanism uses MLA (Multi-Head Latent Attention), which compresses processed data into lightweight mathematical representations, reducing KV cache usage. The activation function is SwiGLU, which is more hardware-friendly than the previous generation. The vision module is a separate 400M parameter encoder supporting image and multimedia input.
The context window is 256K, and INT4 quantization is supported, making local deployment feasible.
300 parallel agents, 12 hours continuous operation
This is the part of Kimi K2.6 that truly interests developers:
- Supports up to 300 sub-agents running in parallel
- Up to 4,000+ tool calls per single task
- Continuous runtime can reach 12 hours without interruption
To support multi-agent collaboration at this scale, Moonshot AI introduced "Claw Groups" — breaking large tasks into sub-groups that allow humans and AI to collaborate. Humans intervene at key decision points, while AI handles the bulk of intermediate execution. The logic is similar to human-machine division of labor on a factory assembly line.
This design is practical for enterprise applications. Real-world workflows are not a binary choice between "full automation" and "full manual" — they require human oversight at critical nodes while AI does the heavy lifting in between.
Open-source ecosystem support from day one
On launch day, the model already supported vLLM, OpenRouter, Cloudflare Workers AI, and MLX. This was the result of ecosystem integration work that began months in advance.
- MLX: Runs on Apple Silicon
- OpenRouter: Direct access via API aggregation platform
- vLLM: High-throughput inference deployment without obstacles
- Cloudflare Workers AI: Edge inference now has a viable path
Completing these integrations on day one was no accident — it shows that Moonshot AI did serious engineering preparation before the release.
The open-source race is catching up seriously
Last year, some still said "open-source models have a clear gap with GPT-4." This year, that statement no longer holds.
Kimi K2.6 outperformed Claude Opus 4.6 and GPT-5.4 on HLE. DeepSeek V3.2 changed its attention to a sparse structure, cutting costs in half. Qwen3.6-35B runs on a MacBook, scoring 73.4% on SWE-bench.
Open source is no longer a low-cost substitute for closed source — it is now leading on specific benchmark dimensions. For enterprises, this means that needs for private deployment, data confidentiality, and cost control finally have flagship-grade options available.
There is still a gap between benchmark scores and real-world production. The 54.0 on HLE and everyday code debugging, long-document understanding, and multi-turn conversations are different things. But the direction is clear: Moonshot AI is fighting this battle seriously, and this round is won.
Sources: Moonshot AI releases Kimi-K2.6 model with 1T parameters,CocoLoop, attention optimizations (SiliconANGLE); [AINews] Moonshot Kimi K2.6: the world's leading Open Model refreshes to catch up to Opus 4.6 (Latent Space)