Kimi K2.6

What it is. Kimi K2.6 is Moonshot AI’s open-weight visual-and-text model, released 2026-04-20: a 1T-parameter MoE with 32B activated parameters and a 256K-token context window. Why it matters. It was Moonshot’s open-weight workhorse before Kimi K3, and it publishes a full architecture card — a rarity among open releases — while remaining far lighter to self-host than the 2.8T K3. Key characteristics. 384 routed experts (8 selected + 1 shared), MLA attention with SwiGLU activation, 61 layers, a 400M MoonViT vision encoder, thinking and non-thinking modes, and a Modified MIT license. What a professional should know. It is superseded as Moonshot’s flagship by K3 but is the practical open-weight Kimi to download; its only recorded benchmarks are two vendor-reported HLE rows.

Architecture and parameters

K2.6 publishes a detailed card: 1T total / 32B activated (~3.2% activation), 384 routed experts with 8 selected + 1 shared per token, MLA (Multi-head Latent Attention) with SwiGLU activation, 61 layers (1 dense), and a 400M-parameter MoonViT vision encoder. China AI Hub analysis indicates this is a mid-scale MoE tuned for economy — the 32B active set keeps per-token cost low while the 1T total still dictates self-hosting memory, making it a multi-GPU but not datacenter-scale deployment. See Mixture-of-Experts and the MoE architectures research.

What the context window actually means

The 262,144-token context with a 98,304 output ceiling is the standard mid-tier split — long-input analysis, not long-form generation. China AI Hub analysis indicates this is the clearest spec-level gap from K3’s 1M-in/1M-out pairing, and it is what keeps K2.6 a code/dialogue tool rather than a long-form synthesis tool. See the context-window research.

Pricing implications

$0.95 input / $4.00 output per 1M tokens, cache hits $0.16 (17% of input). China AI Hub analysis indicates the 4.2x input-to-output ratio is mid-tier, and the cache pricing rewards agent loops that re-read shared context. See choosing-by-price.

API, coding, and agent implications

K2.6 supports thinking and non-thinking modes for dialogue and agent tasks, with native vision. It is the model the K-series coding line (Kimi K2.7 Code) and the Kimi Code agent evolved from, though its own coding benchmark evidence is not separately recorded. See choosing-a-coding-model.

Open weights and license

Open weight under Modified MIT, self-hostable. China AI Hub analysis indicates the Modified MIT label is a softer constraint than plain MIT — permissive for most uses but with conditions that should be read before commercial redistribution, distinct from the revenue-threshold licenses on K3 and the Qwen3.8-Max-class weights. See licensing explained and the open-weights hub.

Benchmark interpretation

The two recorded rows are both vendor-reported HLE scores: 36.4 text-only (no tools) and 55.5 with tools. The 19-point gap is the same tool-access pattern the database records elsewhere — tool use materially raises the ceiling on the hardest knowledge exam.

What the benchmarks do not prove

Two vendor-reported HLE rows do not independently verify K2.6 against competitors, and there are no recorded coding or agent benchmarks for K2.6 despite its positioning. Cross-vendor comparison is invalid without matched harnesses and versions. See how to read vendor-reported benchmarks and the benchmark-methodology research.

Suitable and less suitable workloads

Well-suited: open-weight visual-and-text deployment, self-hosting where a permissive-ish license and a 1T/32B footprint fit, dialogue and agent tasks, and multimodal (vision) input. Less suited: 1M-context long-form work, frontier-scale generation, and any use case that has already standardized on K3.

Where this model fits

Workload	Relevance
Long-context analysis	Moderate
Coding	Moderate
Structured API workflows	Moderate
Agent orchestration	Moderate
Local self-hosted deployment	High
GUI automation	No evidence
Video generation	No evidence
Enterprise cloud	Moderate

Relevance judgments are China AI Hub analysis based on documented capabilities, not vendor claims.

China AI Hub analysis

China AI Hub analysis indicates K2.6’s enduring value is as the practical open-weight Kimi: it retains a full architecture card and a permissive-style license while being roughly a third the total parameters of K3 (1T vs 2.8T), which is why it remains the more tractable self-host choice even after K3 took the flagship role. Its thin benchmark record — two HLE rows and nothing for coding or agents — is the clearest sign that it is a transitional release: capable and transparent, but quickly overtaken by the K2.7 coding line and K3 within the same family. See the state of China’s AI models research and the MoE hub.

What is uncertain

• The Modified MIT license’s exact conditions are less fully spelled out than plain MIT and should be read before redistribution.
• Only two benchmark rows (both HLE, both vendor-reported) are recorded; no coding or agent benchmarks are published.
• The 98,304-token output ceiling is documented as the evaluation-configuration max, not necessarily a hard API cap.

Sources

• Kimi API platform — model list
• Kimi API — pricing (chat)
• Hugging Face model card — Kimi K2.6 (architecture, parameters, benchmarks)

Labels used above: Official fact (architecture, parameters and pricing from Moonshot’s docs and the HF card), Vendor-reported claim (the two HLE benchmark scores), and China AI Hub analysis (our synthesis, always introduced as such). No third-party benchmark evidence is currently recorded for this model.