Kimi K2.5 / K2.6 / K2.7-Code

Overview

The Kimi K2.5 → K2.6 → K2.7-Code line is Moonshot AI's 1-trillion-parameter open-weights MoE family, all three sharing an identical architecture: 1T total parameters, 32B activated per token, 61 layers, MLA attention, 384 routed experts with 8 selected per token plus 1 shared expert, a 256K context window and a 400M-parameter MoonViT vision encoder (Kimi-K2.5 model card; Kimi-K2.6 model card; Kimi-K2.7-Code model card). Moonshot states this explicitly on the later cards — "Kimi-K2.6 has the same architecture as Kimi-K2.5, and the deployment method can be directly reused" and "Kimi-K2.7-Code has the same architecture as Kimi-K2.5/Kimi-K2.6" (Kimi-K2.6 model card; Kimi-K2.7-Code model card). The text backbone is DeepSeek-style: K2.7-Code's config.json declares the text architecture as DeepseekV3ForCausalLM with topk_method: noaux_tc, sigmoid expert scoring, YaRN RoPE scaling and MLA-style q_lora_rank/kv_lora_rank projections (Kimi-K2.7-Code config.json).

Release timeline, per Moonshot's own channels: Kimi K2.5 was published 2026-01-27 and is an open-source native multimodal agentic model built by continual pretraining on ~15 trillion mixed visual and text tokens on top of Kimi-K2-Base, adding instant and thinking modes and the Agent Swarm execution scheme (Kimi Research & Tech Blog; Kimi-K2.5 model card); its card also logs a 2026.1.29 changelog removing the default system prompt and fixing the <|media_begin|> token (Kimi-K2.5 model card). Kimi K2.6 was published 2026-04-20 and pushes long-horizon coding, coding-driven design, proactive autonomous execution and swarm orchestration scaling to 300 sub-agents executing 4,000 coordinated steps (Kimi Research & Tech Blog; Kimi-K2.6 model card). Kimi K2.7-Code was announced on Moonshot's official X account on June 12, 2026 as "our latest coding model … now released and open-sourced," with +21.8% on Kimi Code Bench v2, +11.0% on Program Bench and +31.5% on MLS Bench Lite versus K2.6, 30% lower reasoning-token usage, and availability via the Kimi API and Kimi Code (@Kimi_Moonshot).

Positioning shifts across the family from general multimodal agent to coding specialist. K2.5 is framed around native multimodality, coding from visual specifications, and an agent swarm that decomposes tasks into parallel domain-specific sub-agents (Kimi-K2.5 model card). K2.7-Code is explicitly "a coding-focused agentic model built upon Kimi K2.6," strengthening end-to-end completion of complex software-engineering workflows while cutting thinking tokens ~30% versus K2.6; it forces thinking and preserve_thinking on (they cannot be disabled), unlike K2.6 where preserve-thinking is off by default and an instant (non-thinking) mode exists (Kimi-K2.7-Code model card; Kimi-K2.6 model card). All three ship under the Modified MIT License for both code repository and weights (Kimi-K2.7-Code model card; Kimi-K2.6 model card). Notable techniques include native INT4 quantization ("the same native int4 quantization method as Kimi-K2-Thinking"), interleaved thinking with multi-step tool calls, and Kimi Code CLI as the recommended agent framework; supported inference engines are vLLM, SGLang and KTransformers with transformers >=4.57.1, <5.0.0 (Kimi-K2.7-Code model card; Kimi-K2.6 model card). Moonshot's API platform lists dedicated pricing pages for kimi-k2.7-code and kimi-k2.6, and notes that kimi-k2.5 and the moonshot-v1 series are no longer available to newly registered users following the K3 launch (Kimi API pricing docs; Kimi API model list).

Architecture

Shared by K2.5, K2.6 and K2.7-Code unless noted.

• Architecture family: Mixture-of-Experts (MoE), native multimodal (image + video input, text output) (Kimi-K2.7-Code model card; Kimi-K2.5 model card)
• Total parameters: 1T; Hugging Face reports the K2.7-Code safetensors model size as 1T params (Kimi-K2.7-Code model card)
• Activated parameters per token: 32B (Kimi-K2.7-Code model card; Kimi-K2.6 model card; Kimi-K2.5 model card)
• Layers: 61 including 1 dense layer (Kimi-K2.7-Code model card); num_hidden_layers: 61, first_k_dense_replace: 1 (Kimi-K2.7-Code config.json)
• Attention mechanism: MLA (Multi-head Latent Attention) (Kimi-K2.7-Code model card); config: q_lora_rank: 1536, kv_lora_rank: 512, qk_nope_head_dim: 128, qk_rope_head_dim: 64, v_head_dim: 128 (Kimi-K2.7-Code config.json)
• Attention hidden dimension / heads: 7168 hidden, 64 attention heads (Kimi-K2.7-Code model card); hidden_size: 7168, num_attention_heads: 64, num_key_value_heads: 64 (Kimi-K2.7-Code config.json)
• Experts: 384 routed experts, 8 selected per token, 1 shared expert (Kimi-K2.7-Code model card); n_routed_experts: 384, num_experts_per_tok: 8, n_shared_experts: 1, moe_layer_freq: 1 (Kimi-K2.7-Code config.json)
• MoE dimensions / routing: MoE hidden dimension per expert 2048 (Kimi-K2.7-Code model card); moe_intermediate_size: 2048, dense intermediate_size: 18432, sigmoid expert scoring, topk_method: noaux_tc, norm_topk_prob: true, routed_scaling_factor: 2.827, aux_loss_alpha: 0.001 — the DeepSeek-V3-style MoE recipe (Kimi-K2.7-Code config.json)
• Text backbone class: DeepseekV3ForCausalLM inside KimiK25ForConditionalGeneration (model_type: kimi_k25, text model_type: kimi_k2) (Kimi-K2.7-Code config.json)
• Activation function: SwiGLU per the model cards; hidden_act: silu in config (Kimi-K2.7-Code model card; Kimi-K2.7-Code config.json)
• Context window: 256K per the cards; max_position_embeddings: 262144 (Kimi-K2.7-Code model card; Kimi-K2.7-Code config.json). Evaluations were run at a 262,144-token context length (Kimi-K2.7-Code model card)
• Positional encoding: YaRN RoPE scaling, rope_theta: 50000.0, factor 64.0, original_max_position_embeddings: 4096, beta_fast 32.0 / beta_slow 1.0 (Kimi-K2.7-Code config.json)
• Vocabulary: 160K per the cards; vocab_size: 163840 (Kimi-K2.7-Code model card; Kimi-K2.7-Code config.json)
• Vision encoder: MoonViT, 400M parameters (Kimi-K2.7-Code model card); vision config: 27 layers, 16 heads, hidden size 1152, intermediate 4304, patch_size: 14, patchmerger projector, video_attn_type: spatial_temporal (Kimi-K2.7-Code config.json)
• Precision / quantization: native INT4 — same method as Kimi-K2-Thinking (Kimi-K2.7-Code model card); config declares compressed-tensors pack-quantized, num_bits: 4, group_size: 32, symmetric, with attention, shared experts, dense MLP, lm_head, vision tower and mm projector excluded; published tensor types are BF16/F32/I32 (Kimi-K2.7-Code config.json; Kimi-K2.7-Code model card)
• Novel components (KDA, LatentMoE, SiTU-GLU): n.a. — these are Kimi K3 components and are not present in the K2.x cards or config, which specify MLA + SwiGLU + DeepSeek-style MoE instead (Kimi-K2.7-Code model card; Kimi-K2.7-Code config.json)
• Thinking behavior: K2.7-Code forces thinking and preserve_thinking to True and does not support instant mode; K2.6/K2.5 support instant vs thinking modes with preserve_thinking off by default (Kimi-K2.7-Code model card; Kimi-K2.6 model card)
• Recommended sampling: temperature 1.0, top-p 0.95 in thinking mode (K2.6 instant mode: 0.6) (Kimi-K2.7-Code model card; Kimi-K2.6 model card)
• Pretraining data (K2.5): continual pretraining on ~15 trillion mixed visual and text tokens atop Kimi-K2-Base (Kimi-K2.5 model card)
• Token efficiency (K2.7-Code): ~30% fewer thinking tokens than K2.6 (Kimi-K2.7-Code model card)
• Serving: vLLM, SGLang, KTransformers; transformers >=4.57.1, <5.0.0 (Kimi-K2.7-Code model card)

Official vendor eval scores

Vendor-reported only. K2.7-Code's model card publishes six benchmarks (all with the K2.6 baseline alongside); the broader public suites (SWE-bench Verified, LiveCodeBench, AIME, HMMT, GPQA, MMLU-Pro) are reported by Moonshot for K2.6 and K2.5 but not re-run for K2.7-Code.

Vendor-reported deltas for K2.7-Code versus K2.6, stated in Moonshot's launch post: +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, +31.5% on MLS Bench Lite, and 30% lower reasoning-token usage (@Kimi_Moonshot).

Evaluation conditions: K2.7-Code and K2.6 were tested with thinking mode enabled via Kimi Code CLI at temperature 1.0, top-p 0.95 and 262,144-token context; MCP-Atlas and MCPMark-Verified used a 100 tool-call/step budget with 32k max tokens per step, averaged over 3 runs (Kimi-K2.7-Code model card). K2.6's SWE-Bench family results used an in-house harness adapted from SWE-agent, averaged over 10 runs (Kimi-K2.6 model card).

Benchmark explanations

• Kimi Code Bench v2 — Moonshot's in-house coding-agent benchmark of realistic software-engineering tasks across 10+ mainstream languages and a full production stack (backend, infra, performance, systems, security, frontend, ML/data) (Kimi-K2.7-Code model card).
• Program Bench — 200 tasks asking an agent to recreate a program's behavior from only a compiled binary plus documentation (no source, decompilation or internet), judged against 248,000+ fuzz-generated behavioral tests (Kimi-K2.7-Code model card).
• MLS Bench Lite — official 30-task subset of MLS-Bench testing whether an agent can invent generalizable, scalable ML methods (LLM pretraining/post-training, robotics, world models, CV, RL, ML systems), with 5 hours per task (Kimi-K2.7-Code model card).
• Kimi Claw 24/7 Bench — Moonshot's in-house long-horizon agentic benchmark for persistent, multi-day coworking tasks: 17 professional scenarios and 610 evaluation points, executed through the OpenClaw harness (Kimi-K2.7-Code model card).
• MCP Atlas — realistic tool-use tasks served through scalable MCP servers, measuring multi-step tool orchestration (Kimi-K2.7-Code model card).
• MCP Mark Verified — human-verified MCPMark: MCP tool use across five real server environments (Notion, GitHub, Filesystem, Postgres, Playwright) (Kimi-K2.7-Code model card).
• SWE-Bench Verified — human-validated real GitHub issues where the model must produce a patch that passes hidden tests; measures repository-level bug fixing.
• SWE-Bench Pro — harder, contamination-resistant SWE-bench variant with larger, more complex real-world tasks.
• SWE-Bench Multilingual — SWE-bench-style issue resolution across repositories in multiple programming languages.
• Terminal-Bench 2.0 — terminal/command-line agent tasks run under the default Terminus-2 agent framework; measures end-to-end shell task completion (Kimi-K2.6 model card).
• LiveCodeBench (v6) — competitive-programming problems drawn from recent contests to limit training contamination; measures algorithmic code generation.
• BrowseComp — hard web-search/browsing agent benchmark for locating obscure facts online over many tool calls; K2.6 also reports an Agent Swarm variant (Kimi-K2.6 model card).
• HLE-Full — Humanity's Last Exam full set (text and image): extremely hard expert-written questions, reported with and without search/code tools (Kimi-K2.6 model card).
• AIME 2026 / AIME 2025 — American Invitational Mathematics Examination: olympiad-qualifier math problems, averaged over many runs (Kimi-K2.5 model card).
• HMMT 2026 (Feb) / HMMT 2025 (Feb) — Harvard-MIT Mathematics Tournament February contest: harder competition math than AIME.
• GPQA-Diamond — hardest subset of graduate-level, Google-proof science multiple-choice questions in physics, chemistry and biology.
• MMLU-Pro — harder, reasoning-heavier successor to MMLU for broad multiple-choice knowledge.
• MMMU-Pro — massive multi-discipline multimodal understanding, "Pro" version: college-level reasoning over images.
• tau2-bench — customer-service-style tool-use agent benchmark with domain policies and simulated users (not reported by Moonshot for this family).