Skip to content
Bud's Wiki

Kimi K2.6: Advancing Open-Source Coding

TL;DR

Section titled “TL;DR”

Moonshot ships Kimi K2.6, the successor to K2.5, foregrounding long-horizon coding (4,000+ tool calls across 12+ hours per run), a 3× scale-up of the Agent Swarm (300 sub-agents × 4,000 coordinated steps vs K2.5’s 100 × 1,500), and a new heterogeneous-collaboration mode called Claw Groups. Open weights ship on Hugging Face alongside the chat/agent/code product surfaces. Headline numbers: SWE-Bench Pro 58.6, SWE-Bench Verified 80.2, Terminal-Bench 2.0 (Terminus-2) 66.7, HLE-with-tools 54.0, BrowseComp 83.2 (86.3 with Agent Swarm), Toolathlon 50.0, OSWorld-Verified 73.1 — placing K2.6 inside the GPT-5.4 / Claude Opus 4.6 / Gemini 3.1 Pro frontier band on agentic and coding axes while remaining open-source. Two case studies anchor the long-horizon-coding claim: deploying Qwen3.5-0.8B on a Mac in Zig over 4,000+ tool calls / 14 iterations (15 → 193 tok/s, ~20% faster than LM Studio), and overhauling the 8-year-old exchange-core matching engine over 13 hours / 1,000+ tool calls / 4,000+ lines, extracting a 185% median throughput gain.

Key claims

Section titled “Key claims”
  • Kimi K2.6 reports SWE-Bench Pro 58.6, SWE-Bench Verified 80.2, SWE-Bench Multilingual 76.7, and Terminal-Bench 2.0 (Terminus-2) 66.7 — placing it inside the GPT-5.4 xhigh / Claude Opus 4.6 max-effort / Gemini 3.1 Pro thinking-high band on the open-MoE agentic-SWE frontier (K2.5 baseline: 50.7 / 76.8 / 73.0 / 50.8) [Benchmark table].
  • On the agentic search axis, K2.6 reports BrowseComp 83.2 (single-agent) → 86.3 (Agent Swarm) — vs 82.7 GPT-5.4, 83.7 Claude Opus 4.6, 85.9 Gemini 3.1 Pro; K2.5 baseline was 60.6 → 78.4 — i.e. an 8.5-point Agent-Swarm step-up at the K2.5→K2.6 boundary [Benchmark table].
  • Long-horizon coding: a single K2.6 run autonomously deployed Qwen3.5-0.8B locally on a Mac in Zig (a deliberately out-of-distribution language) over 4,000+ tool calls, 12+ hours, and 14 iterations, improving inference throughput from ~15 tok/s to ~193 tok/s — claimed ~20% faster than LM Studio [§Long-horizon coding].
  • A second long-horizon case study overhauled exchange-core, an 8-year-old open-source financial matching engine, over 13 hours / 1,000+ tool calls / 4,000+ lines of code; K2.6 used CPU and allocation flame graphs to reconfigure the core thread topology (4ME+2RE → 2ME+1RE) and extracted a 185% median throughput leap (0.43 → 1.24 MT/s) and 133% performance throughput gain (1.23 → 2.86 MT/s) [§Long-horizon coding].
  • Agent Swarm scales horizontally to 300 sub-agents executing across 4,000 coordinated steps per run — up from K2.5’s 100 × 1,500 — described as a “qualitative leap” from the prior swarm preview [§Agent Swarm].
  • HLE-Full-with-tools 54.0 vs GPT-5.4 xhigh 52.1 / Claude Opus 4.6 53.0 / Gemini 3.1 Pro 51.4 — K2.6 leads the cohort on tool-augmented reasoning [Benchmark table].
  • Toolathlon 50.0 vs GPT-5.4 54.6 / Claude Opus 4.6 47.2 / Gemini 3.1 Pro 48.8 — K2.6 is best-of-open on Toolathlon and second only to GPT-5.4 across the cohort [Benchmark table].
  • OSWorld-Verified 73.1 vs Claude Opus 4.6 72.7 / GPT-5.4 75.0 — narrows the computer-use gap that K2.5 already opened (K2.5 had reported 63.3 OSWorld-Verified vs GPT-5.2’s 8.6 / Gemini 3 Pro’s 20.7), and is now comparable to the closed frontier [Benchmark table].
  • Vision benchmarks show K2.6 substantially behind GPT-5.4 / Gemini 3.1 Pro on raw MMMU-Pro and MathVision, but the “with python” variants close most of the gap — CharXiv-RQ-with-python 86.7 vs 90.0 / 89.9; MathVision-with-python 93.2 vs 96.1 / 95.7; V*-with-python 96.9, matching Gemini 3.1 Pro [Benchmark table].
  • Reasoning-only HLE-Full (no tools) is 34.7 vs GPT-5.4 39.8 / Gemini 3.1 Pro 44.4 — i.e. K2.6’s tool-using delta (HLE-w/tools − HLE-no-tools) is +19.3 points, similar in size to the +20.1 K2.5 had reported and larger than any closed-cohort entry [Benchmark table].
  • Proactive Agents — autonomous background agents like OpenClaw — are powered by K2.6 for 24/7 ops; the post reports an internal RL-infra agent running autonomously for 5 days managing monitoring, incident response, and system operations [§Proactive Agents].
  • Claw Groups is introduced as a research-preview heterogeneous-collaboration mode: a new instantiation of the Agent Swarm where multiple agents (any device, any model, with their own toolkits and persistent memory) and humans operate as collaborators, with K2.6 acting as an “adaptive coordinator” that matches tasks to agents, detects failures, and re-routes [§Claw Groups].
  • “Skills” let K2.6 ingest PDFs / spreadsheets / slides / Word docs and reproduce their structural and stylistic DNA in future tasks — i.e. document-as-style-primitive rather than document-as-context [§Agent Swarm].
  • Released across a multi-surface launch package: model weights at huggingface.co/moonshotai/Kimi-K2.6, API at platform.moonshot.ai, web app at kimi.com (chat mode + agent mode), and Kimi Code at kimi.com/code for production-grade coding workflows [§Overview, launch tweet].

Method

Section titled “Method”

K2.6 is positioned by the blog post as a coding-and-agentic capability scale-up on top of the K2 / K2.5 lineage; no architectural or training-recipe details are disclosed in the public post, so the methodology section is narrower than for the K2.5 release. What the post does describe is the capability surface and the deployment package.

Agent Swarm scale-up. The orchestrator-trains-sub-agents framing carried over from K2.5’s PARL (Kimi K2.5: Visual Agentic Intelligence) but scaled — 300 concurrent sub-agents across 4,000 coordinated steps per run, vs K2.5’s 100 × 1,500. The post characterizes the swarm as “seamlessly coordinat[ing] heterogeneous agents to combine complementary skills” and reports end-to-end outputs spanning documents, websites, slides, and spreadsheets within a single autonomous run. No new training-method names are introduced — the inference-time orchestrator dispatch graph is the same shape PARL trained, just larger.

Long-horizon coding capability. Two concrete case studies are reported in detail: (a) 4,000+ tool calls / 12+ hours / 14 iterations for the Mac-Zig-Qwen3.5 deployment task; (b) 1,000+ tool calls / 13 hours / 4,000+ lines edited for the exchange-core matching-engine overhaul, including CPU/allocation flame-graph analysis as part of the agent’s optimization loop. The post does not disclose how this was trained — whether via longer-horizon trajectories during RL, longer context, a new tool-call format, or a new reward shape.

Skills. K2.6 can convert high-quality input documents (PDFs, spreadsheets, slides, Word docs) into reusable “Skills” — internal style-and-structure templates the swarm draws on when producing later deliverables. This is a document-as-prior pattern rather than document-as-context-window.

Claw Groups. A research preview where multiple agents and humans operate as peers in a shared workspace, with K2.6 as the adaptive coordinator. Agents can run on any device, any model, with their own toolkits and persistent memory. K2.6’s role is task-to-agent matching, failure detection, automatic re-routing, and lifecycle management — extending the Agent Swarm pattern from one-orchestrator-with-Kimi-spawned-children to one-orchestrator-over-arbitrary-third-party-agents.

Eval suites. The post references three internal benchmarks alongside the public ones: Kimi Code Bench (end-to-end coding), Kimi Design Bench (visual input / landing pages / full-stack / creative — compared head-to-head with Google AI Studio with K2.6 “performing well across these categories”), and Claw Bench (five domains: coding, IM ecosystem integration, information research & analysis, scheduled task management, memory utilization), reported as where K2.6 “significantly outperforms K2.5 in task completion rates and tool invocation accuracy.”

Results

Section titled “Results”

Headline benchmark table (with the strongest cohort comparisons). Columns are K2.6 / GPT-5.4 xhigh / Claude Opus 4.6 max-effort / Gemini 3.1 Pro thinking-high / K2.5:

  • Agentic / search / tools: HLE-Full-w/tools 54.0 / 52.1 / 53.0 / 51.4 / 50.2; BrowseComp 83.2 / 82.7 / 83.7 / 85.9 / 74.9; BrowseComp(Agent Swarm) 86.3 / — / — / — / 78.4; DeepSearchQA-f1 92.5 / 78.6 / 91.3 / 81.9 / 89.0; DeepSearchQA-accuracy 83.0 / 63.7 / 80.6 / 60.2 / 77.1; WideSearch-item-f1 80.8 / — / — / — / 72.7; Toolathlon 50.0 / 54.6 / 47.2 / 48.8 / 27.8; MCPMark 55.9 / 62.5* / 56.7* / 55.9* / 29.5; Claw Eval pass^3 62.3 / 60.3 / 70.4 / 57.8 / 52.3; APEX-Agents 27.9 / 33.3 / 33.0 / 32.0 / 11.5; OSWorld-Verified 73.1 / 75.0 / 72.7 / — / 63.3.
  • Coding: Terminal-Bench 2.0 Terminus-2 66.7 / 65.4* / 65.4 / 68.5 / 50.8; SWE-Bench Pro 58.6 / 57.7 / 53.4 / 54.2 / 50.7; SWE-Bench Multilingual 76.7 / — / 77.8 / 76.9* / 73.0; SWE-Bench Verified 80.2 / — / 80.8 / 80.6 / 76.8; SciCode 52.2 / 56.6 / 51.9 / 58.9 / 48.7; OJBench-python 60.6 / — / 60.3 / 70.7 / 54.7; LiveCodeBench v6 89.6 / — / 88.8 / 91.7 / 85.0.
  • Reasoning & knowledge: HLE-Full (no tools) 34.7 / 39.8 / 40.0 / 44.4 / 30.1; AIME 2026 96.4 / 99.2 / 96.7 / 98.3 / 95.8; HMMT 2026 (Feb) 92.7 / 97.7 / 96.2 / 94.7 / 87.1; IMO-AnswerBench 86.0 / 91.4 / 75.3 / 91.0* / 81.8; GPQA-Diamond 90.5 / 92.8 / 91.3 / 94.3 / 87.6.
  • Vision: MMMU-Pro 79.4 / 81.2 / 73.9 / 83.0* / 78.5; CharXiv-RQ 80.4 / 82.8* / 69.1 / 80.2* / 77.5; CharXiv-RQ-w/python 86.7 / 90.0* / 84.7 / 89.9* / 78.7; MathVision 87.4 / 92.0* / 71.2* / 89.8* / 84.2; MathVision-w/python 93.2 / 96.1* / 84.6* / 95.7* / 85.0; BabyVision 39.8 / 49.7 / 14.8 / 51.6 / 36.5; BabyVision-w/python 68.5 / 80.2* / 38.4* / 68.3* / 40.5; V*-w/python 96.9 / 98.4* / 86.4* / 96.9* / 86.9.

Headline absolute results worth flagging: K2.6 leads the cohort on HLE-with-tools, DeepSearchQA (both f1 and accuracy), WideSearch, and SWE-Bench Pro; matches Gemini 3.1 Pro on V*-with-python; closes the OSWorld-Verified gap to the closed cohort; and is second only to GPT-5.4 on Toolathlon. K2.5 → K2.6 deltas are largest on Toolathlon (+22.2 absolute), MCPMark (+26.4), APEX-Agents (+16.4), Terminal-Bench 2.0 (+15.9), OSWorld-Verified (+9.8), and BrowseComp (+22.6 single-agent / +7.9 Agent Swarm).

Why it’s interesting

Section titled “Why it’s interesting”

K2.6 is the direct successor to Kimi K2.5: Visual Agentic Intelligence — same Agent Swarm framing, same multi-surface release package, scaled-up swarm capacity (300 × 4,000 vs 100 × 1,500), and a deliberate doubling-down on long-horizon coding. For the Agentic Software Engineering frontier this page tracks, K2.6 lands as the new highest-reported open-cohort number on SWE-Bench Pro (58.6, ahead of Qwen3-Coder-Next Technical Report‘s 44.3 and Step 3.5 Flash‘s peer-table entries) and is the first open release to claim parity-or-better with GPT-5.4 / Claude Opus 4.6 / Gemini 3.1 Pro across the full SWE-Bench-Pro + Terminal-Bench 2.0 + OSWorld + HLE-with-tools quadruple. For Inference-Time Scaling, the 3× swarm scale-up (100→300 sub-agents, 1,500→4,000 coordinated steps) is the largest reported datapoint of learned parallel-agent orchestration (the fourth axis K2.5’s PARL named), making it the natural extension of the K2.5 contribution to that page. The lack of a training-recipe disclosure means K2.6 is filed primarily as a capability and benchmark datapoint; the open weights at huggingface.co/moonshotai/Kimi-K2.6 will be the surface where the recipe ultimately gets reverse-engineered.

See also

Section titled “See also”