Intelligence
30.9
31
General reasoning and benchmark headroom.
LimitedMoonshotAI: Kimi K2 0905 is a budget text-first model from moonshotai with a fast runtime profile, extended context posture, and the clearest fit around long-context research / agent workflows.
Benchmark blend
Dev workflow signal
Extended
Budget tier
MoonshotAI: Kimi K2 0905 currently reads as a budget text-first option with extended context and a fast runtime profile.
Decision Strip
Core buy-side signals stay in one pass. The rest of the page expands only after intelligence, speed, context, and price are clear.
General reasoning and benchmark headroom.
LimitedTTFT 0.52s
CompetitiveHow much prompt and task state can stay in view.
Competitive$2.50 output / 1M
EfficientEditorial Profile
Positioning, tradeoffs, and fit are consolidated into one read instead of repeating the same story across separate cards.
Kimi K2 0905 is the September update of [Kimi K2 0711](moonshotai/kimi-k2). It is a large-scale Mixture-of-Experts (MoE) language model developed by Moonshot AI, featuring 1 trillion total parameters with 32 billion active per forward pass. It supports long-context inference up to 256k tokens, extended from the previous 128k. This update improves agentic coding with higher accuracy and better generalization across scaffolds, and enhances frontend coding with more aesthetic and functional outputs for web, 3D, and related tasks. Kimi K2 is optimized for agentic capabilities, including advanced tool use, reasoning, and code synthesis. It excels across coding (LiveCodeBench, SWE-bench), reasoning (ZebraLogic, GPQA), and tool-use (Tau2, AceBench) benchmarks. The model is trained with a novel stack incorporating the MuonClip optimizer for stable large-scale MoE training.
moonshotai text-first profile
Long-context research / Agent workflows with extended context and fast runtime.
Efficient spend profile. More comfortable for sustained prompt volume if the capability fit is right.
Large context headroom supports repo-wide prompts and long research sessions.
Latency and throughput look responsive enough for interactive loops.
Budget-friendly input pricing is a strength, but raw capability may vary by workload.
Current metadata points to a text-first profile rather than a broad multimodal one.
Long-context summarization, repo analysis, and policy or document review.
Benchmarks
Only benchmark categories with actual signal are shown. Secondary values stay as simple definitions instead of nested micro-cards.
Broad reasoning, knowledge depth, and flagship benchmark posture.
Software implementation, debugging quality, and coding benchmark signal.
Formal reasoning, structured problem solving, and competition-style math.
Long-horizon execution quality and interactive benchmark evidence.
Specs & Pricing
Specs stay neutral, pricing gets emphasis through values rather than extra containers. Raw provider internals remain in metadata at the end.
This model is relatively efficient on price. It is the easier fit when sustained prompt volume matters.
Metadata
Verification details remain available, but the page no longer forces them ahead of the editorial read.