Comparisons / Kimi K3 vs Kimi K2.7 Code: Flagship vs Specialist
Kimi K3 vs Kimi K2.7 Code: Flagship vs Specialist
Updated: 2026-09-29
Compared entities
Comparison dimensions
Decision rule
Route coding work to K2.7-Code and reserve K3 for genuinely long inputs: K2.7-Code is $0.95/$4.00 against $3.00/$15.00, a third of the price, and its 262K window covers most repositories. K3 earns its premium above 262K tokens or on general reasoning tasks that are not code. Running everything on K3 is the most common and most expensive misconfiguration in this family.
All figures below are from the China AI Hub database, last verified 2026-09-20 (K2.7 Code last verified 2026-09-27). Where a field is not publicly disclosed, we say so rather than estimating.
At a glance
| Dimension | Kimi K3 | Kimi K2.7 Code |
|---|---|---|
| Status (database) | active | active |
| Context window | 1,048,576 tokens | 262,144 tokens |
| Maximum output | 1,048,576 tokens | Not stated |
| Input price (per 1M) | $3.00 | $0.95 |
| Output price (per 1M) | $15.00 | $4.00 |
| Open weights | Yes (Kimi K3 License) | Not confirmed |
| API available | Yes | Yes |
| Reasoning | Yes | Yes |
| Coding | Yes | Yes |
| Vision / Video | Yes / Yes | Not listed |
| Tool calling | Yes | Not listed |
| Structured output | Yes | Not listed |
| Agent capability | Yes | Not listed |
| Benchmark records in DB | 6 | 0 |
Pricing
Kimi K3 lists $3.00 input / $15.00 output per 1M tokens; Kimi K2.7 Code lists $0.95 input / $4.00 output — roughly a third of the input price and just over a quarter of the output price. A highspeed K2.7 Code variant (same model, higher throughput) is billed at 2x ($1.90 / $8.00). The K2.7 Code tier is the lower-cost coding-specific option in Moonshot’s lineup.
Context and output
Kimi K3 lists a 1,048,576-token context with a full-window 1,048,576-token maximum output. Kimi K2.7 Code lists a 262,144-token context and does not state a maximum-output ceiling in the fetched docs. For very long repository-scale work, K3’s 1M-input/1M-output pairing is the distinguishing listed difference.
Architecture and parameters
Kimi K3 publishes the fullest architecture card in the database: a 2.8T-total / 104B-activated MoE with 93 layers (1 dense), 896 experts (16 selected + 2 shared per token), a 69 KDA + 24 Gated MLA attention split, hidden dim 7168, SiTU-GLU, a 401M-parameter MoonViT-V2 vision encoder, and MXFP4 weights / MXFP8 activations. Moonshot does not disclose architecture or parameter counts for the K2.7 series — the K2.7 Code model is API-only with that information not publicly documented.
Coding
Both list coding. Kimi K3 publishes vendor-reported DeepSWE (67.5) and Terminal-Bench 2.1 (88.3); Kimi K2.7 Code has no benchmark records in the database. The K2.7 Code model is positioned by Moonshot as a dedicated coding model (with always-on thinking and fixed sampling parameters), but its coding performance is not benchmark-documented in the database — the flagship K3 is the benchmark-backed coding model on record.
Vision
Kimi K3 lists vision and video input; Kimi K2.7 Code lists neither. For multimodal coding (e.g., screen-recording-to-code workflows), only K3 lists the capability. Note K3’s modality documentation is itself inconsistent across Moonshot’s sources.
Agent capabilities
Kimi K3 lists agent capability, tool calling and structured output; Kimi K2.7 Code lists none of these in its recorded set (its frontmatter records only reasoning and coding). For a documented tool/agent surface, the database records a clear difference.
License and openness
Kimi K3 is open weight under the Kimi K3 License (MIT-style with revenue-triggered obligations above $20M / 100M MAU). Kimi K2.7 Code has no confirmed open-weight release — the K2.7 series is API-only, and Moonshot does not disclose architecture or parameter counts for it. For self-hosting, only K3 has a documented option.
Deployment
Both are served through the Moonshot AI Platform API. Kimi K3 is self-hostable open weight; Kimi K2.7 Code is API-only with no confirmed open weights. K3’s temperature is fixed at 1.0 and top_p at 0.95; K2.7 Code’s thinking is always on with temperature/top_p/n/penalties fixed and must not be passed.
Benchmarks
The database holds 6 benchmark records for Kimi K3 (GPQA Diamond, HLE-Full, DeepSWE, Terminal-Bench 2.1, MMMU-Pro, Video-MME) and 0 for Kimi K2.7 Code. All K3 rows are vendor-reported.
Benchmark comparability is limited: Kimi K2.7 Code has no benchmark records, so no head-to-head score comparison is possible.
Why each difference matters
China AI Hub analysis indicates the following per-dimension implications, drawn from the listed facts above.
- Pricing: Kimi K2.7 Code has an advantage at roughly a third of K3’s input and just over a quarter of its output price — relevant for high-volume coding at lower cost.
- Context: Kimi K3 has an advantage in long-context work at 1M input with a 1M output ceiling, versus K2.7 Code’s 262,144 input and unstated output.
- Coding: both list coding, but K3 is the benchmark-backed coding model on record (DeepSWE 67.5, Terminal-Bench 2.1 88.3) while K2.7 Code has no benchmark records.
- Vision / media: K3 is more relevant when vision/video input is required — it lists both where K2.7 Code lists neither.
- Agent: K3 is more relevant when a documented tool/agent/structured-output surface is core — it lists these where K2.7 Code lists none.
- License: K3 is open weight (Kimi K3 License); K2.7 Code has no confirmed open-weight release — only K3 offers a self-hosting option.
- Deployment: both API-available; K3 is self-hostable, K2.7 Code is API-only.
- Benchmark: K3 records 6 and K2.7 Code records 0; the evidence base is not comparable.
Trade-off summary
- Price: Kimi K2.7 Code lists roughly a third of K3’s input and a quarter of its output price.
- Context and output: Kimi K3 lists 1M-in/1M-out; K2.7 Code lists 262,144 in and no stated output ceiling.
- Capabilities: Kimi K3 lists vision, video, tool calling, structured output and agent capability; K2.7 Code lists none of these.
- Openness: K3 is open weight; K2.7 Code is API-only.
Choose by workload and budget: cost-sensitive coding favors K2.7 Code’s price; long-context, multimodal, agent-heavy and self-hosting needs favor K3’s listed capabilities. Verify current prices on the official pages before committing.
Decision context
China AI Hub analysis indicates the following decision-context implications, drawn from the listed facts above.
For API developers. Kimi K2.7 Code lists a lower price for a dedicated coding model, but has no benchmark records; Kimi K3 lists benchmarks plus a wider capability surface at a higher price. Both have fixed sampling parameters (K3: temperature 1.0/top_p 0.95; K2.7 Code: always-on thinking with fixed params). Latency is not publicly documented on this page.
For self-hosting. Kimi K3 is open weight (Kimi K3 License, revenue-triggered obligations); Kimi K2.7 Code has no confirmed open-weight release and is API-only. Hardware requirements and quantization are Not publicly documented on this page.
For coding agents. Both list coding. K3 is the benchmark-backed model (DeepSWE 67.5, Terminal-Bench 2.1 88.3) with a 1M output ceiling for long-horizon whole-repo work; K2.7 Code is the lower-cost coding specialist with no benchmark records.
For enterprise. Region, SLA, data-residency and compliance terms are not publicly documented — confirm with the vendor. Deployment differs: K3 offers self-hosting under a conditional license; K2.7 Code is API-only.
What is uncertain
- Kimi K2.7 Code has no benchmark records and no documented maximum-output ceiling.
- Kimi K2.7 Code’s architecture and parameter counts are not publicly disclosed.
- Kimi K3’s input-modality boundary is inconsistent across Moonshot’s official sources.
Sources
Labels used above: Official fact (prices, context windows, capabilities and license terms from primary provider sources), Vendor-reported claim (benchmark scores), and China AI Hub analysis (the “why each difference matters” reasoning, introduced as analysis). No third-party benchmark evidence is currently recorded for these models.