Kimi K2.7 Code vs Kimi K3
Different efforts: thinking · max
Humanity's Last Exam Humanity's Last Exam: Very hard expert questions across many fields, written to stump today’s models
Kimi K2.7 Code35
Kimi K346.9
Terminal-Bench 4.0 Terminal-Bench 4.0 · source setting 2: Agents use a terminal to complete executable coding, system administration and research tasks in the 4.0 suite
Kimi K2.7 Code1
Kimi K312.6
AutomationBench-AA AutomationBench-AA: Agents orchestrate SaaS workflows through APIs; objectives receive credit only when the task has no guardrail violation
Kimi K2.7 Code24.5
Kimi K358.3
Arena Text · creative writing Arena Text · creative writing: Human pairwise preferences in the stated text vote pool
Kimi K2.7 Code—
Kimi K31,461
MMMU-Pro (vision) MMMU-Pro (vision): College-level questions that require understanding images: charts, diagrams, photos
Kimi K2.7 Code—
Kimi K380.5
| Benchmark | Kimi K2.7 Code | Kimi K3 |
|---|---|---|
| Humanity's Last Exam | 35 | 46.9 |
| Terminal-Bench 4.0 | 1 | 12.6 |
| AutomationBench-AA | 24.5 | 58.3 |
| Arena Text · creative writing | — | 1,461 |
| MMMU-Pro (vision) | — | 80.5 |
Specs
| Spec | Kimi K2.7 Code | Kimi K3 |
|---|---|---|
| Input, per 1M | $0.95 | $3 |
| Output, per 1M | $4 | $15 |
| Cost per task? | $0.54 | $2.00 |
| Speed, tok/s | 63 | 112 |
| Context | 256K | 1M |
| Weights | Open | Open |
| Released | Jun 2026 | Jul 2026 |
| Memory (4-bit) | 616 GB | 1,668 GB |
| Licence | modified-mit | kimi-k3 |
| Commercial use | Restricted | Restricted |
| See images | ✓ | ✓ |
| Read PDFs | Partial | Partial |
| Hear audio | — | — |
| Watch video | — | ✓ |
| Call tools | ✓ | ✓ |
| Force a tool call | Partial | ✓ |
| Return strict JSON | ✓ | ✓ |
| Search the web | ? | Partial |
| Run code | ? | ? |
| Use a computer | ? | ? |
What breaks when you switch
Kimi K2.7 Code → Kimi K3
Memory (4-bit): 616 GB → 1,668 GB
Sampling parameters are fixed›
temperature, top_p, n and penalty values are fixed; omit them from requests.
No public image URLs›
Vision input rejects public image URLs; use base64 or ms:// file references in a content array.
Web search is being updated›
Web search is not recommended for near-term production workflows; avoid depending on it in production.
Thinking mode is always enabled›
K3 always has thinking enabled; control effort with the top-level reasoning_effort field using low, high or max.
Kimi K3 → Kimi K2.7 Code
Context: 1M → 256K
Instant mode is not supported · Self-hosted / vLLM / SGLang›
The documented third-party vLLM and SGLang deployments do not support Instant mode.