Kimi K2 Thinking
Moonshot AIOpen weightsOpen agentic reasoner shipped with int4 quantization-aware training — the official serving IS int4. Collapses hardest when you cut effort: a cliff-curve model.
Index @ High
75.3
Context
256K
$/M in · out
$0.60 · $2.50
Median speed
50 tok/s
Released
Nov 6, 2025
02·Which effort for which task
The sweet spot, per task
Cheapest effort tier that keeps ≥96% of this model's peak quality on the benchmarks behind each task type. Below the floor number, quality is not okay — it just fails quietly.
Everyday writing & Q&A
MediumChat, drafting, summaries, general questions.
Code assist
HighAutocomplete, single functions, small fixes.
Coding agents
HighReal repos, multi-file changes, end-to-end tickets.
⚠ at Low: only 72% of peak — avoid
Math & quantitative
HighDerivations, statistics, financial models.
⚠ at Low: only 58% of peak — avoid
Hard research problems
HighNovel, frontier, genuinely difficult questions.
⚠ at Low: only 68% of peak — avoid
Agents & tool use
HighLong workflows with tools, policies, and state.
03·The effort ladder
Quality and cost, tier by tier
On everyday tasks, Medium keeps 97.0% of High quality at $0.011 vs $0.032 per task.
Intelligence Index by effort
Cost per typical task by effort
04·Benchmark detail
| Benchmark | Low | Peak | What it predicts |
|---|---|---|---|
| MMLU-Pro | 76.8 | 84.0 | Everyday Q&A · Drafting emails & docs · Summarizing articles |
| GPQA Diamond | 57.2 | 84.0 | Technical analysis · Root-cause debugging · Scientific writing |
| Humanity's Last Exam | 15.6 | 23.0 | Novel research questions · Frontier problem solving |
| AIME 2025 | 42.6 | 89.0 | Financial modeling · Statistics · Algorithm design |
| SWE-bench Verified | 52.4 | 71.0 | Coding agents · Bug fixing in real repos · Multi-file refactors |
| LiveCodeBench | 56.9 | 77.0 | Writing functions from scratch · Algorithms · Code autocomplete |
| τ²-bench | 56.1 | 80.0 | Tool-using agents · Customer-facing automation · Multi-step workflows |
| IFEval | 81.3 | 87.0 | Structured output (JSON) · Templated generation · Batch pipelines |
05·Host variants
Same weights, different model
Hosts serve Kimi K2 Thinking at different quantizations and stacks. Index below is measured per host at High effort — the headline number only applies to the reference serving.
| Host | Quant | Index | Δ vs reference | $/M in · out | tok/s | TTFT | Uptime | Gateway id |
|---|---|---|---|---|---|---|---|---|
| Moonshot (official)reference | int4-qat | 75.3 | — | $0.60 · $2.50 | 50 | 1.10s | 99.0% | kimi-k2-thinking |
| Together AI | bf16 | 75.1 | -0.2 | $1 · $3 | 90 | 0.50s | 98.8% | kimi-k2-thinking@together |
| Groq | int8 | 72.6 | -2.6 | $1 · $3 | 720 | 0.25s | 98.5% | kimi-k2-thinking@groq |
06·Drift
Is it still the same model this week?
Weekly Intelligence Index · dashed rules mark serving-change events
Speculative-decoding aggressiveness increased for throughput; quality dip still in effect.
07·Call it
One key, this model, your effort
The gateway speaks the OpenAI Chat Completions dialect. Pick a host with kimi-k2-thinking@host syntax, set reasoning_effort, and the sweet spot above becomes one line of config. Full docs →
curl https://models.fyi/api/v1/chat/completions \
-H "Authorization: Bearer $MFYI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "kimi-k2-thinking",
"reasoning_effort": "medium",
"messages": [{"role": "user", "content": "Hello"}]
}'
# local dev: replace https://models.fyi with http://localhost:3000