Kimi K3 on QuickSilver Pro
Kimi K3 is Moonshot's 2.8T-parameter open-weight multimodal reasoning model — suited for complex coding, knowledge work, and long-horizon agentic workflows, and particularly strong at navigating large repositories, using tools, debugging, and iterating against images, logs, tests, and runtime feedback. Its architecture uses KDA and Attention Residuals for computational efficiency. On QuickSilver Pro it's $2.40 input / $12.00 output per 1M tokens, ~20% below OpenRouter's $3.00 / $15.00, with a 1M-token context window.
At a glance
Frontier multimodal reasoning for long-horizon coding and agentic workflows over a 1M-token context.
Pricing comparison ($/1M tokens)
| Provider | Input | Output | vs QSP |
|---|---|---|---|
| QuickSilver Pro | $2.40 | $12.00 | lowest-cost |
| OpenRouter (moonshotai/kimi-k3) | $3.00 | $15.00 | 20% lower |
When to use
Reach for K3 when a task spans a large repository or corpus and needs frontier reasoning across it: complex coding agents that debug against logs, tests, and runtime feedback; multimodal workflows that iterate against images alongside code; and long-horizon agents that plan and coordinate many tool calls over a 1M-token working set. Its KDA + Attention Residuals architecture keeps that long-context reasoning computationally efficient.
When to use something else
For routine chat, short-context codegen, or single-shot tasks the per-token price is overkill — DeepSeek V4 Flash ($0.112/$0.224) or V4 Pro ($0.435/$0.87) land those far lower. For coding-only agent loops where reasoning-token discipline matters most, Kimi K2.7 Code is the leaner sibling to A/B against. For pure mathematical reasoning, DeepSeek V4 Pro.
Quickstart (curl)
curl https://api.quicksilverpro.io/v1/chat/completions \
-H "Authorization: Bearer $QSP_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "kimi-k3",
"messages": [{"role": "user", "content": "Hello!"}]
}'OpenAI-compatible. One-line migration via base_url.
FAQ
K3 is a larger, natively multimodal step up — a 2.8T-parameter open-weight reasoning model with a 1M-token context, versus the K2 line's trillion-parameter text models at 256K. It's tuned for complex coding, knowledge work, and long-horizon agentic workflows, and accepts images as input so it can iterate against screenshots, diagrams, and rendered output alongside code. Its architecture uses KDA and Attention Residuals for computational efficiency at that scale. If your workload is coding-only and cost-sensitive, K2.7 Code's leaner reasoning-token budget may still win — A/B the two on your evals.
Yes — K3 is an OpenAI-compatible chat completions endpoint on QuickSilver Pro. Set base_url=https://api.quicksilverpro.io/v1 and use model="kimi-k3". QSP's current OpenRouter route honors `reasoning: { enabled: false }`; Moonshot's direct API does not, so do not assume that toggle is portable across providers.
OpenRouter lists K3 at $3.00 input / $15.00 output per 1M tokens; QuickSilver Pro is $2.40 / $12.00 — ~20% below on both legs. Same OpenAI-compatible surface; migration is a base_url + key swap, dropping the `moonshotai/` provider prefix from the model ID.
Append the complete assistant message returned by the API on every later turn and tool call; do not keep only `content`, and do not switch into K3 midway through an existing reasoning conversation. Omit sampling overrides unless they match K3's fixed temperature=1.0, top_p=0.95, n=1, and zero penalties. Use top-level `reasoning_effort` with low, high, or max when reasoning is enabled.