Muse Spark 1.2 on QuickSilver Pro
Muse Spark 1.2 is Meta's coding-focused reasoning model — extended thinking tuned for whole-repository generation, complex debugging, codebase understanding, and long-horizon developer workflows, over a 1M-token context with multimodal input. On QuickSilver Pro it's $1.00 input / $3.40 output per million tokens — ~20% below Meta's published API list price of $1.25 / $4.25 on both legs.
At a glance
Repository-scale coding and long-horizon agentic development with extended reasoning over a 1M-token context.
Pricing comparison ($/1M tokens)
| Provider | Input | Output | vs QSP |
|---|---|---|---|
| QuickSilver Pro | $1.00 | $3.40 | lowest-cost |
| Meta list price (meta/muse-spark-1.2) | $1.25 | $4.25 | 20% lower |
When to use
Reach for Muse Spark 1.2 on coding work that spans a whole repository and rewards deliberate reasoning: end-to-end feature generation, complex multi-file debugging, codebase understanding, and long-horizon agents that plan and iterate across many tool calls. Its 1M-token context holds large working sets, and it accepts images alongside code so an agent can iterate against screenshots, diagrams, and rendered output.
When to use something else
It reasons at length and is verbose, so for routine chat, short single-shot codegen, or latency-sensitive loops the token budget is overkill — DeepSeek V4 Flash ($0.112/$0.224) or Qwen3.7 Flash land those far lower. Reasoning is mandatory here, so it spends output tokens even on trivial turns; rein that in with `reasoning_effort: "low"` where cost matters (a small `max_tokens` only truncates the answer, not the thinking). Weights are closed and API-only; if you need self-hosting, choose an open-weight model such as Kimi K3.
Quickstart (curl)
curl https://api.quicksilverpro.io/v1/chat/completions \
-H "Authorization: Bearer $QSP_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "muse-spark-1.2",
"messages": [{"role": "user", "content": "Hello!"}]
}'OpenAI-compatible. One-line migration via base_url.
FAQ
Coding. Meta trained it on long-horizon coding tasks — whole-repository generation, large end-to-end projects, complex debugging, and codebase understanding — and pairs it with Muse Code, its coding agent. On Meta's own launch charts it sits just behind Claude Opus 5 on Terminal-Bench 2.1 and DeepSWE. It accepts text and images as input and returns text, over a 1M-token context.
Reasoning is always on — it is mandatory on this endpoint and cannot be disabled (a request that tries to turn it off is rejected). Use `reasoning_effort` (`low`, `high`, or `max`) to trade depth against latency and cost; `low` roughly halves reasoning-token spend on simple turns. Reasoning tokens are drawn from your `max_tokens` budget before the visible answer, so set `max_tokens` generously — 1024 or more — or the reply can come back empty because the budget was spent thinking.
Meta's published API list price is $1.25 input / $4.25 output per 1M tokens; QuickSilver Pro is $1.00 / $3.40 — ~20% below on both legs. Same OpenAI-compatible surface: point base_url at https://api.quicksilverpro.io/v1 and use model="muse-spark-1.2".
Yes — it's an OpenAI-compatible chat completions endpoint, with the Responses API available too. Set base_url=https://api.quicksilverpro.io/v1 and model="muse-spark-1.2", and point coding agents (the same ones you'd run Muse Code-style workflows with) at that base URL.