Muse Spark 1.3 Limits

Muse Spark 1.3 limits center on a 1-million-token context window and partner-only access gating.

At Layer3Labs, we help engineering teams evaluate model constraints and integrate coding agents into production workflows without unexpected throttling.

Meta released the model through Meta Superintelligence Labs (MSL) on 2026-09-02. Specific API (Application Programming Interface) quotas and output caps remain unlisted.

Context Window Capacity and Long-Context Retrieval

Muse Spark 1.3 maintains a 1-million-token context window.

This matches the raw capacity established by earlier releases in the product line.

A 1M-token window allows the model to ingest multi-file codebases, technical documentation, and dependency graphs in a single prompt.

Meta reported a 98.5% score on long-context retrieval benchmarks.

High recall accuracy prevents critical configuration flags from dropping out of active context during large audits.

Teams can verify retrieval testing in the Meta AI Research announcement.

Our Muse Spark 1.3 explained guide provides a complete architectural overview.

• Total context window: 1 million tokens (1M)
• Long-context retrieval benchmark: 98.5%
• Architectural context baseline: Identical to Muse Spark 1.1 and Muse Spark 1.2

Variant Gating Between Max and Xhigh Tiers

Meta divides Muse Spark 1.3 into two tiers with distinct availability limits.

The accessible production model is Muse Spark 1.3 (xhigh).

In contrast, Muse Spark 1.3 (max) remains restricted to a limited preview for select Meta partners.

Meta gated the extended reasoning mode in the max variant behind ongoing safety testing.

Weights remain closed.

Keeping weights closed prevents self-hosting on private servers and confines all workloads to remote infrastructure.

External benchmarks show meaningful variance between tiers.

Bloomberg reported that the max variant scored 62 on the Artificial Analysis Intelligence Index, placing it just behind Claude Fable 5.1 and Claude Opus 5.

The release page on Artificial Analysis reports lower individual scores for accessible variants.

Because reported figures diverge across industry reports, operators should validate performance directly on internal codebases rather than treating one index as definitive.

• Muse Spark 1.3 (xhigh): Broadly accessible production tier
• Muse Spark 1.3 (max): Partner-only limited preview with gated reasoning
• Model weights: Closed architecture with no self-hosted option

Unpublished API Rate Limits and Output Caps

Meta has not published per-minute token limits or output caps at launch.

Quotas depend on individual account standing within the Meta Model API and Muse Code.

Developers should check their account dashboard on the Meta Muse Code product page to verify active limits.

Efficiency gains partly offset these unannounced caps.

Meta engineers recorded approximately 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 on identical developer workflows.

Lower token demand helps developers run automated agent loops without exhausting concurrency allowances.

Published benchmark scores reflect these coding capabilities: 75.4% on DeepSWE 1.1 for Software Engineering (SWE) tasks, 88.8% on Terminal-Bench 2.1, and 59.4% on SWEAtlas CodeBase Question and Answer (QnA).

Our Muse Spark 1.2 coding guide examines how earlier execution limits affected multi-file refactoring.

• Requests per minute: Unpublished at launch
• Maximum output tokens: Unpublished at launch
• Tool call reduction: ~20% fewer tool calls than Muse Spark 1.2
• Token efficiency: ~25% fewer tokens required per task

Cost Considerations and Billing Boundaries

Meta has not published a finalized per-token pricing schedule for Muse Spark 1.3.

Estimates on Artificial Analysis suggest a blended rate near $0.80 per million tokens.

The prior generation used a two-tier pricing structure.

Muse Spark 1.2 charged $1.25 per million input tokens, $0.15 for cached inputs, and $4.25 for outputs on its Standard tier.

Its Contributor tier dropped prices to $0.10 input and $0.20 output in exchange for data-training rights.

Our Muse Spark 1.2 pricing analysis breaks down those baseline figures.

While Meta may adopt a comparable structure for 1.3, teams must confirm active billing rates on the Meta Muse Code product page before budgeting production workloads.

• Published 1.3 rate card: No official input or output token split at launch
• Third-party blended estimate: ~$0.80 per million tokens
• Predecessor baseline: Standard tier ($1.25/$4.25) and Contributor tier ($0.10/$0.20)

Production Planning Around Muse Spark 1.3 Limits

Planning around Muse Spark 1.3 limits requires building defenses against unstated concurrency ceilings.

Engineering groups with strict air-gapped security requirements should not adopt Muse Spark 1.3, because Meta maintains closed weights with no on-premise hosting.

Those organizations can review alternative architectures covered in our Muse Glimmer guide.

Our recommendation would change if Meta provides open weights or introduces binding enterprise SLAs (Service Level Agreements) for dedicated throughput.

To protect active workflows from unexpected rate throttles, engineering leads should introduce local linters, prompt caching, and request queuing.

Check active quota thresholds in your developer dashboard and run small-batch integration tests to manage Muse Spark 1.3 limits before expanding your pipeline.

• Who this is not for: Regulated teams requiring air-gapped hosting or open weights
• What would change our answer: Enterprise concurrency SLAs or open-weight releases
• Recommended mitigation: Local syntax parsing, aggressive prompt caching, and batched CLI agent calls

How to use Muse Spark 1.3

A hosted model runs on the provider's servers, so using it is really about the tool you access it through.

The fastest way to put Muse Spark 1.3 to work day to day is inside an AI IDE, and Cursor is the most popular — it supports every major model, so you can be working in minutes. Each major maker also ships a first-party tool — Claude Code, Codex, or Antigravity — worth trying for the native experience. Prefer a different editor? Windsurf, Zed, and GitHub Copilot drive these models too.