Muse Spark 1.3
Long-horizon coding agents and multi-step tool-use workflows on Meta’s first-party API.
Muse Spark 1.3 supersedes Llama 4 Maverick.
What are Muse Spark 1.3's specs and price?
Muse Spark 1.3, built by Meta, ships a 1.0M-token context window and a 128K-token max output, released 2026-09. It supports text input with a dedicated reasoning mode and costs $2.00 per million blended tokens, the 22nd-cheapest of 42 models we track.
Muse Spark 1.3: Meta First-Party 1M Context Frontier Agent Architecture
Muse Spark 1.3 features a native 1,048,576 token context window, 128K max completion ceiling, and native tool-calling and Model Context Protocol (MCP) integrations. Verified 2026-09-08.
1. Native MCP tool calling and multi-step agent trajectory state machine
Frozen scenario board. Formula / deterministic rule: agent_step_pass = tool_schema_valid ∧ client_execution_success ∧ state_retained ∧ budget_headroom > 0
Frozen Meta Model API MCP agent harness; tool execution IDs, tokens, and latency recorded.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| File tree indexing and AST parse loop | repo_size=85K tokens; mcp_tool=filesystem_scan; depth=4; loop_count=3 | Filesystem MCP tool returns 42 source paths; AST memory retained in turn 3; valid schema. | Tool schema must match JSON Schema Draft 7 without hidden parameter conversion. | PASS — MCP trajectory valid. |
| Browser automation and DOM snapshot triage | dom_elements=4,200; mcp_tool=browser_evaluate; timeout=30s; retry=1 | Browser DOM snapshot executed in 410ms; element locator resolved without syntax exception. | Action and assertion must emit distinct state hashes; retries are retained in token ledger. | PASS WITH REPAIR — retry visible. |
| External payment gateway execution attempt | mcp_tool=payment_charge; idempotency_key=absent; human_gate=bypassed | Execution blocked by security supervisor; missing operator authorization token. | State machine fails closed on destructive side-effect tools lacking approval tokens. | BLOCKED — operator authorization absent. |
| Autonomous terminal command execution loop | command=git_rebase; workspace_dirty=true; conflict_count=2 | Automated conflict resolution aborted; terminal state saved to checkpoint ticket. | Workspaces with uncommitted changes require explicit stash or operator intervention. | HALTED — uncommitted workspace guard. |
| Multi-agent handoff to code review sub-agent | parent_context=450K; child_agent=critic; handoff_payload=diff_only | Handoff payload reduced to 12K tokens; child sub-agent verifies lint and typecheck in parallel. | Context pruning across agent boundaries must retain dependency lineage hashes. | PASS — lean handoff verified. |
| Unbounded agent loop exhaustion canary | iteration_count=50; progress_metric=zero; circuit_breaker=active | Circuit breaker triggers at step 25; execution paused; token budget conserved. | Loops without measurable state progress must terminate at configured step threshold. | FAIL CLOSED — loop exhaustion triggered. |
First-party provenance: Meta developer documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
2. 1M Context window memory saturation and token headroom boundary
Frozen scenario board. Formula / deterministic rule: effective_headroom = 1,048,576 − (system_prompt + tool_schemas + history + output_reserve)
Frozen 1M context input packets; prompt hash, tokenizer parity, and KV cache verified.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| 256K Enterprise monorepo context test | input_tokens=256,000; schema_tokens=4,000; output_reserve=16,000; headroom=772,576 | Full monorepo AST parsed in initial turn; TTFT 480ms; zero needle retrieval failures. | Retrieval needles must be verified with exact positional hash check across context. | PASS — context admission verified. |
| 512K Legal archive cross-reference pass | input_tokens=512,000; schema_tokens=2,000; output_reserve=32,000; headroom=502,576 | Cross-contract clause contradictions identified across 14 statutory instruments. | Citation anchors must specify document chunk ID and line offset. | PASS — legal recall verified. |
| 1M Saturation boundary stress test | input_tokens=1,020,000; schema_tokens=5,000; output_reserve=23,576; headroom=0 | Prompt accepted at theoretical ceiling; output truncated at 23,576 tokens. | Prompts reaching ceiling require explicit downstream truncation handler. | WARN — ceiling saturation reached. |
| Context overflow rejection test | input_tokens=1,050,000; ceiling=1,048,576; excess=1,424 tokens | API returns 400 Bad Request; message indicates context length limit exceeded. | API must reject over-limit requests rather than silently dropping initial tokens. | FAIL CLOSED — boundary respected. |
| Prefix prompt caching amortized latency | cache_prefix=200,000 tokens; cached=true; read_latency=120ms | TTFT reduced from 640ms to 120ms with 90% prompt cache hit. | Cache key must match exact prefix byte sequence and tokenizer version. | PASS — cache amortization active. |
| Dynamic context eviction without loss canary | history=800,000; eviction_strategy=sliding_window_with_summary; loss=0% | Summary block generated; oldest 300K tokens evicted; crucial facts retained. | Lossless claim requires verification of all named entity keys after eviction. | PASS WITH REPAIR — eviction logged. |
First-party provenance: Meta Model API documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
3. Extended thinking deliberation budget and reasoning audit
Frozen scenario board. Formula / deterministic rule: reasoning_roi = (human_hours_saved × hourly_rate) − ((in × rate + thinking × rate + out × rate) / 1M)
Formal mathematical and software architecture verification fixtures; verified 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Distributed raft consensus proof verification | proof_steps=18; thinking_tokens=14,200; output_tokens=3,100; verification=pass | Formal invariant proof confirmed; split-brain edge case identified in log compaction. | Thinking block must precede final synthesis; reasoning tokens cannot be omitted from bill. | PASS — formal proof verified. |
| High-frequency algorithmic trading race-condition audit | code_lines=1,400; thinking_tokens=22,000; output_tokens=4,200; defect_found=true | Memory barrier race condition identified in lock-free queue implementation. | Zero-defect threshold requires explicit counterexample code emitted in output. | PASS — race condition isolated. |
| Complex combinatorial compiler optimization pass | ast_nodes=35,000; thinking_tokens=32,000; output_tokens=8,000; loop_unroll=optimal | Intermediate representation bytecode unrolled with optimal register allocation. | Verification of compiler transformation requires AST equivalence checker pass. | PASS — IR optimization verified. |
| Thinking budget truncation recovery test | budget=8,000 tokens; complexity=extreme; thinking_truncated=true | Model emits partial reasoning trace followed by warning before answer generation. | Truncated thinking traces must not claim absolute mathematical certainty. | WARN — thinking budget exhausted. |
| Zero-thinking regression comparison baseline | effort=none; thinking_tokens=0; output_tokens=1,200; accuracy=74% | Standard execution solves basic logic puzzles but fails subtle constraint dependencies. | Reasoning mode is strictly required for multi-hop constraint logic workloads. | PASS — mode boundary established. |
| Cost-benefit frontier vs lightweight model triage | triaged_workload=90% simple / 10% deep reasoning; cost_reduction=82% | Lightweight models filter trivial syntax checks; Muse Spark 1.3 reserved for proofs. | Multi-tier routing delivers optimal enterprise ROI over monolithic deployment. | PASS — tiering policy confirmed. |
First-party provenance: Meta developer documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
muse-spark-1.3What are Muse Spark 1.3's specs?
| Context window | 1.0M tokens |
| Max output | 128K tokens |
| Modalities | text |
| Extended thinking | Yes |
| Released | 2026-09 |
| Knowledge cutoff | Not published |
| Provider | Meta |
| Tools | tool calling, MCP |
Verified 2026-08-14 — source.
Where does Muse Spark 1.3 rank?
What are Muse Spark 1.3's strengths?
- Built for long-horizon coding and multi-step agentic work
- Improved tool, browser, and computer use across harnesses
- Native tool calling and MCP support
What else should you know about Muse Spark 1.3?
What are common questions about Muse Spark 1.3?
What is Muse Spark 1.3's context window?
Muse Spark 1.3 has a 1.0M-token context window and a 128K-token max output — the 9th-largest context of the 42 current models we track. Source: https://developer.meta.com/ai/models/muse-spark/, verified 2026-08-14.
Does Muse Spark 1.3 support vision or audio input?
No — Muse Spark 1.3 is text-only as of 2026-08-14.
Does Muse Spark 1.3 have a reasoning or extended-thinking mode?
Yes — Muse Spark 1.3 exposes a dedicated reasoning mode for multi-step problems.
When was Muse Spark 1.3 released, and what is its knowledge cutoff?
Muse Spark 1.3 was released 2026-09.
How much does Muse Spark 1.3 cost, and who provides it?
Muse Spark 1.3 is served by Meta at $2.00/M blended tokens (3:1 input:output) — the 22nd-cheapest of 42 current models. Full pricing breakdown: /llm-api-pricing/muse-spark-1-3.