Meta (Meta Superintelligence Labs)

Muse Spark Open

Meta's open-weight Muse line: 816B / 42B active under the Muse Model License, not the same record as API-only Muse Spark 1.3. Artificial Analysis at xhigh (checked 2026-10-17) reads Humanity's Last Exam 43.88 and Terminal-Bench 4.0 8.08 — below 1.3's AA max-effort 48.7 and 33.33 on that same runner (this site's Spark 1.3 Terminal-Bench 4.0 row remains vals.ai's 24.75).

Muse Spark Open benchmarks and pricing, every number sourced: 3 tracked Muse Spark Open benchmark scores (3 independently run, 0 still resting on a vendor’s own claim), priced at $1.25 per million input tokens and $4.25 per million output.

Muse Spark Open architecture: MoE, natively multimodal (text, image, video, speech in → text out); 816B total parameters (42B activated per token); 1M-token context window.

Released: 2026-10-15
License: open-weights
Context window: 1M tokens
Knowledge cutoff: Not disclosed
Verified: source checked 2026-10-17
Parameters: 816B (42B active)
Architecture: MoE, natively multimodal (text, image, video, speech in → text out)

Muse Spark Open’s verified record

Muse Spark Open’s most-compared rival is DeepSeek V4 Pro (0813): 1 lead and 1 not callable across their 2 shared comparisons. Muse Spark Open is priced at $1.25/$4.25 per 1M tokens (in/out) vs DeepSeek V4 Pro (0813)’s $1.32/$3.96.

Against the 67 head-to-head comparisons Muse Spark Open shares with other tracked models:

13 real gaps3 inside the band51 not called

A gap counts for Muse Spark Open only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Muse Spark Open trails on 5 of them.

HLE · no tools Reasoning

±2 is noise

Behind

Step 5 Preview −2.6 · Kimi K3 −3.0 · Gemini 3.1 Pro Preview −3.1 · MiMo-V2.6-Pro −5.5 · 1 previous: Claude Opus 4.8 −4.8

Ahead

DeepSeek V4 Pro (0813) +2.9 · GLM-5.3-Flash +4.0 · MiniMax M3 +4.9 · Qwen3.8-Flash-Next +5.9 · Mistral Large 4 +8.9 · Beam 1.0 +9.2 · Nemotron 3.5 Super VL +13.2 · 1 previous: GLM-5.2 +2.8

Tie

3 models within ±2

Unverified

1 model — vendor-reported on one side

Setup-dependent

22 models — scored on a different harness or effort tier

No verdict for Muse Spark Open anywhere on GPQA Diamond ( saturated); Terminal-Bench 4.0 · xhigh ( nothing independently confirmed on both sides).

None of Muse Spark Open’s agentic comparisons are independently confirmed on both sides yet.

Muse Spark Open API pricing

$1.25 in / $4.25 out per 1M tokens — official pricing source

What Muse Spark Open costs per job

Muse Spark Open cost for three reference workloads, computed from its list rates
Workload	Tokens in / out	Cost
One long chat turn	100K / 10K	$0.168
A codebase review	1,000K / 100K	$1.68
A day of agent work	10,000K / 1,000K	$16.75

Computed from Muse Spark Open’s list rates above — cache discounts and batch tiers are not applied.

Muse Spark Open is one of 4 Meta (Meta Superintelligence Labs) models tracked on this site, at these official list prices.

Meta (Meta Superintelligence Labs) model pricing, official list rates
Model	In / 1M	Out / 1M
Muse Spark Open	$1.25	$4.25
Muse Wonder	$4.00	$20.00
Muse Spark 1.3	$1.25	$4.25
Muse Spark 1.2(previous version)	$1.25	$4.25

Muse Spark Open benchmark scores

Muse Spark Open benchmark scores, provenance, and source links
Benchmark	Score
HLE(no tools)^[1] Reasoning · ±2 is noise	43.9artificialanalysis.ai · 2026-10-17
Terminal-Bench 4.0(xhigh)^[2] Terminal ops · ±12.4 is noise	8.1artificialanalysis.ai · 2026-10-17
GPQA Diamondsaturated^[3] Expert science Q&A — not ranked at any gap size	89.0artificialanalysis.ai · 2026-10-17

Who ran these numbers: 3 of 3 independent — artificialanalysis.ai (3).

HLE: AA's own run, board row 'Muse Spark Open (Xhigh)': hle 0.438832252085264 = 43.88, no tools. Read from the Muse Spark Open page on 2026-10-17.
Terminal-Bench 4.0: AA's own run, board row 'Muse Spark Open (Xhigh)': terminalBench40 0.0808080808080808 = 8.08. Read from the Muse Spark Open page on 2026-10-17.
GPQA Diamond: AA's own run, board row 'Muse Spark Open (Xhigh)': gpqa 0.88989898989899 = 88.99. Read from the Muse Spark Open page on 2026-10-17. GPQA Diamond is saturated on this site — stored for the record; pairwise comparisons on it are tainted.

Notes on the record

Released 2026-10-15 (Meta research post "Introducing Muse Spark Open", checked 2026-10-17). Weights are on Hugging Face under meta-models/Muse-Spark-Open-816B-A42B (BF16) and an NVFP4 sibling, under the Muse Model License with limited commercial terms. Specs from Meta and Artificial Analysis: 816B total / 42B active, 1,048,576-token context, natively multimodal. This is a separate model record from Muse Spark 1.3, which remains proprietary API-only on this site.

Hosted API price on Artificial Analysis's muse-spark-open page is $1.25 / $4.25 per 1M tokens (checked 2026-10-17) — the same sticker AA shows for Muse Spark 1.3. Serving the Muse Spark Open weights yourself has no token fee.

Independent scores from that AA page at xhigh effort (embedded data, checked 2026-10-17): Humanity's Last Exam 43.88; Terminal-Bench 4.0 8.08; GPQA Diamond 88.99 (GPQA is saturated on this site, so the row is stored but pairwise comparisons on it are tainted). AA-AnalystAgent is null. Against Muse Spark 1.3's max-effort AA row (HLE 48.7, Terminal-Bench 4.0 33.33), Open trails on both — open weights are not a free upgrade on those boards.

Compare with

Muse Wonder Muse Spark 1.3 Beam 1.0 Nemotron 3.5 Super VL any model…

FAQ

Is Muse Spark Open the same as Muse Spark 1.3?

No. Muse Spark 1.3 is Meta's proprietary API model on this site. Muse Spark Open is a separate 2026-10-15 open-weight release (816B/42B) under the Muse Model License. On Artificial Analysis, Open's xhigh HLE 43.88 sits below 1.3's max-effort 48.7, and Open's xhigh Terminal-Bench 4.0 8.08 sits below 1.3's AA max 33.33 — different effort tiers, same runner. This site's stored Spark 1.3 Terminal-Bench 4.0 row is vals.ai's 24.75, not that AA figure.

Is Muse Spark Open open source?

Open-weights, not MIT. Meta released BF16 and NVFP4 checkpoints on Hugging Face under the Muse Model License with limited commercial terms (research post and HF collection, checked 2026-10-17). This page's license field is open-weights because the weights are public.

How much does Muse Spark Open cost?

Downloading Muse Spark Open and serving it yourself has no token price. Artificial Analysis lists a hosted rate of $1.25 input / $4.25 output per million tokens (checked 2026-10-17), matching the Muse Spark 1.3 sticker on that site. Partner hosts (Fireworks, Together, Groq and others named in Meta's post) set their own rates.

What are Muse Spark Open's independent scores?

Artificial Analysis at xhigh, checked 2026-10-17: Humanity's Last Exam 43.88, Terminal-Bench 4.0 8.08, and GPQA Diamond 88.99. AA-AnalystAgent has no row. GPQA is saturated here, so that number does not support a Real/Tie judgment against other models.

How does Muse Spark Open compare with Muse Wonder?

Different products. Muse Wonder is a proprietary API model at $4/$20 with AA max-effort HLE 50.09 and Terminal-Bench 4.0 52.02. Muse Spark Open is open-weight at a lower hosted sticker, with weaker AA terminal and HLE numbers at xhigh. They are not a previous-version pair on this site.

What is Muse Spark Open good at?

Meta positions it for coding and agentic desktop work, including Apple MLX on dual M5 Ultra Mac Studios. Independent evidence on this site so far is mixed: HLE 43.88 is usable mid-pack, while Terminal-Bench 4.0 at 8.08 is weak on that board.