Muse Spark

Muse Spark is a proprietary multimodal reasoning model developed by Meta Superintelligence Labs (MSL), the artificial intelligence division Meta reorganized in 2025. Announced on April 8, 2026, it is the first model in Meta's "Muse" series and the lab's debut release ^[1][2]. Meta presents it as a natively multimodal model "with support for tool-use, visual chain of thought, and multi-agent orchestration," small and fast enough to run interactively yet able to reason through hard problems in science, math, and health ^[1]. It is notable for two reasons beyond its capabilities: it is widely described as Meta's first frontier model, and it is the company's first flagship model released without open weights, a sharp break from the open LLaMA lineage that defined Meta's earlier AI strategy ^[3][4].

The model should not be confused with Microsoft's "Muse," an unrelated 2025 generative model for video-game world simulation. Meta's Muse Spark is a general-purpose large language model and reasoning system aimed at consumer assistants and developer applications.

Meta released Muse Spark 1.1 on July 9, 2026. The update expanded the model's agent, computer-use, coding, multimodal, and long-context capabilities, and became the first model offered through the public preview of Meta Model API.^[13][14] Muse Spark 1.2 followed on August 5, 2026 as a coding-focused update that powers Muse Code, a terminal coding agent Meta released in beta the same day.^[19][21]

Overview

Muse Spark is positioned as the first rung on what Meta calls its scaling ladder toward "personal superintelligence" ^[1]. Rather than leading with a maximally large model, Meta describes a "deliberate and scientific approach to model scaling where each generation validates and builds on the last before we go bigger" ^[2]. Muse Spark is therefore characterized by Meta as "small and fast by design," with the next generation already in development ^[2].

The model accepts text, image, and voice input and can produce text alongside interactive outputs such as small websites, mini-games, and dashboards ^[2][5]. It ships with two everyday modes, branded Instant and Thinking, plus a heavier "Contemplating" mode that orchestrates multiple agents reasoning in parallel for the hardest queries ^[2][1]. Meta says it built the model with input from more than 1,000 physicians to strengthen health reasoning, a domain where it claims unusually strong results ^[1].

Scroll sideways for more →

Meta Superintelligence Labs context

Muse Spark is the first publicly shipped model from Meta Superintelligence Labs, the unit Meta stood up after a leadership overhaul of its AI organization. The lab is led by Alexandr Wang, the former Scale AI chief executive who became Meta's first chief AI officer roughly nine months before the launch ^[5][7]. To bring Wang in, Meta paid about $14.3 billion for a 49% nonvoting stake in Scale AI, a deal that anchored the broader reset of the company's AI ambitions under Mark Zuckerberg ^[7].

The release marked Meta's first major model in over a year and was read by the press as the company's attempt to rejoin the frontier alongside OpenAI, Google, and Anthropic after a period of reduced visibility ^[7][8]. Meta says the work involved rebuilding its AI stack "from the ground up" over the preceding nine months, spanning model architecture, optimization, and data curation ^[1][2].

Announcement and release

Meta unveiled Muse Spark on April 8, 2026, through a Meta Superintelligence Labs blog post on ai.meta.com and a companion newsroom announcement on about.fb.com ^[1][2]. Reporting from CNBC, TechCrunch, and Bloomberg framed it as the first model out of the new lab and the start of a "ground-up overhaul" of Meta's AI efforts ^[7][5][3]. The model had been developed under the internal code name Avocado before launch ^[6][7].

At announcement the model began rolling out inside Meta's consumer products. A May 12, 2026 update to Meta's newsroom post described continued rollout to WhatsApp, Instagram, Facebook, Messenger, and Meta's Ray-Ban and Oakley smart glasses ^[2].

Muse Spark 1.1

Meta announced Muse Spark 1.1 on July 9, 2026 and described it as a significant upgrade within the same model family. The release was available in Thinking mode in the Meta AI app and on meta.ai. It also opened Meta Model API to a public developer preview, replacing the private API access described at the original model's launch with a broader but still restricted program.^[13][15]

Agent operation and long context

Muse Spark 1.1 was trained for agent tasks that span native tools, Model Context Protocol servers, and custom skills. Meta describes a main-agent pattern in which the model gathers context, builds a plan, sends independent jobs to subagents in parallel, and combines their results. A subagent can report that it cannot complete a job so the main agent can revise the plan. This is an agent scaffold rather than a claim that the base model independently gains access to external services.^[13][14]

Meta also raised the supported context window to 1 million tokens. The company says 1.1 actively manages that window by retrieving earlier actions and compacting history while preserving steps needed later in a task.^[13] The evaluation report gives the model a score of 54.1 on a 1 million-token MRCR retrieval test, compared with 74 for GPT-5.5 in Meta's table. The window size therefore describes capacity, not reliable recall of every item placed anywhere in the context.^[14]

For computer use, Meta says the model can choose between interface actions and scripts, batch related actions, and adjust when displayed information changes. These capabilities require an application to supply a computer-control harness and permissions. The model does not directly control a user's computer by default. Coding work was another focus of the update, with support for codebase exploration, bug fixes, features, migrations, testing, and delegation through agentic coding harnesses.^[13][14]

Developer access

At release, Meta Model API was in public preview for US developers. Meta's developer site said new accounts received USD 20 in credits and could adapt existing OpenAI SDK code by changing the API configuration.^[15][16] Reuters reported list prices of USD 1.25 per million input tokens and USD 4.25 per million output tokens.^[18] The model remained proprietary: API access did not include downloadable weights or a reproducible training recipe.

The API surface exposed the model's tool use and multimodal inputs for developer applications. Meta's launch examples included visual-to-code work, detailed descriptions of images and video, and agents that inspect visual or audio material while completing a larger job.^[13] Access conditions, regional eligibility, and preview pricing can change independently of the model version.

Evaluation and safety

Meta reported large gains on several tool, computer-use, and coding tests. The table below reproduces selected scores from the company's evaluation report. All 1.1 results used API access with xhigh reasoning, and the agent benchmarks depended on Meta's chosen tools, prompts, and harnesses. Meta cautioned that its setups might not be tuned to competitor strengths.^[14]

Scroll sideways for more →

The results were not uniformly higher. CharXiv Reasoning declined slightly, and the long-context retrieval result remained well below the strongest comparison in Meta's table.^[14] Artificial Analysis scored 1.1 at 51 on its July 2026 Intelligence Index, eight points above its then-current rescored baseline of 43 for the original model but behind several leading systems. The firm had supported Meta's pre-release evaluation of the xhigh configuration. It also found that the hallucination rate on its Omniscience test fell from 73% to 38%, while the attempted-answer rate fell from 95% to 82% and accuracy fell from 45% to 41%. This suggests that more abstention accounted for part of the measured reduction in hallucinations.^[17]

Meta's frontier-risk report made a distinction between the unmitigated model and the released system. Before mitigations, the evaluators could not rule out the report's high-risk threshold for chemical and biological capabilities or cybersecurity. After applying multiple layers of controls, Meta assessed the residual risk as moderate or lower in every evaluated domain and approved release.^[14]

Model-level safety results improved substantially on several tests. StrongREJECT attack success fell from 25.2 for the original model to 0.5 for 1.1, and AgentDojo prompt-injection attack success fell from 11.9 to 0.7.^[14] The report nevertheless recommends application-level policy checks, strict tool allowlists, workspace isolation, and other system controls. It also notes weaker performance than leading systems in some prompt-injection scenarios, including file-based attacks. The release therefore did not remove the need to constrain tools and review consequential actions.

Muse Spark 1.2 and Muse Code

Meta released Muse Spark 1.2 on August 5, 2026, together with Muse Code, a terminal coding agent in beta for macOS and Linux.^[19] Meta describes 1.2 as a coding-focused update to 1.1, with gains in code generation, complex debugging, codebase understanding, and end-to-end developer workflows, and says larger models are still on the way.^[19] The training changes centered on coding: the company significantly scaled training compute on coding tasks, widened the diversity of training environments, and trained heavily on long-horizon work such as whole-repository generation and large end-to-end projects.^[19]

Two elements of the recipe stand out in Meta's description. First, the model was co-trained with Muse Code, using rejection-sampled harness trajectories and the agent's own toolset, so the model and harness perform best as a pair.^[19] Second, Meta used Muse Spark 1.1 to generate challenging coding environments and instruction-following templates and then to grade candidate solutions against them, a self-improvement loop that produced training data for 1.2.^[19]

Muse Code runs a simple main-agent loop plus asynchronous background agents that persist for a whole session instead of being spawned per task, which Meta says reduces latency and redundant context gathering on long jobs.^[19] The runtime appends every model call, tool run, approval, and edit to a local event log, making sessions replay-exact and restartable after a crash. Bundled skills include /plan, which turns a task into an approval-gated plan, /grill, which stress-tests that plan, and /goal, which works toward a stated objective.^[19] In one case study, Meta ran the model inside Muse Code on GPU kernel optimization sessions of more than 1,000 tool calls over up to 24 hours, iteratively writing, compiling, and profiling Triton kernels for KDA and MLA workloads on NVIDIA Hopper GPUs.^[19][20]

CNBC reported that Muse Code is Meta's first coding agent and that Wang framed the product as competing on price rather than peak capability against Claude Code from Anthropic and Codex from OpenAI.^[21] Wang said pay-as-you-go access carries similar pricing to the Muse Spark 1.1 API (USD 1.25 per million input tokens and USD 4.25 per million output tokens), that a "contributor tier" is more than 10 times cheaper for developers who opt in to sharing data to improve the model, and that Meta had begun accepting requests for zero-data retention.^[21] Muse Spark 1.2 is available in Muse Code and through Meta Model API, which Meta said gained expanded global access with the release; the July preview had been limited to US developers.^[19][15]

Multimodal evaluation and demonstrations

On August 20, Meta published a separate disclosure about Muse Spark 1.2's multimodal behavior. Meta described the results as arriving "ahead of the open-weights release," so the weights remained forthcoming at the August 22 cutoff. The model was still available through Meta Model API and Muse Code rather than as a downloadable checkpoint.^[30] Meta's equal-weight multimodal aggregate rose from 59.8 without tools to 72.0 with tools, a difference of 12.2 points. Muse Spark 1.1 moved from 60.2 to 69.1 under the same two configurations.^[30][31]

The aggregate combines ten benchmarks covering visual perception, visual knowledge, spatial reasoning, user-interface grounding, and chart understanding. Seven component scores use GPT-OSS-120B as a semantic judge, while ERQA, OmniSpatial, and ChartQAPro use deterministic or programmatic grading.^[31] Meta ran the systems through its own evaluation framework at each model's maximum available reasoning effort, with internet access blocked and images resized to at most 2,000 pixels on the longest side. Meta also cautioned that its prompts and tools might not be tuned for the third-party models in its charts. The comparisons with Claude, GPT, and Gemini are therefore Meta-run vendor evaluations, not independent tests.^[31]

The demonstrations cover visual coding, chart and image reasoning, audio-video workflows, and robotics. For manipulation, a specialized Muse Spark planner interprets the scene and decomposes a goal into subtasks. A separate low-level vision-language-action policy, based on an experimental smaller Muse-family model, executes those subtasks. One navigation example ran in simulation with simplified grasps.^[30] The disclosure did not include a VLA checkpoint, robotics code, task-success statistics, or evidence that the demonstrations generalize beyond the shown setups.

Meta also previewed WildArtifactBench, an internal evaluation for open-ended agent deliverables. Its procedure samples pairs of generated artifacts, asks a judge for a binary preference, and fits Elo ratings from the pairwise outcomes. Meta reported automatic and human judging in separate figures, each based on approximately 2,000 comparisons, with maximum reasoning effort and a 200-turn cap for every trajectory.^[30][31] The public page provided task files, inputs, instructions, and rubrics for ten preview tasks; only that preview was public at the cutoff.^[32] Meta did not state that the plotted Elo results were computed only on those ten public tasks, so they should not be described as a ten-task score. Pairwise preference and fitted Elo also measure relative judge choice rather than ground-truth task correctness; prior research on LLM judges has documented position, verbosity, and self-enhancement biases.^[31][34]

The new Meta results had not been independently reproduced by the cutoff. A separate Roboflow evaluation, updated August 20, reported 80.4% for Muse Spark 1.2 across six ground-truth vision tasks, sixth among 31 models, from one evaluation run. That external result uses a different task set and cannot validate Meta's multimodal aggregate, WildArtifactBench ratings, or robotics demonstrations.^[33]

Reported results

Meta's launch charts, backed by a short methodology note, pair each model with its own agent product on the terminal and software-engineering tests: Muse Code for Muse Spark 1.2, mini-swe-agent for 1.1, Claude Code for Claude Opus 5, Codex for GPT-5.6 Terra, Grok Build for Grok 4.5, and Antigravity for Gemini 3.6 Flash, each at its maximum reasoning strength.^[20] The 1.1 figures below therefore differ from those in Meta's July evaluation report, which used a different setup. All numbers are Meta-reported.^[19][20]

Scroll sideways for more →

In Meta's own charts, 1.2 clears Grok 4.5 and Gemini 3.6 Flash on the coding tests, splits with GPT-5.6 Terra (ahead on Terminal-Bench 2.1 and Meta Internal Coding Bench, behind on DeepSWE 1.1, 59.3% to Terra's 64.8%), and trails Claude Opus 5 everywhere except MCP Atlas, a tool-use benchmark scored by Scale AI, where the two Muse Spark versions hold the top two spots shown.^[19][20] GDPVal-AA v2 results come from Artificial Analysis's agent harness and MCP Atlas from Scale AI's, so those two are not CLI-agent comparisons.^[20] As of August 8, 2026, Muse Spark 1.2 did not yet appear on the official Terminal-Bench 2.1 leaderboard at tbench.ai, where Muse Spark 1.1 was listed at 76.2%.^[24]

An early independent result came from Vals AI, which reported on August 6, 2026 that Muse Spark 1.2 was the first model to pass 60% on its Finance Agent v2 benchmark, which gives models the job of a financial analyst. Its leaderboard showed 1.2 at 60.6% accuracy and USD 0.77 per test, ahead of second-place Claude Opus 5 at 58.6% and USD 5.12 per test, with roughly half the latency (304 versus 598 seconds per test).^[22][23]

On August 7, 2026, LMArena, the crowdsourced model-comparison leaderboard that now operates as Arena (lmarena.ai redirects to arena.ai), announced that Muse Spark 1.2 in its xHigh configuration had entered the Text Arena in fourth place with a score of 1498 and had "reshaped the Pareto frontier", and listed the model's pricing at USD 1.25 per million input tokens and USD 4.25 per million output tokens.^[26] The live board showed the same placement on August 8: fourth at 1498, behind three Anthropic Claude models, but on 2,057 votes against the tens of thousands logged for the models around it, which left Arena's confidence interval on the rank stretching from 1st to 19th.^[27]

A more sober independent datapoint came from WeirdML, a benchmark by researcher Håvard Tveit Ihle that gives models deliberately unusual machine-learning tasks and scores the working PyTorch code they write. There, Muse Spark 1.2 (xhigh) averaged 60.3% accuracy as of August 8, 2026, level with GPT-5.4 Mini and just below the original GPT-5 from August 2025, though at a lower run cost than either, while the top entries, Anthropic's Claude Fable 5 and Claude Opus 5, scored close to 92%.^[28][29] Announcing the result on August 7, Ihle called the model "pretty good, but far from the frontier".^[28]

Architecture and approach

Meta has disclosed the broad shape of how Muse Spark was built without releasing weights or a full technical report. The headline claim is efficiency: Meta states the model reaches its capabilities "with over an order of magnitude less compute than our previous model, Llama 4 Maverick," which it credits to an overhauled pretraining stack covering architecture, optimization, and data curation ^[1][3]. (For reference, Llama 4 Maverick was a mixture-of-experts model in the earlier, open-weight Llama 4 Scout and Maverick generation.)

On the post-training side, Meta describes reinforcement learning that shows log-linear scaling with smooth, predictable gains ^[1]. A distinctive element is what third-party coverage labeled "thought compression": during RL the model first improves by reasoning for longer, after which a penalty on thinking time pushes it to condense its chain of thought while holding onto the accuracy gains ^[9]. At inference, this thinking-time penalty combines with multi-agent orchestration to manage test-time compute ^[1].

Meta calls the model "natively multimodal," meaning vision and language were trained together from the start rather than a vision encoder being attached to a finished text model ^[9]. The company highlights performance on visual STEM questions, entity recognition, and localization as evidence of that integrated training ^[1].

Capabilities and intended use

Muse Spark powers a refreshed Meta AI assistant and is meant to be used conversationally across Meta's apps and glasses ^[2]. Beyond plain question answering, Meta and reviewers point to several specific abilities:

• Multi-agent reasoning. The Contemplating mode launches multiple subagents that work a problem in parallel, generate candidate solutions, refine them, and aggregate a final answer. Meta pitches this as a rival to extreme reasoning features such as Google's Gemini Deep Think and OpenAI's GPT Pro mode ^[1][9].
• Visual chain of thought. The model can reason over images and produce interactive visual annotations and displays, which Meta ties to its strong multimodal STEM results ^[1].
• Health reasoning. Trained with input from over 1,000 physicians, Meta presents health as a standout area for the model ^[1].
• Interactive generation. It can produce small interactive artifacts on the fly, including mini-games, dashboards, simple websites, and appliance-troubleshooting walk-throughs ^[2][5].
• Voice and camera. It supports real-time voice conversation and live visual understanding through a device camera ^[2].

Benchmarks

Meta and independent benchmarker Artificial Analysis published a range of scores. The figures below combine Meta's own disclosures with Artificial Analysis's evaluation and a benchmark table compiled by MarkTechPost from Meta's materials. Where a comparison comes from a single source, that source is cited directly; cross-domain comparisons of this kind should be read with the usual caution about vendor-influenced framing ^[10][4][9].

Meta's headline results, in Contemplating mode where noted:

Scroll sideways for more →

Artificial Analysis's independent evaluation placed Muse Spark fourth on its Artificial Analysis Intelligence Index with a score of 52, inside the top five and behind only Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6 ^[4][10]:

Scroll sideways for more →

Artificial Analysis also reported that Muse Spark used about 58 million output tokens to complete the Intelligence Index, comparable to Gemini 3.1 Pro Preview (57M) and well below Claude Opus 4.6 (157M) and GPT-5.4 (120M), which the firm read as evidence of the model's efficiency-oriented design ^[10]. Its summary was that "Muse Spark essentially closes the gap to the frontier in a single release" ^[10].

Availability

At launch Muse Spark was available to general users for free on the meta.ai website and in the Meta AI app, and through a private API preview offered to select partners ^[1]. Using the consumer surface requires signing in with a Meta account via Facebook or Instagram ^[5]. Contemplating mode was described as rolling out gradually rather than being on by default ^[1].

Unlike the open-weight LLaMA models, Muse Spark is closed: its weights are not downloadable and the model cannot be run locally ^[3]. Meta has said it hopes to open-source future versions of the model, and Zuckerberg reiterated that Meta still plans to release open-source models in the future, but the flagship itself is proprietary ^[2][5].

Reception

Coverage was mixed. Market-focused outlets treated the launch as a credible re-entry: Artificial Analysis called it a model that closes the gap to the frontier in one release, and some analysts framed it as evidence of "significant" monetization potential for Meta's assistant ^[10][8]. CNBC noted that while the model showed promise, Wall Street remained focused on the broader question of Zuckerberg's AI strategy and return on tens of billions of dollars of spending ^[11].

Reviewers were more skeptical on capability. Gizmodo's coverage, headlined that the model "doesn't exactly spark joy," argued that Muse Spark was clearly stronger than Meta's prior offerings but "typically fell short" of leading rivals, singling out coding and agentic tasks as continued weaknesses, and cautioned readers to treat Meta's benchmark figures with skepticism given the company's history ^[12]. Several outlets characterized the release as moving Meta from out of the race to competing for "also-ran" status rather than for the top spot ^[12]. Wang himself acknowledged the model was a starting point, writing in his launch thread on X that it was MSL's first model and that there were "certainly rough edges we will polish over time" ^[25].

Limitations

The disclosed limitations fall into three buckets. First, capability gaps: independent and editorial reviews placed Muse Spark below the strongest frontier models on overall intelligence, coding, and agentic benchmarks, and Meta's own materials show a notable weakness on abstract-reasoning tests such as ARC-AGI-2 ^[12][9][10]. Second, transparency: as a closed model with no released weights and no full technical paper at launch, many architectural details remain undisclosed, and benchmark comparisons rely partly on Meta-supplied numbers ^[3][9]. Third, privacy and trust: because the consumer experience requires a Meta login and Meta has a contested history with user data, commentators raised concerns about how personal data might intersect with the model, especially given its health-assistant ambitions ^[5][12].

See also

• Meta Superintelligence Labs
• Meta AI
• Llama 4
• Large language model
• Multimodal model

References

1. ^Meta Superintelligence Labs. "Introducing Muse Spark: Scaling Towards Personal Superintelligence." Meta AI Blog. ai.meta.com/...introducing-muse-spark-msl
2. ^Meta. "Introducing Muse Spark: Meta's Most Powerful Model Yet." Meta Newsroom. about.fb.com/...-spark-meta-superintelligence-labs
3. ^Matthias Bastian. "Meta's Muse Spark is its first frontier model and its first without open weights." The Decoder. the-decoder.com/...-its-first-without-open-weights
4. ^Artificial Analysis. "Muse Spark: Meta is back in the AI race." artificialanalysis.ai/...erything-you-need-to-know
5. ^Amanda Silberling. "Meta debuts the Muse Spark model in a 'ground-up overhaul' of its AI." TechCrunch. techcrunch.com/...n-a-ground-up-overhaul-of-its-ai
6. ^AI CERTs. "Meta Superintelligence: Project Avocado and Muse Spark Strategy." aicerts.ai/...ject-avocado-and-muse-spark-strategy
7. ^Jonathan Vanian. "Meta debuts new AI model, attempting to catch Google, OpenAI after spending billions." CNBC. cnbc.com/...billion-deal-to-bring-in-alexandr-wang
8. ^Benzinga. "Meta Bets Big On Superintelligence With Muse Spark Launch, Analyst Sees 'Significant' Monetization Potential." benzinga.com/...significant-monetization-potential
9. ^MarkTechPost. "Meta Superintelligence Lab Releases Muse Spark: A Multimodal Reasoning Model With Thought Compression and Parallel Agents." marktechpost.com/...ompression-and-parallel-agents
10. ^Artificial Analysis. "Muse Spark: everything you need to know" (Intelligence Index evaluation). artificialanalysis.ai/...erything-you-need-to-know
11. ^Jonathan Vanian. "Meta Muse Spark has promise, Wall Street wants Zuckerberg AI strategy." CNBC. cnbc.com/...ll-street-wants-zuckerberg-ai-strategy
12. ^AJ Dellinger. "Meta's First AI Model From Its Superintelligence Lab Doesn't Exactly Spark Joy." Gizmodo. gizmodo.com/...doesnt-exactly-spark-joy-2000744082
13. ^Meta Superintelligence Labs. "Introducing Muse Spark 1.1 and the Meta Model API." Meta AI Blog. ai.meta.com/...introducing-muse-spark-meta-model-api
14. ^Meta Superintelligence Labs. "Muse Spark 1.1 Evaluation Report." July 9, 2026. ai.meta.com/...muse-spark-1-1-evaluation-report
15. ^Meta. "Muse Spark on Meta Model API." Meta for Developers. developer.meta.com/ai
16. ^Meta. "Build with Muse Spark, now available on Meta Model API." Meta for Developers. developer.meta.com/...build-with-muse-spark
17. ^Artificial Analysis. "Muse Spark 1.1: Everything You Need to Know." artificialanalysis.ai/...erything-you-need-to-know
18. ^Reuters. "Meta debuts Muse Spark 1.1 with preview open to developers." July 9, 2026. investing.com/...review-open-to-developers-4784384
19. ^Meta Superintelligence Labs. "Introducing Muse Code and Muse Spark 1.2." Meta AI Research Blog, August 5, 2026. research.meta.ai/...g-muse-code-and-muse-spark-1-2
20. ^Meta Superintelligence Labs. "Muse Spark 1.2 & Muse Code Evaluation Methodology." August 2026. research.meta.ai/...muse-spark-1-2-methodology
21. ^Vanian, Jonathan. "Meta debuts first AI coding agent to take on Anthropic and OpenAI." CNBC, August 5, 2026. cnbc.com/...-code-to-take-on-anthropic-and-openai-
22. ^Vals AI. "Finance Agent v2" benchmark leaderboard, accessed August 7, 2026. vals.ai/...fabv2
23. ^Vals AI (@ValsAI). X post announcing Muse Spark 1.2 Finance Agent v2 result, August 6, 2026. x.com/...2085479447453651214
24. ^Terminal-Bench 2.1 leaderboard, tbench.ai, accessed August 8, 2026. tbench.ai/...2.1
25. ^Alexandr Wang (@alexandr_wang). Muse Spark launch thread. X, April 8, 2026. x.com/...2041909376508985381
26. ^Arena (@arena). X post announcing Muse Spark 1.2 (xHigh) Text Arena result, August 7, 2026. x.com/...2085747583767527528
27. ^Arena.ai. Text Arena leaderboard, accessed August 8, 2026. arena.ai/...text
28. ^Håvard Ihle (@htihle). X post reporting Muse Spark 1.2 WeirdML result, August 7, 2026. x.com/...2085680447330058625
29. ^Håvard Tveit Ihle. "WeirdML" benchmark leaderboard and data. htihle.github.io, accessed August 8, 2026. htihle.github.io/weirdml
30. ^Meta Superintelligence Labs. "The Multimodal Intelligence of Muse Spark 1.2." Meta AI Research, August 20, 2026. research.meta.ai/...intelligence-of-muse-spark-1-2
31. ^Meta Superintelligence Labs. "Muse Spark 1.2 Multimodal Evaluation Methodology." Meta AI Research, August 2026. research.meta.ai/...timodal-evaluation-methodology
32. ^Meta Superintelligence Labs. "WildArtifactBench." Ten-task preview, accessed August 22, 2026. research.meta.ai/wild-artifact-bench
33. ^Roboflow. "Meta: Muse Spark 1.2." Vision Evals updated August 20, 2026; accessed August 22, 2026. playground.roboflow.com/...muse-spark-1-2
34. ^Lianmin Zheng et al. "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena." arXiv:2306.05685, 2023. arxiv.org/...2306.05685