Entity trail

Muse Spark

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-07-10 / sources-researcherllm 0.31.1 Fixes a Tool Call Crash That Only Shows Up on Non-OpenAI ProvidersSimon Willison shipped llm 0.31.1 on July 9 to fix issue #1521: the default OpenAI plugin converted `arguments=None` to an empty string and then called `json.loads("")`, crashing on any provider that returns null arguments for a zero-argument tool call. He found it running `llm -m meta-ai/muse-spark-1.1 -T llm_time 'time?'` against the llm-meta-ai plugin — Meta's new API surfaced the bug within a day of launch. If you maintain an OpenAI-compatible client that talks to anything other than OpenAI, this is the exact class of bug to audit for right now.
2026-07-10 / sources-researcherSimon Willison Breaks Down the GPT-5.6 Price Ladder: Luna, Terra, SolWillison's July 9 post lays out the three-size family from smallest to largest — Luna $1/$6, Terra $2.50/$15, Sol $5/$30 per 1M input/output tokens — and treats general availability as the notable event rather than the benchmark claims. His framing is the useful one for builders: the family is a cost ladder, not a capability jump, and the interesting question is which tier is the new default for agentic loops. Worth reading alongside his same-day llm 0.31.1 release, which was triggered by a real tool-calling bug against Meta's brand-new Muse Spark API.
2026-07-10 / sources-researcherMeta Charges for a Frontier Model for the First Time: Muse Spark 1.1 at $1.25/$4.25Meta released Muse Spark 1.1 on July 9 — a multimodal reasoning model built for tool use, computer use, and multi-agent orchestration with a 1M-token context — priced at $1.25 per 1M input and $4.25 per 1M output via a new Meta Model API, in public preview for US developers. It is the first Meta frontier model behind a paid API, a direct reversal of the open-weights posture the company built its AI reputation on. Zuckerberg posted on X for the first time in three years to promote it, promising "aggressive pricing" against OpenAI and Anthropic; the pricing undercuts GPT-5.6 Terra on both sides of the meter.
2026-07-10 / thought-leaders-researcherllm 0.31.1 Released — Point Update to the Command-Line LLM HarnessWillison pushed llm 0.31.1 on July 9, a maintenance release of the CLI tool for accessing language models from the terminal. Released in the same window as the llm-meta-ai plugin and his Muse Spark 1.1 and GPT-5.6 writeups, suggesting the point release absorbs plumbing needed for the week's new model APIs. Low individual signal, but llm remains the most direct way to script multi-provider model calls without SDK lock-in.
2026-07-10 / thought-leaders-researcherSimon Willison Ships llm-meta-ai 0.1, Bringing Meta's API Into the llm CLIA same-day plugin release on July 9 giving the llm command-line tool access to Meta's AI API, landing alongside the Muse Spark 1.1 launch. Willison's plugin cadence is a reliable leading indicator of which new model APIs are actually accessible and worth wiring into a local toolchain. Single-source and small in scope, but it makes Meta's newest models addressable from the same CLI harness as OpenAI and Anthropic endpoints.
2026-07-09 / arxiv-researcherMeta Ships Muse Image, Superintelligence Labs' First Image ModelMeta announced Muse Image on July 7, 2026, the first image-generation model from Alexandr Wang's Meta Superintelligence Labs (codenamed Mango, following April's Muse Spark LLM). It generates and edits images from text, writes and executes code to produce accurate plots and QR codes, and searches the web to ground outputs; it launched free in Meta AI, WhatsApp, and Instagram Stories and drew immediate pushback over training on users' personal photos, with a Muse Video model planned next.
2026-06-16 / reddit-researcherMeta's 'Hatch' Developer Agent Heads to Internal Testing by End of JuneReporting on the personal-agent race notes Meta is targeting end-of-June internal testing for Hatch, a developer-focused agent product under Alexandr Wang's Superintelligence Labs, as Google and Meta scramble to catch Anthropic and OpenAI on agentic workflows. It's an early, unreleased signal rather than a shipping product, but it indicates Meta is moving from foundation-model catch-up (Muse Spark, April) toward agent tooling. Builders should treat this as a watch-item, not yet an adoptable tool.
2026-05-06 / sources-researcherMeta Developing 'Hatch' Agentic AI Assistant on Muse Spark — Targeting Instagram Launch Q3 2026The Financial Times reports Meta is building a personalized agentic AI assistant codenamed 'Hatch' on its Muse Spark model, mirroring OpenAI's agentic consumer strategy. The assistant is being tested internally with a target Instagram launch before Q4 2026 and plans to integrate agentic shopping capabilities. Internal testing is expected to complete by end of June 2026, with Muse Spark serving as the assistant-role model family alongside Llama's foundation-research role.
2026-05-02 / hn-researcherMeta Abandons Open-Source Llama for Proprietary Muse Spark — No Migration Path, Cloud-Only, No Downloadable WeightsMeta has formally pivoted from its open-weight Llama strategy to the fully proprietary Muse Spark, its first model from the newly formed Meta Superintelligence Labs. Unlike Llama, Muse Spark offers no downloadable weights, no self-hosting capability, and only private API preview access to select partners — making it more proprietary than offerings from OpenAI or Anthropic. There is no migration path for Llama users; Meta justified the shift by noting $115–135B in guided 2026 AI infrastructure spend with no frontier-competitive open model to show for it.
2026-04-13 / news-researcherStratechery: Does Aggregation Theory Survive in a World of Constrained Compute?Ben Thompson published 'Mythos, Muse, and the Opportunity Cost of Compute' on April 13, analyzing whether his Aggregation Theory holds when AI companies face genuine compute scarcity. The piece examines the strategic implications of Anthropic restricting Mythos to Project Glasswing's ~40 organizations and Meta launching Muse Spark, arguing that controlling demand will still give power over supply even when compute is the binding constraint. This is the first major framework-level analysis of how AI's physical resource limits reshape platform economics.
2026-04-12 / news-researcherMeta Debuts Muse Spark — First Proprietary AI Model Under Chief AI Officer Alexandr WangMeta launched Muse Spark (code-named Avocado), the first model from its new Muse series built by Meta Superintelligence Labs under Alexandr Wang, whom Meta hired for $14.3B in June 2025. The model is fully proprietary with no weights or fine-tuning access — a stark departure from Meta's open-source Llama strategy — and achieves Llama 4 midsize performance at an order of magnitude less compute. Muse Spark will roll out inside Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta AI glasses, backed by Meta's planned $115-135B AI capex for 2026.
2026-04-09 / thought-leaders-researcherMuse Spark Benchmarks: GPQA 89.5% Trails Opus 4.6 (92.7%), GPT-5.4 (92.8%), Gemini 3.1 Pro (94.3%) — Leads on HealthBenchFortune's analysis of Meta's published benchmarks shows Muse Spark is competitive but does not surpass leading models across the board. On GPQA Diamond (PhD-level reasoning), it scores 89.5% — trailing Gemini 3.1 Pro (94.3%), GPT-5.4 (92.8%), and Claude Opus 4.6 (92.7%). However, it outperforms all rivals on HealthBench Hard at 42.8%, suggesting Meta focused training on healthcare and medical reasoning capabilities. Meta claims improved training techniques enable smaller models matching older midsize Llama 4 capabilities at 'an order of magnitude less compute.'

Graph relationships

RELEASED
Meta -> Muse Spark
Meta released Muse Spark as closed-source from Meta Superintelligence Labs.
Source finding
RELEASED
Meta -> Muse Spark
Meta released Muse Spark.
Source finding
BENCHMARKED AGAINST
Muse Spark -> Gemini 3.1 Pro
Muse Spark scores 89.5% on GPQA vs Gemini 3.1 Pro at 94.3%
Source finding
BENCHMARKED AGAINST
Muse Spark -> GPT-5.4
Muse Spark scores 89.5% on GPQA vs GPT-5.4 at 92.8%
Source finding
BENCHMARKED AGAINST
Muse Spark -> Claude Opus 4.6
Muse Spark scores 89.5% on GPQA vs Opus 4.6 at 92.7%
Source finding
RELEASED
Meta -> Muse Spark
Meta launched Muse Spark first model in Muse series.
Source finding

Source trail

Muse Spark 1.1 fully proprietary Muse Spark massive backlash Muse Spark Gizmodo's take Google is going more open Gemma 4 now runs reliably

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues