Muse Spark 1.1

Why this oneWhat it is actually for

You are here because a row somewhere said Meta shipped a 600B open-weight model with a five-million-token window, and you were about to plan a self-hosted deployment around it. Stop. As of 2026-07-25 there is no Llama 5 on Meta's blog, on Meta's developer site, or in the meta-llama organisation on Hugging Face, whose newest repository of any kind was created on 2025-04-28. The real decision underneath the question splits two ways. If you need weights you control, the last Llama you can actually download is Llama 4, and Mistral Large 3 is the stronger open-weight frontier option today. If you want Meta's current best model, it is Muse Spark 1.1, and it is closed.

What it isArchitecture, lineage, training

There is nothing to describe, because the model does not exist. Every check pointed the same way. llama.com now issues a 301 to developer.meta.com/ai/. Meta's AI blog carries no Llama 5 post — its 2026 model announcements are Muse Spark, Muse Spark 1.1, Muse Image and Muse Video. The Hugging Face models API for the meta-llama organisation returns thirteen Llama 4 repositories created between 2025-04-01 and 2025-04-04, then Llama-Guard-4-12B and the Prompt-Guard-2 pair in late April 2025, and nothing after that.

What Meta does have as open weights is the Llama 4 herd: Llama-4-Scout-17B-16E-Instruct and Llama-4-Maverick-17B-128E-Instruct. The organisation card describes them as natively multimodal mixture-of-experts models — Scout with 17B active parameters across 16 experts, Maverick with 17B active across 128 — under the Llama 4 Community License, a custom commercial licence rather than an OSI one. Both repositories are gated behind an access request, so their config.json would not load for me and I am not quoting hidden sizes, head counts or rope theta for them.

Meta's frontier line moved elsewhere. Muse Spark was announced on 2026-04-08 by Meta Superintelligence Labs and updated to 1.1 on 2026-07-09. Meta discloses no parameter count and no architecture for it. What it does publish: the model id is muse-spark-1.1, the context window is 1,048,576 tokens, inputs are text, image, video, audio and PDF, and output is text. Weights are not released; the April announcement says only that Meta hopes to open-source future versions. The lineage the site's row assumed — open weights carrying forward from Llama into a bigger Llama — is precisely the thing Meta stopped doing.

At a glanceSee it

Meta's newest open weights are Llama 4, from April 2025; its newest frontier model is Muse Spark 1.1, announced July 2026 and served closed. There is no Llama 5.

CapacityContext, output and what fits

The parametersEvery knob, and what moving it does

Three knobs decide almost everything here. reasoning_effort is the one that moves cost and latency by multiples, and it is where you should start tuning — but note the floor is minimal, not none, because Muse Spark always reasons. prompt_cache_retention set to 24h is the difference between paying $1.25 and $0.15 per million on a fixed system prompt. And max_completion_tokens is your only circuit breaker, because stop is rejected. Note also that logprobs, logit_bias, verbosity, prediction, web_search_options, modalities and audio all return 400 — a request copied verbatim from an OpenAI integration will often fail on one of them, and any evaluation harness that scores token-level confidence will not work against this model at all.

SamplingShaping the output distribution

Meta exposes a conventional two-knob sampler and says almost nothing about how to set it. temperature runs 0 to 2 with a default of 1.0; top_p runs above 0 to 1 with a default of 1.0. At those defaults there is no nucleus truncation at all, so the raw distribution is sampled at full temperature — a noticeably looser starting point than several competing APIs. The documentation does not state whether the two are applied jointly or whether one should be left alone, and it publishes no recommended per-task values, so treat any advice you have read on that as unverified for this model. A sensible starting point is temperature 1.0 with top_p lowered to about 0.9 for anything you intend to parse, and temperature nearer 0.2 for extraction and classification. There is no top_k and no logit_bias, so you cannot ban a token or hard-bound the candidate pool; the only shaping available is the nucleus. On long agentic runs, reasoning_effort changes output character far more than either sampling knob.

ReasoningThinking, effort and budgets

Yes, and it is a first-class request field rather than a prompt trick. reasoning_effort accepts six values — none, minimal, low, medium, high and xhigh — with the default chosen by the model rather than pinned in the docs. That is a finer ladder than most vendors publish, and xhigh in particular has no equivalent elsewhere. There is no numeric token budget: you pick a rung, not a count. On the consumer side the same capability surfaces as a Thinking mode in the Meta AI app and on meta.ai. What Meta does not document on any page I could reach is whether the reasoning trace is returned to the caller, whether it is redacted or summarised, and whether reasoning tokens are billed as output. Those three answers materially change your cost model, and none of them could be confirmed from a primary source as of 2026-07-25 — budget for reasoning tokens being billable until Meta says otherwise.

ToolsFunction calling and server tools

The Meta Model API is built to be swapped in under existing tooling. Meta's own overview states it is drop-in compatible with the OpenAI SDK, the Anthropic SDK, and agent CLIs including OpenCode and Claude Code, and that both a Responses-style API and a Chat Completions endpoint are offered; LangChain, LlamaIndex and the Vercel AI SDK are named as working clients. tools, tool_choice, parallel_tool_calls and response_format all appear in the documented request schema, but the reachable pages do not print their shapes or allowed values, so you should expect OpenAI-compatible semantics and verify against a live call rather than trust a spec. Server-side web search grounding exists and is billed separately per query. The failure modes are concrete and easy to hit: n above 1 is rejected, stop and top_k return HTTP 400, and web_search_options — the OpenAI-shaped way of configuring search — is also a 400, so grounding must be enabled Meta's way, not OpenAI's.

CostPrice, caching, batching, what drives the bill

Meta publishes one flat price for the API rather than per-model rates, which makes the arithmetic unusually simple.

Figures are from the Meta Model API pricing and rate limits page, read 2026-07-25. The cache discount is the lever that matters: cached input is 12% of the uncached rate, and prompt_cache_retention set to 24h keeps a prefix warm long enough for that to bite on real traffic. No batch discount is documented. Whether reasoning tokens count as billable output is not stated anywhere I could read — assume they do. Rate limits are 60 requests and 2,000,000 tokens per minute on the free tier, 3,000 requests and 4,000,000 tokens per minute on paid. Llama 4 weights carry no per-token price at all; the cost there is GPUs, and Meta publishes no guidance on how many.

Where it runsSurfaces and availability

StrengthsWhat it is good at

• Llama 4 Scout and Maverick weights remain downloadable and commercially licensable under the Llama 4 Community License, so the on-prem path Meta opened is still open — just frozen at April 2025.
• The Meta Model API is documented as drop-in compatible with the OpenAI SDK, the Anthropic SDK, and agent CLIs including OpenCode and Claude Code, which makes evaluating it a base-URL change rather than a rewrite.
• A 1,048,576-token window on muse-spark-1.1, with Meta describing the model as actively compacting that context during long agentic runs rather than merely accepting it.
• Cached input at $0.15 against $1.25 uncached is a 88% discount, and prompt_cache_retention: "24h" is a longer published retention than most vendors offer.
• reasoning_effort exposes six rungs including xhigh, a finer ladder than the three-or-four-level controls that are more common.

LimitsWhere it falls down

• There is no Llama 5. Claims of 600B parameters and a 5M-token window trace to aggregator articles and press-release syndication, not to any Meta source; Meta's own April 2026 announcement was Muse Spark.
• Muse Spark weights are closed and the API is in public preview, so the open-weight promise that made Llama interesting does not carry forward to Meta's current best model.
• stop, top_k, logit_bias, verbosity, prediction, web_search_options, modalities and audio all return HTTP 400, so OpenAI-shaped requests frequently fail on a field you did not think about.
• n is capped at 1 — no best-of-k sampling in a single call.
• Meta discloses no parameter count, no architecture and no knowledge cutoff for Muse Spark, so you cannot reason about its behaviour from first principles at all.

Against its neighboursHow it compares

Against Mistral Large 3, this is not a close call if weights matter to you: Mistral Large 3 ships 675B parameters under Apache 2.0 with a published params.json, and costs $0.50 in and $1.50 out per million on the vendor API — cheaper on both sides than Meta's $1.25 and $4.25. Meta's advantages are the 1M window against Mistral's 256k, and a documented reasoning control that Mistral does not list for its Large model. Against GPT-5.6 Sol the comparison is closer in kind, since both are closed frontier models sold per token, and the decision comes down to price, window and whether your agent stack already speaks one dialect. What Meta no longer competes on is the thing this page was filed under. If your requirement was literally an open-weight frontier model you can run yourself, Meta has not had one since April 2025.

Getting startedThe smallest call that works

POST https://api.meta.ai/v1/chat/completions
Authorization: Bearer $MODEL_API_KEY
Content-Type: application/json

{
  "model": "muse-spark-1.1",
  "messages": [
    {"role": "user", "content": "What are three differences between TCP and UDP?"}
  ]
}

Change reasoning_effort first — it is the only field that moves latency and cost by multiples, and the model picks its own default if you leave it out. Then add prompt_cache_retention set to "24h" once your system prompt stabilises. Do not add stop or top_k; both return HTTP 400. Bound long runs with max_completion_tokens instead.

SourcesWhere every claim above came from

• AI at Meta Blog — post index, checked 2026-07-25; no Llama 5 entry, newest model posts are Muse Spark 1.1 and Muse Image / Muse Video
• Introducing Muse Spark: Meta's Most Powerful Model Yet — announcement of 2026-04-08, states weights are not released
• Introducing Muse Spark 1.1 — 1 million token context, public preview on the Meta Model API
• Meta Model API — Overview — base URL, bearer auth, OpenAI and Anthropic SDK compatibility, muse-spark-1.1 as the only model
• Meta Model API — Models — 1,048,576 token window, input and output modalities
• Meta Model API — Chat completion — parameter names, defaults, ranges, and the list of fields that return HTTP 400
• Meta Model API — Create a chat completion — endpoint, headers and the verbatim curl example reproduced above
• Meta Model API — Pricing and rate limits — $1.25 / $0.15 / $4.25 per 1M tokens, web search at $2.50 per 1,000 queries, tier rate limits. Checked 2026-07-25 at ai.developer.meta.com/docs/getting-started/pricing-rate-limits; that page stopped resolving by 2026-08-04, so it is named rather than linked. The figures stand as verified on the date shown — a citation records where a number came from, and the source moving does not unmake that.
• meta-llama on Hugging Face — organisation card and models API, sorted by creation date; newest repository created 2025-04-28, Llama 4 is the latest family
• Could not confirm: any model named Llama 5, in any form, from any Meta source — the 600B / 5M-context / April 2026 figures in the site's row are not supported and should be treated as wrong. Also unconfirmed: Muse Spark's parameter count, architecture, knowledge cutoff and max output tokens; whether its reasoning traces are returned or billed; the schemas for tools, tool_choice and response_format; and Muse Spark availability on Bedrock, Vertex AI or Azure. Llama 4's config.json could not be read because the repositories are access-gated.

RelatedGo deeper