Meta's Muse Spark 1.2: A Genuinely Impressive Mid-Tier Model at a Bargain Price

Meta has unveiled Muse Spark 1.2, their new mid-tier reasoning model that aims to offer frontier-level performance at a fraction of the operating cost.

Muse Spark 1.2 comes equipped with a 1,048,576-token context window (1M tokens) with a maximum output of roughly 131,072 tokens, and it shipped on August 5, 2026 alongside Muse Code, Meta's first terminal coding agent. Meta has not published the parameter count for the Muse Spark family, but the model clearly sits in mid-tier territory: powerful enough for serious agentic work, cheap enough to run all day. Reasoning is always on, with five effort levels (minimal, low, medium, high, and xhigh) trading latency and cost for deliberation — medium is the default, and there is no configuration that skips reasoning entirely.

According to Meta, 1.2 is a coding-focused point release of Muse Spark 1.1 rather than a brand-new base model. It was co-trained with the Muse Code agent using rejection-sampled harness trajectories, plus a self-improvement loop in which Spark 1.1 generated the training environments and graded candidate solutions for 1.2. On the API side it is a drop-in for the OpenAI and Anthropic SDKs at api.meta.ai/v1, which makes swapping it into an existing stack a configuration change instead of a rewrite.

In our evaluations across standard AI benchmarks, Muse Spark scored impressively: - MMLU-Pro: 72.4% - HumanEval: 88.6% - MATH 500: 81.2% - SWE-bench Lite: 44.8%

To put those numbers in context: MMLU-Pro is the hardened successor to the classic MMLU knowledge test — roughly 12,000 questions with ten answer choices instead of four — and frontier models typically score 16–33 points lower on it than on plain MMLU. HumanEval asks a model to write 164 Python functions and checks them against unit tests (pass@1); it has long been saturated at the top of the market, with frontier models scoring above 90%, so an 88.6% from a mid-tier model is genuinely close to flagship territory. SWE-bench Lite uses real GitHub issues rather than toy prompts, which is why its numbers look lower across the board for every model.

On the independent side, Artificial Analysis currently scores Muse Spark 1.2 at 54 on the Intelligence Index — tied with xAI's Grok 4.5, and trailing Claude Opus 5 (61) and GPT-5.6 Sol (59). Independent evaluator Vals ranked it 5th overall on the Vals Index at $0.69 per test, and 14th on the common-harness Terminal-Bench — a useful reminder that Meta's own headline agent numbers (like its 82.9% Terminal-Bench claim) are self-reported and had not landed on the official verified leaderboard at release.

When deployed on cloud infrastructure, Muse Spark achieves 120+ tokens per second on 2x H100 SXM GPUs. Officially, the Meta Model API prices it at $1.25 per 1M input tokens and $4.25 per 1M output tokens — unchanged from Spark 1.1, and roughly a quarter of what the flagship tier charges. For comparison: Claude Opus 5 runs $5/$25, GPT-5.6 Sol runs $5/$30, and Gemini 3.1 Pro runs $2/$12 on prompts up to 200K tokens.

The worked math makes the gap obvious. A mid-size agentic coding task that consumes 50K input and 50K output tokens costs about $0.28 on Muse Spark 1.2. The same task costs about $1.50 on Claude Opus 5 and about $3.00 on Anthropic's Fable 5 ($10/$50). Multiply that by thousands of runs a month and the choice of mid-tier model becomes a line item your CFO will actually notice. Consumers can also use it free inside Meta AI, which makes it one of the cheapest ways to poke at near-frontier quality.

One correction to early coverage of this release (including the first version of this article): Muse Spark 1.2 shipped with closed, proprietary weights, not the open-weights license that some early listings carried. Meta has not published the parameter count, and at launch the model was available only through the Meta Model API — no Hugging Face repository, no self-hosting, no fine-tuning. Artificial Analysis still lists it as a proprietary model.

The open-weights story is still moving, though. On August 10, Mark Zuckerberg announced that Meta will release Muse Spark 1.2's weights "in the coming weeks" — a reversal of the closed strategy Meta adopted when the Muse family replaced Llama back in April 2026. As of this update the repository has not landed, so treat that as a promise rather than a product; the exact date, parameter count, and license are all still unconfirmed.

If you need Meta open weights today, the model you want is Muse Glimmer 30B, which landed on August 10 under Apache 2.0: a 29.6-billion-parameter dense model distilled from Muse Spark, with a vision encoder and a 131K context window, small enough to run on a single 24GB GPU, and already supported in Ollama, LM Studio, vLLM, and OpenRouter.

Muse Spark 1.2 is a tremendous option for development teams needing robust coding and reasoning capabilities without the enterprise pricing of frontier APIs — with the honest caveat that "frontier API" now describes this model too, at least until the promised weights ship. What you are buying is near-frontier agentic performance at a quarter of the price, through an API that is a drop-in replacement for your existing OpenAI or Anthropic integration. If Meta actually follows through on the open weights, this model goes from great value to category-defining almost overnight.