Muse Spark 1.3

Original Article ↗ Hacker News Discussion ↗

Meta’s release of the Muse Spark 1.3 language model draws attention for its strong benchmark performance (often near or above current frontier models) and a heavily discounted “contributor” tier that trades price for permission to train on user data. Commenters see it as a highly capable, fast, and cost-effective coding and reasoning model—especially when used with Meta’s own agent harness—while noting that models from Anthropic, OpenAI, Google, and others still lead in some real-world software engineering tasks. Much of the debate centers on data privacy, trust in Meta’s business model, and whether rapidly improving benchmarks (including long‑running meme tests like “pelican on a bicycle” SVGs) still say anything meaningful about how these systems behave in practice.

Pelican SVG benchmark & image behavior

  • Multiple examples of “pelican riding a bicycle” SVGs show clear improvement from 1.2 to 1.3 (better anatomy, bike, accessories).
  • Images tend to use similar composition: side-on, 2D, pelican biking left→right on flat ground, often with hats/scarves.
  • Commenters link this to training data biases (product shots of bikes, cultural left→right “progression”), generic prompts regressing to the “mean,” and SVG favoring flat vector style.
  • Some note that many models now associate this exact prompt with a well-known benchmark; several models can even infer the benchmarker’s identity from the prompt alone.

Benchmarks, saturation, and “pelicanmaxxing”

  • Debate over whether labs might be overfitting to the pelican benchmark; linked analyses suggest no obvious targeted overfitting so far.
  • Many see the pelican task as a quick, intuitive visual smoke test and good for tracking improvements within a model family; others call it close to useless for predicting real-world usefulness.
  • Skepticism extends to coding benchmarks like DeepSWE: several report that benchmark rankings don’t match perceived long-horizon software engineering quality.

Coding performance & harnesses

  • Some found Muse Spark 1.2 weak vs frontier models; others report good experiences for reverse engineering, everyday coding, and Latin translation.
  • Users emphasize that 1.2 followed instructions closely and felt “tool-like,” unlike more “helpful” but intrusive frontier models.
  • 1.3 scores very high on DeepSWE and is described as fast and strong for UI and coding, but real-world quality is disputed.
  • Meta’s own coding harness (Muse Code) is said to be co-trained with the model and more efficient than generic agents; several recommend using it or community harnesses (e.g., OpenCode).

Pricing, contributor tier & data usage

  • Standout feature: extremely cheap “contributor” endpoint (~20× discount) that explicitly trains on user data; standard tier is much pricier.
  • Some praise the transparency and see this as smart price segmentation; others view it as making users “the product.”
  • Concern that users will inevitably send secrets (keys, configs) into contributor models; questions on how well providers can scrub or detect sensitive data.
  • Observation that the discount quantifies how valuable user interaction traces are for RL and model improvement.

Comparisons, progress, and plateau talk

  • Muse Spark 1.3 (especially Max reasoning) appears competitive with top models on some aggregate benchmarks, sometimes edging certain OpenAI models.
  • Many still prefer other frontier models (e.g., Sol, high-end Anthropic, some Chinese or Google models) for complex, long-horizon coding work.
  • Discussion about whether the field is hitting a sigmoid plateau; others argue a new idea (like modern RL-style reasoning) can quickly change the curve again.

Trust, ethics, and provider preferences

  • Strong anti-Meta sentiment: some refuse to use Meta models at all due to past harms (surveillance, manipulation, social damage) and the recent large lawsuit.
  • Similar distrust is expressed toward other major labs and founders; no clear consensus “good actor.”
  • Debate over Anthropic’s and other labs’ ethics, safety posturing, and attempts at regulatory positioning; views are highly polarized.
  • Some argue open-weight models are the only truly trustworthy option; others accept hosted models but try to avoid data retention.

Open weights and deployment

  • Repeated interest in whether Muse Spark 1.2/1.3 will have open weights; hints suggest this may come, which would make the model attractive for self-hosting.
  • Several users are already using Muse Spark via third-party harnesses and note that swapping models in these setups is now relatively easy.