Muse Code and Muse Spark 1.2

Original Article ↗ Hacker News Discussion ↗

Meta’s release of Muse Spark 1.2 and its Muse Code harness is seen as a solid but not frontier-level step in AI-assisted software development, with performance generally below top models from OpenAI, Anthropic, and Chinese labs like DeepSeek. The most notable aspect for many is Meta’s aggressive “contributor” pricing, offering steep discounts if users allow their code and prompts to be used for training—raising sharp concerns about privacy, data retention, and Meta’s trustworthiness. Commenters also criticize the login and account model, the lack of open weights, and question whether enterprises will adopt a mid-tier, closed model from a company with Meta’s track record.

Model positioning & benchmarks

  • Muse Spark 1.2 is framed as a coding-focused, mid-tier model; commenters note it trails top models (e.g., OpenAI’s higher-end tiers, Anthropic’s frontier models, DeepSeek V4 Pro) and is “not SOTA but decent.”
  • Some criticize comparisons: Meta mostly shows results against mid-tier models (e.g., Terra, some Anthropic models) and often loses, especially when better models are included.
  • Debate on benchmarks:
    • One side: benchmarks are biased/gamed and don’t reflect real-world performance.
    • Other side: if Meta can’t find any benchmark where they beat Sol/Fable, it signals they’re behind.
  • Kernel optimization case study is seen as interesting; models show stepwise improvements reminiscent of genetic algorithms.

Release cadence & product quality

  • 1.2 follows 1.1 by only a few weeks; some see this as a “do-over” after a weak 1.1 launch, others say frequent minor updates are now normal and just reflect releasing checkpoints.
  • Some users report Muse Code + Spark 1.2 is roughly comparable to other mid-code tools but still behind favorites like Claude Code/Codex in practice.

Muse Code harness

  • Muse Code is a CLI-like coding agent, apparently implemented in Rust and reminiscent of Codex CLI but with different config and features.
  • Interesting harness capabilities mentioned: multiple workers in separate worktrees, crash recovery, built-in orchestrator/subagent pattern.
  • Questions remain about how it ranks versus other harnesses; no clear, shared benchmark standard for “code agents” exists.

Pricing, contributor mode & data use

  • Dual pricing: standard tier vs. “contributor” tier that is ~10–20x cheaper if you allow training on your data.
  • Many find this pricing very attractive and competitive with DeepSeek V4 Flash; others are excluded because contributor mode appears US-only.
  • Some praise the transparency (“pay less if we train on your data”), but there is deep distrust that Meta will honor non-training promises; others argue contractual and legal risk make abuse unlikely.
  • Comparisons are drawn to OpenAI’s free/discounted usage for data sharing, with complaints that OpenAI’s program is more complex and limited.

Trust, privacy & login friction

  • Strong skepticism about giving Meta access to proprietary code, even via a coding agent, due to its past data practices.
  • Concerns about:
    • Mandatory login and some flows asking for selfie verification.
    • Tying dev usage to Facebook/Instagram domains/accounts, which can be blocked by corporate firewalls and disliked by developers.
  • Some users say they would tolerate Meta’s offer only for non-sensitive or personal projects.

Open source & availability

  • Repeated calls for Meta to open-source model weights again; many say they’d care more if Spark were open-weight.
  • Questions about whether Muse Code itself is open source; no clear answer in the thread.
  • Contributor pricing and certain features are reported unavailable in EU, Australia, and other regions.

Broader ecosystem commentary

  • Frustration that every AI lab is launching its own coding agent; reasons suggested include telemetry, marketing, controlling the harness, and owning the customer relationship.
  • Some feel Meta’s main play is undercutting on price using profits from its core business.
  • A few see this as decent progress given Meta’s recent reboot in LLMs but still uncompetitive with the absolute frontier today.