Introducing Muse Spark 1.3

We’re excited to merchandise Muse Spark 1.3, which delivers improved capacity crossed agentic and coding tasks. Drawing connected what we learned from months of wide take of Muse Code and Meta Model API, we’ve besides made this exemplary easier to usage successful real-world settings. Smarter and much practically useful, Muse Spark 1.3 advances our activity toward individual superintelligence.

Muse Spark 1.3 is rolling retired coming successful Muse Code and Meta Model API. Previously disposable reasoning modes are disposable coming pinch max reasoning coming soon aft we decorativeness further information testing.

For much specifications astir our evaluations, spot our report.

Agentic Workflows

Muse Spark 1.3 is designed to amended prolong longer-horizon activity by collaborating pinch users and juggling aggregate workflows successful a single, agelong thread. When fixed an open-ended objective, it uses devices to make its ain discourse crossed messy and conflicting sources, proactively corrects gaps successful its plan, and keeps way of what it has learned to nutrient a last deliverable. We trained the exemplary crossed a divers group of harnesses to generalize to various agentic environments.

Trained to much actively collaborate pinch the user, Muse Spark 1.3 asks clarifying questions erstwhile prompts are ambiguous, invokes thief from the personification erstwhile stuck, and confirms earlier taking consequential actions. When moving connected agelong tasks, it adapts to personification preferences, either providing predominant updates aliases moving silently successful the background.

Muse Spark 1.3 follows complex, long-form instructions much reliably than earlier Muse Spark models. Across multi-step tasks, it’s amended astatine preserving elaborate requirements without dropping constraints aliases drifting from the requested workflow.

We’ve besides improved the multitasking capabilities of Muse Spark 1.3. For example, it now much accurately maps incoming prompts to the correct task wrong messy, single-threaded contexts, sloppy of whether the personification is steering past requests aliases interrupting them.

The exemplary has amended consciousness of its ain capabilities and limitations. We trained Muse Spark 1.3 to person a amended consciousness of what it tin and can’t do, what it knows and doesn’t know, and erstwhile it hits hurdles alternatively of hallucinating outcomes.

Prompt and task context

You are a Mechanical Engineer astatine a mini aerospace patient designing an experimental X-Wing assembly for a next-generation aircraft. To support the creation review, create a draught flow-simulation study based connected the attached: (1) the preliminary CFD simulation results, and (2) STEP record containing a CAD exemplary of the helping assembly utilized for simulation. Use the CFD post-processing information to outline the study objectives, picture the computational domain and mesh, statement the worldly properties, inlet/outlet bound conditions, and engineering goals utilized to thrust convergence. Summarize cardinal capacity metrics specified arsenic highest axial velocity, maximum turbulence intensity, turbulent kinetic energy, and the forces acting connected the wing. Include a array of world extremity values and a 2nd array showing minimum and maximum values for important section variables (e.g., density, pressure, temperature, velocity components, Mach number, and comparative pressure). Discuss the implications of these results for aerodynamic capacity (e.g., assistance vs. drag, daze formation, travel separation, and turbulence) and reason pinch preliminary recommendations to amended the design. Overall, the study should beryllium concise, well-structured, and exported arsenic a PDF. Organize your findings into the pursuing sections: "Objective," "Simulation environment," "Boundary conditions," "Results," "Discussion," and "Conclusion." Present numerical results successful tabular form. Ultimately, this study will beryllium utilized internally to little the creation squad and guideline further optimization work.

Muse Spark 1.3 output

Prompt and task context

You’re an audio operation technologist moving astatine a reputable signaling studio. A caller creator has brought successful a demo opus to way vocals over, but it needs to beryllium somewhat cleaned up earlier this tin happen. The creator likes the semi-rough, earthy value of the performances, and would for illustration to clasp that 70's era quality; but they want to fix, correct, aliases edit immoderate evident mistakes successful the bass guitar part. These mistakes tin beryllium the incorrect note, played dissonant aliases retired of key; these notes should beryllium replaced pinch an due statement copied from different constituent successful the opus that’s successful tune pinch the cardinal of the opus and arrangement. This should beryllium reasonably easy, arsenic the chords and statement repetition respective times passim the song. Some mistakes whitethorn beryllium violative drawstring noise, clicks, aliases pops. You should edit these retired and switch them pinch silence, without changing the wide magnitude of the track. The creator has provided immoderate timecode references for spots that will request to beryllium fixed successful reference record attached (Bass Edit Spots.docx). The spots are referenced successful “minutes: seconds: milliseconds”, for example: 01:44.375 You’ve besides been provided pinch Stems of each instrumentality way successful the song, including the raw, unedited Bass way that you’ll need. After each edits and corrections are made to the soloed bass track, please operation your edited bass way backmost successful pinch the different instruments. Be judge to operation the bass measurement astatine a comparable measurement to the Rough Mix provided arsenic a reference, without altering the measurement levels of the different instruments since the creator likes really they sounded successful the Rough Mix. All way lengths should stay the aforesaid earlier and aft editing to guarantee that each instruments sync up. Your last delivered record should beryllium a Stereo operation of each the stems exported successful 48k/24b .WAV, named “State of Affairs_FULL_EDIT_MIX”, pinch the recently edited Bass way replacing the earthy Bass track. The extremity consequence will beryllium a much master sounding demo track, free of awesome bass mistakes, that is suitable for the creator to way their vocals with.

Muse Spark 1.3 output

Coding

Muse Spark 1.3 was trained connected much long-horizon coding tasks and shows improved usability successful communal engineering workflows. Relative to Muse Spark 1.2, it takes less turns wherever not needed and is little verbose, while having a cleaner wide coding style. In comparisons by Meta engineers, it proved to beryllium importantly faster and much efficient, utilizing ~20% less instrumentality calls and ~25% less tokens.

Availability

Muse Spark 1.3 is disposable coming successful Muse Code and successful Meta Model API.

Get started pinch Muse Code

Safety

We’ve improved information on respective axes astir applicable to agentic and coding capabilities. Muse Spark 1.3 shows stronger adversarial robustness, pinch improved guidance to adversarial inputs and punctual injections. On analyzable agentic tasks, the exemplary has amended calibration connected what constitutes irreversible actions and proceeds accordingly. Together, these changes bespeak amended discretion and judgement successful long-horizon agentic tasks.

Looking Forward

We person an breathtaking roadmap lined up, including bigger models, the Muse Spark unfastened weights release, and more. Stay tuned.