Meta’s next artificial intelligence model, reportedly identified as Muse Spark 1.4, has surfaced in a public GitHub pull request associated with the company’s RocksDB project. The reference appeared in technical testing notes that documented an actual model call using an internal version identified as muse-spark-1.4-internal. The discovery suggests that Meta is already testing a newer version of its Muse Spark model family within its own infrastructure.
The finding was first highlighted on November 13, 2026, after AI observers spotted the model identifier in a pull request involving RocksDB’s experimental live-tuning feature. The notes also revealed details about the model’s token consumption and processing time. Meta has not officially announced Muse Spark 1.4, and there is no confirmed information about its release date, pricing, public availability or performance improvements.
Muse Spark 1.4 Appears in Meta’s RocksDB GitHub Pull Request
The model identifier appeared in pull request number 15674 in Meta’s public RocksDB repository.
RocksDB is an open-source key-value storage engine originally developed by Meta. It is designed to store and retrieve data efficiently and is used in large-scale infrastructure systems.
The pull request, titled around recording the advisor’s model and usage during live tuning, introduced a small logging change to the experimental feature.
The change was designed to record two additional pieces of information whenever an AI model provides a database configuration recommendation:
- Model identification: The name of the AI model generating the recommendation.
- Usage information: The number of tokens consumed and the time taken to produce the response.
The purpose was to compare advanced AI models with smaller models running directly on a host machine. By recording the model name, token consumption and response time, Meta’s engineers can evaluate the cost and performance of different approaches under similar database-tuning conditions.
The model name muse-spark-1.4-internal appeared in the accompanying test notes, providing evidence that the version had been called by Meta’s internal tooling.
Internal Test Reveals 7,867 Reasoning Tokens
One of the most interesting details in the pull request was the usage information recorded during a single model call.
The test notes showed that the internal Muse Spark 1.4 model processed more than 33,000 input tokens and generated over 8,000 output tokens.
| Test metric | Recorded usage |
|---|---|
| Model identifier | muse-spark-1.4-internal |
| Total input tokens | 33,535 |
| Cached input tokens | 17,780 |
| Total output tokens | 8,169 |
| Reasoning tokens | 7,867 |
| Processing time | Approximately 65 seconds |
| Model calls | 1 |
The figures indicate that most of the generated output tokens were classified as reasoning tokens.
This suggests that the model was performing a reasoning-intensive task rather than simply generating a short response. The database-tuning process required the AI advisor to examine information about the running system and recommend configuration changes.
However, the figures represent only one recorded call. They do not establish how fast Muse Spark 1.4 is compared with its predecessor or how well it performs on general reasoning, coding or other benchmarks.
The test was primarily designed to record model usage during database optimisation, not to evaluate the model through a comprehensive performance comparison.
Why Meta Is Using AI Models to Tune RocksDB
The pull request was connected to RocksDB’s experimental live-tuning functionality.
Live tuning allows an automated advisor to monitor a running database and recommend configuration changes that could improve performance. Instead of relying entirely on engineers to adjust settings manually, the system uses an AI model to examine available information and suggest changes.
The AI advisor operates through an internal service called Muse, which provides access to language models.
The system can evaluate database conditions, generate recommendations and track the results of proposed adjustments.
For large technology companies, database performance is important because inefficient configurations can increase infrastructure costs and affect application performance.
Meta operates services that process enormous volumes of information. Improving the efficiency of its storage infrastructure can therefore have financial and operational benefits.
The new logging fields are intended to help engineers compare frontier AI models with smaller models that can run locally on the host machine.
This comparison is important because advanced models may produce more sophisticated recommendations but consume more computing resources and take longer to respond. Smaller models could offer lower costs and faster responses for routine optimisation tasks.
How Muse Spark 1.4 Compares With Muse Spark 1.3
Muse Spark 1.3 is the latest publicly announced version of the model family in Meta’s documentation.
Meta introduced Muse Spark 1.3 on September 2, 2026, highlighting improvements in coding and long-running agentic workflows.
The company positioned the model for tasks involving multiple tool calls, software engineering assistance and long-context reasoning. It also highlighted lower token consumption compared with the previous version.
The model supports a context window of approximately one million tokens and is available through the Meta Model API and Muse Code.
| Model version | Status | Key information |
|---|---|---|
| Muse Spark 1.1 | Publicly available | Original version of the model family |
| Muse Spark 1.2 | Publicly available | Previous generation |
| Muse Spark 1.3 | Latest officially documented version | Improved coding and agentic workflows |
| Muse Spark 1.4 | Unofficially spotted | Internal model identifier found in RocksDB test notes |
The appearance of the 1.4 identifier suggests that Meta is continuing development of the Muse Spark family.
However, there is no evidence yet that the internal version will be released publicly under the same name or that it will include specific improvements over Muse Spark 1.3.
Meta’s Expanding Muse AI Model Family
The discovery comes as Meta continues expanding its Muse family of artificial intelligence models.
The company has been investing heavily in AI infrastructure, research and model development as it competes with companies such as OpenAI, Google and Anthropic.
Muse Spark is positioned around coding, reasoning and agentic workflows, while other models in the broader Muse family target different workloads.
Meta has also introduced Muse Ember for lower-cost agentic tasks and Muse Wonder for more complex, long-running agent workflows.
This broader strategy allows the company to offer different models for different computing requirements rather than relying on one system for every task.
The potential arrival of Muse Spark 1.4 could strengthen this portfolio, particularly if the new version delivers improvements in reasoning efficiency, coding or the ability to handle complex multi-step tasks.
For now, the RocksDB reference provides evidence of internal testing rather than a complete picture of Meta’s development plans.
No Official Launch Date or Performance Benchmarks Yet
Despite the discovery, Meta has not announced Muse Spark 1.4 or confirmed when it might become available.
The pull request does not contain benchmark results, pricing information or a comparison with Muse Spark 1.3.
It also does not establish whether the internal model will be released through the Meta Model API, Muse Code or the company’s consumer-facing AI products.
The recorded token usage cannot be used to calculate the future model’s API pricing. That would require official pricing information and confirmation of how the model’s input, output and reasoning tokens are billed.
The internal identifier also includes the suffix “internal”, which indicates that the reference concerns an in-house model build rather than a confirmed public product.
The Bigger Picture
The discovery highlights how Meta is integrating its AI models into the company’s own engineering infrastructure. Rather than limiting AI models to chatbots and consumer applications, Meta is experimenting with using them to improve database operations, monitor systems and recommend technical changes.
The RocksDB test also demonstrates an important challenge for companies deploying advanced AI models internally: balancing performance against computing costs. A powerful model may generate sophisticated recommendations, but a smaller model could be more economical for routine tasks. Recording token consumption and response times allows engineers to evaluate these trade-offs using operational data.
For Meta, this approach could help determine where advanced models are genuinely necessary and where smaller AI systems can deliver comparable results at lower cost.
Looking Ahead
The next important development will be whether Meta officially confirms Muse Spark 1.4 and provides information about its capabilities. Developers will be particularly interested in potential improvements to coding, reasoning, agentic workflows and API performance compared with Muse Spark 1.3. Until an announcement is made, the model’s public availability and technical specifications remain unknown.
The appearance of muse-spark-1.4-internal in RocksDB test notes nevertheless indicates that a newer Muse Spark version has been used within Meta’s infrastructure. If the company follows its previous release pattern, a public update could eventually follow, but the pull request alone does not establish a launch schedule.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
