Editorial illustration for Meta's New AI Model Scores Solutions to Complex Instructions
Meta's Muse Spark 1.2 Tackles Complex Coding Tasks
Meta's New AI Model Scores Solutions to Complex Instructions
Meta released Muse Spark 1.2 this week, its latest open-weights model, paired with the company's first dedicated coding agent. The pitch is familiar: cheaper access, wider availability, and a direct challenge to closed rivals. The base tier runs 20 cents per million output tokens, a price that undercuts most competitors, though Meta collects user data in exchange for that discount.
Spark 1.2 is billed as a coding-focused update to Spark 1.1, which launched earlier this year. Meta says it poured more compute into programming tasks during training and expanded the number of training environments, aiming at long-running work like building out full repositories or running independent research sessions. The model reportedly plans ahead, holds a fixed goal, and compresses earlier context rather than dropping it, a design meant to keep it steady over extended sessions. Some training data even came from Spark 1.1 itself, which generated coding tasks and graded candidate solutions against them.
Meta backs the release with benchmark results across Terminal-Bench 2.1, DeepSWE v1.1, and 440 tasks pulled from its own codebase, stacking Spark 1.2 against Grok 4.5, Claude Opus 5, GPT-5.6 Terra, and Gemini 3.6 Flash. The comparisons raise a question about how that gap gets measured.
Meta released its new Muse Spark 1.2 model along with its first dedicated coding agent. The cheapest tier runs 20 cents per million output tokens, but users pay for it with their data.
Why this matters
Meta wants Muse Spark 1.2 judged on Terminal-Bench 2.1 and DeepSWE v1.1, benchmarks it picked and partly built the training pipeline around. That's the gap worth flagging: a model trained by scoring its own attempted solutions against its own generated tasks, then graded on suites Meta selected, tells us less about real-world coding reliability than Meta's press materials suggest. The 20-cent-per-million-token price is the real headline for founders weighing infrastructure costs, but that price comes with a data-use tradeoff most teams won't read closely before flipping the switch.
For developers, the coding agent is worth testing against your own messy, ambiguous tickets, not the curated 440-task set Meta cites. For researchers, the self-generated scoring loop is a technique worth watching closely, since it could just as easily produce a model that's good at satisfying its own training rubric as one that's good at satisfying yours. Cheap and fast doesn't mean neutral.
Read the benchmarks Meta didn't cite before deciding this replaces whatever you're using now.
Related Reading
Browse all Meta news →
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
Related Topic OpenAI hires 630 ex-Meta staff as ChatGPT memory may turn data into ads
Business & Startups
Related Topic Meta AI Update Pulls From Your Calendar for Daily Briefings
LLMs & Generative AI
From Archives Prime Intellect's Open-Source Agent Runs Sub-Agents as IPython Kernel Functions
From Archives OpenAI Developer Warns Million AI Models Will Hunt Exposed API Keys
From Archives Meta’s Muse Code AI agent autonomously manages large software projects
Business & Startups
From Archives Meta Adds Persistent Async Agents to Muse AI Coding Tools
Business & Startups
Common Questions Answered
What is the pricing advantage of Meta's Muse Spark 1.2 compared to competitors?
Meta's Muse Spark 1.2 base tier costs 20 cents per million output tokens, which undercuts most competitors in the market. However, users exchange their data with Meta in return for this discounted pricing, making it a trade-off between cost savings and data privacy.
What new capability does Meta introduce alongside Muse Spark 1.2?
Meta released its first dedicated coding agent alongside Muse Spark 1.2 to enhance the model's ability to handle complex coding tasks. This coding agent is specifically designed to work with the updated model to improve solution generation for programming instructions.
How does Meta evaluate the performance of Muse Spark 1.2?
Meta judges Muse Spark 1.2's performance using Terminal-Bench 2.1 and DeepSWE v1.1 benchmarks, which Meta selected and partly built the training pipeline around. However, this evaluation approach has limitations since the model was trained on its own generated tasks and is graded on Meta-selected benchmarks, which may not fully reflect real-world coding reliability.
How does Muse Spark 1.2 differ from its predecessor Spark 1.1?
Muse Spark 1.2 is billed as a coding-focused update to Spark 1.1, which launched earlier in the year, suggesting Meta has concentrated development efforts on improving the model's coding capabilities. The new version represents an evolution of the earlier model with enhanced performance on coding-related tasks.
What is Meta's competitive strategy with Muse Spark 1.2 against closed rivals?
Meta's pitch for Muse Spark 1.2 centers on cheaper access, wider availability through open-weights distribution, and direct competition with closed proprietary models. By offering significantly lower pricing and an open-source approach, Meta positions itself as an accessible alternative to expensive closed-source AI solutions.
Further Reading
- Meta launches new AI coding tool powered by Muse Spark 1.2 - Reuters
- Introducing Muse Code and Muse Spark 1.2 - Meta Research
- Meta debuts Muse Code to take on Anthropic and OpenAI - CNBC
- Meta enters the AI coding wars with Muse Spark 1.2 and Muse Code - VentureBeat
- Meta wants to get inside your terminal with its new coding agent - The Register
