Meta enters the AI coding wars with Muse Spark 1.2 and Muse Code with persistent async background agents

Meta right this moment released Muse Code, a terminal-based AI coding agent now in beta, alongside Muse Spark 1.2, a coding-focused replace to its Muse Spark household of frontier fashions — a one-two punch that places the corporate in direct competitors with Anthropic's Claude Code, OpenAI's Codex, and the rising discipline of agentic coding harnesses which have quickly grow to be the first means {many professional} builders ship software program.

"Releasing Muse Code in beta right this moment," Meta co-founder and CEO Mark Zuckerberg wrote in a post on rival social network X (below his longtime deal with @finkd). "It's a terminal coding agent that takes on full software program engineering duties throughout massive repos: planning adjustments, writing code, validating the outcomes."

The launch marks Meta's most severe entry but right into a class it has largely watched from the sidelines.

Whereas Anthropic and OpenAI turned their coding brokers into flagship merchandise — and startups like Cursor constructed billion-dollar companies on the workflow — Meta's developer story lengthy centered on Llama, the open-weight mannequin household it gave away to the tune of greater than a billion downloads.

Muse Code adjustments that in additional methods than one: it's a full harness, installable on macOS or Linux with a single curl command, co-trained with the mannequin that powers it — and, just like the Muse Spark fashions behind it, fully proprietary.

Nevertheless, Zuckerberg teased that open supply could also be within the playing cards for Muse Spark or maybe one other product fully, in a reply to a question on X, saying "I'll have extra to share on that quickly."

Builders and potential customers can set up it now on their Terminal utilizing the next one-line command — however be warned, if that's you, you'll have to log in with a Meta account and supply billing particulars first with the intention to start: curl -fsSL https://dev.meta.ai/set up.sh | bash

Persistent background brokers and parallel worktrees

Muse Code's headline architectural guess is what Meta calls async background brokers.

Quite than spawning helper brokers recent for every process — the sample most rival harnesses use — Muse Code retains a set of specialised background brokers alive for your entire session.

In line with Meta's weblog publish, these brokers "stay energetic all through every session, fairly than being spawned for particular person duties, serving to keep away from redundant info gathering," finishing up subsequent steps on their very own and selecting when to report again to the principle agent.

The sensible pitch is much less latency and fewer babysitting: an agent that already is aware of the repository doesn't need to re-explore it each time the developer asks for one thing new.

When a job is massive sufficient, Muse Code followers out to separate sub-agents working in parallel, every in its personal remoted git worktree, so the developer's working copy is rarely touched.

"In testing we had it construct six options for a recreation concurrently with no collisions," Zuckerberg wrote on X.

Worktree isolation and parallel sub-agents exist in competing instruments, however Meta is leaning on the mix of persistence plus parallelism as its differentiator.

The second notable design alternative is auditability. Each mannequin name, instrument run, approval, and edit is appended to a native occasion log earlier than it executes — a single supply of reality that Meta says makes the runtime "replay-exact and restart-safe."

If Muse Code crashes 20 hours right into a long-running process, it resumes exactly the place it stopped, with no misplaced work and no re-prompting. For engineering leaders who’ve been burned by opaque agent runs, a whole native audit path could show to be the characteristic that issues most in enterprise evaluations.

Muse Code additionally ships with bundled "expertise" that can look acquainted to customers of rival instruments: /plan turns a process into an approval-gated plan, /grill stress-tests that plan till it holds up, and /objective drives the agent towards completion of a said goal.

Muse Spark 1.2: co-trained with its personal harness

Beneath the hood is Muse Spark 1.2, which Meta describes as a coding-focused replace to Muse Spark 1.1 with "considerably scaled up coaching compute on coding duties" and broader coaching atmosphere variety, enhancing code era, advanced debugging, and codebase understanding whereas sustaining basic agentic functionality.

The replace lands squarely on the Muse household's weakest flank. When the unique Muse Spark debuted in April, it vaulted Meta again into the highest 5 on frontier reasoning and imaginative and prescient benchmarks — however trailed on the agentic coding evaluations that matter most to this market, scoring 77.4 on SWE-Bench Verified in opposition to Claude Opus 4.6's 80.8 and Gemini 3.1 Professional's 80.6, and lagging nicely behind GPT-5.4 on GDPval's measure of long-horizon work duties.

4 months later, a coding-specialized checkpoint paired with a purpose-built harness reads as Meta's direct reply to that hole.

Two coaching particulars stand out. First, Meta co-trained the mannequin with Muse Code itself, utilizing rejection-sampled harness trajectories and recipe optimizations for objectives, context compaction, and sub-agents — that means the mannequin was explicitly tuned to carry out greatest inside this explicit instrument. That mirrors an industry-wide shift away from treating fashions and harnesses as separable merchandise.

Second, Meta used a self-improvement loop: Muse Spark 1.1 generated difficult coding environments and instruction-following templates, then graded candidate options in opposition to these necessities, producing a scalable coaching dataset for its successor. Meta credit the loop with making 1.2 measurably higher at following advanced directions.

Meta revealed benchmark charts evaluating Muse Spark 1.2 in opposition to different coding fashions on Terminal-Bench 2.1, DeepSWE 1.1, and an inside Meta coding benchmark, pointing readers to a separate methodology report for particulars — although the announcement textual content itself doesn't tout any placements, an uncommon reticence in a discipline the place rivals trumpet leaderboard wins. The charts clarify why: they present a robust however clear second place.

On Terminal-Bench 2.1, Muse Spark 1.2 working in Muse Code scored 82.9%, edging OpenAI's GPT-5.6 Terra in Codex (81.8%) and xAI's Grok 4.5 in Grok Construct (81.6%) however trailing Anthropic's Opus 5 at max effort in Claude Code, which leads at 86.7%.

On DeepSWE 1.1, Muse Spark 1.2 posted 59.3% — third, behind Opus 5 (65.0%) and GPT-5.6 Terra (64.8%). Most placing is Meta's personal inside coding benchmark, the place Muse Spark 1.2's 70.6% comfortably beats GPT-5.6 Terra (65.4%) and Gemini 3.6 Flash (63.9%) but nonetheless sits almost 9 factors behind Opus 5's 79.4% — an unusually candid admission that even on the check Meta designed itself, Anthropic's mannequin wins. Certainly, Claude tops all three charts.

The generational features are actual, although: Muse Spark 1.2 improves on 1.1 by 6.7 factors on Terminal-Bench and 6.3 on DeepSWE. One caveat buried within the chart labels — the 1.1 scores have been recorded within the generic mini-swe-agent harness whereas 1.2 ran in Muse Code, so a few of that leap belongs to the brand new harness fairly than the brand new mannequin.

The corporate's most placing demonstration is a long-horizon case research: Meta pointed Muse Spark 1.2 at GPU kernel optimization and let it run for greater than 1,000 instrument calls over as much as 24 hours on NVIDIA Hopper {hardware}.

Working in Triton and barred from merely wrapping present third-party kernel libraries, the agent wrote, compiled, and profiled its option to what Meta calls "substantial enhancements" over baseline implementations of KDA and MLA kernels — together with genuinely non-obvious optimizations like re-centering gated cumulative decay at a bit midpoint.

"It saved discovering substantial enhancements nicely past the preliminary exploration part," Zuckerberg wrote. Sustained enchancment over a 24-hour autonomous run, if it holds up exterior Meta's demos, addresses some of the persistent criticisms of coding brokers: that they plateau or drift as soon as previous their preliminary burst of progress.

Your knowledge for a reduction?

The pricing construction stands out as the most consequential — and most scrutinized — a part of the launch. Meta is providing Muse Spark 1.2 by way of its Meta Model API in two tiers.

The customary tier is priced at $1.25 per million enter tokens and $4.25 per million output tokens (with cached enter at $0.15), and Meta commits that prompts and completions on this tier usually are not used to coach its fashions. There is no such thing as a long-context premium, and fee limits run to three,000 requests and 4 million tokens per minute, per group. It's about mid-range value, in comparison with different main AI fashions accessible over API.

The contributor tier is the place Meta's technique diverges sharply from its rivals: $0.10 per million enter tokens and $0.20 per million output tokens — roughly 12x and 21x cheaper than customary, respectively, with cached enter at a near-free $0.002 — in alternate for express permission to make use of your prompts and completions to coach future Meta fashions. It's the most cost effective accessible in the marketplace, however you pay along with your knowledge — as described beneath.

That is the tier Zuckerberg is steering new customers towards: "It's straightforward and low-cost to get began," he wrote. "Set up Muse Code with one line and you can begin on our contributor tier."

In VentureBeat's personal testing on a Mac mini, the one-line installer labored as marketed — a 97 MB obtain and a sign-in — however the agent stopped in need of working something, reporting that no fashions have been seen and that fee was "required to complete establishing your account."

In different phrases, even the closely discounted contributor tier requires a fee methodology on file earlier than Muse Code will do any work: low-cost is correct, however free is just not.

Meta frames the contributor tier as reducing the barrier for prototyping and experimentation "the place coaching in your knowledge is appropriate."

However it additionally means the default on-ramp for Muse Code sends builders' code and prompts into Meta's coaching pipeline — a tradeoff enterprises with proprietary codebases might want to consciously choose out of by transferring to plain pricing.

The contributor tier additionally carries a lot tighter fee limits (60 requests per minute versus 3,000), a transparent sign it's geared toward people and small experiments fairly than manufacturing workloads.

The strategy is classically Meta: subsidize entry, harvest knowledge at scale, and use it to shut the hole with the frontier. Zuckerberg made no secret of the ambition, calling Muse Spark 1.2 "our subsequent step as we push towards frontier, with bigger, extra succesful fashions on the way in which."

Nevertheless, for builders and enterprises who need or are required legally to maintain their code safe, the tradeoff will not be one they're prepared or capable of make.

No Llama in sight

What right this moment's announcement conspicuously lacks is any point out of open supply — a placing omission from the corporate that spent three years positioning itself because the standard-bearer of open AI.

From the unique LLaMA's debut in February 2023 — whose weights famously leaked onto 4chan inside weeks, inadvertently kickstarting the motion to run succesful fashions on shopper {hardware} — by way of Llama 2's commercially usable license, the coding-specialized Code Llama, and the 405-billion-parameter Llama 3.1, which Zuckerberg launched in July 2024 with a manifesto titled "Open Source AI Is the Path Forward," Meta's whole pitch to builders was that frontier-class weights needs to be free to obtain, self-host, and fine-tune.

The technique labored: by early 2026, the Llama household had been downloaded roughly 1.2 billion times, averaging about 1,000,000 downloads a day, with self-hosting providing enterprises value reductions VentureBeat has beforehand reported at as a lot as 88% versus proprietary API suppliers.

Then got here the unraveling. Llama 4 debuted in April 2025 to mixed reviews and, ultimately, admissions that its benchmark outcomes had been fudged — whereas Chinese language open-weight rivals from DeepSeek, Alibaba, and Zhipu AI surged to account for some 41% of downloads on Hugging Face by late 2025, eroding Llama's declare to management of the very motion it began. The rocky rollout spurred Zuckerberg's summer time 2025 overhaul of Meta's AI operations into Meta Superintelligence Labs (MSL), with Scale AI co-founder Alexandr Wang recruited as chief AI officer.

The Llama period successfully ended this previous April 8, when MSL shipped the original Muse Spark — "essentially the most highly effective mannequin that meta has launched," in Wang's phrases — as Meta's first proprietary mannequin: cloud-only, with no downloadable weights and no self-hosting, initially confined to Meta's apps and a personal API preview.

Requested instantly on the time whether or not Llama growth would proceed, a Meta spokesperson advised VentureBeat solely that "our present Llama fashions will proceed to be accessible as open supply" — pointedly silent on future ones.

Wang, for his half, stated bigger models were already in development "with plans to open-source future versions" — however 4 months on, right this moment's launch does nothing to advance that promise: no weights, no license, and neither the weblog publish nor Zuckerberg's thread a lot as makes use of the phrase "open."

The reversal is all of the sharper as a result of Meta's rivals have been transferring in the other way. OpenAI launched its Codex CLI as open source below the permissive, enterprise-friendly Apache 2.0 license and adopted with its gpt-oss open-weight models; Google's Gemini CLI harness is likewise Apache-licensed.

With Muse Code, Meta lands closest to the posture of Anthropic — whose Claude Code stays proprietary — whereas the corporate that when argued open supply was the trail ahead now asks builders to pay per token for a mannequin they can’t examine, or to subsidize that entry with their very own knowledge.

Seen in that gentle, the contributor tier reads because the successor to the Llama technique itself: the ecosystem flywheel is now not free weights in alternate for mindshare, however low cost tokens in alternate for coaching knowledge.

However Zuck's reply on X — requested instantly by AI developer Luckey Farady, "Will Muse Code be open supply?" he responded "I'll have extra to share on that quickly" — does preserve hope alive that Meta will return to the open supply AI ballgame.

Why it issues

Terminal coding brokers have grow to be the fastest-growing floor in enterprise AI, and till right this moment the class has successfully been a two-horse race between Anthropic and OpenAI, with Google and a crowd of startups in pursuit.

Meta's entry brings a genuinely completely different structure (persistent background brokers, an append-only native occasion log), a reputable long-horizon demo, and an aggressive pricing wedge.

The open questions are those benchmarks charts can't reply: whether or not Muse Spark 1.2 truly matches Claude and GPT-class fashions on real-world repositories, whether or not builders belief Meta with their code, and whether or not the contributor tier's low cost is sufficient to make them cease asking. Muse Code is on the market in beta right this moment; Muse Spark 1.2 is stay within the Meta Mannequin API with expanded world entry.

Source link