Meta released Muse Code on August 5, 2026, described by CEO Mark Zuckerberg in a social media post as able to "accomplish 'complete software engineering tasks across large repos,'" according to TechCrunch (E1). The tool is currently available in beta (E1). Zuckerberg said the agent's tasks include "planning changes, writing code, validating the results" (E2).
According to TechCrunch, Muse Code can be installed with a single command and is powered by Meta's previously released coding model, Muse Spark (E3). VentureBeat's own hands-on test on a Mac mini confirmed the one-line installer worked as advertised, downloading a 97 MB package and prompting a sign-in — but the agent stopped short of running any task, reporting that no models were visible and that payment was "required to finish setting up your account" (E20).
Zuckerberg said that in testing, Muse Code "build six features for a game simultaneously with no collisions," according to TechCrunch (E4). VentureBeat reported that Muse Code is a full harness, installable on macOS or Linux with a single curl command, co-trained with the model that powers it, and — like the Muse Spark models behind it — entirely proprietary (E8).
Per Meta's blog post as cited by VentureBeat, Muse Code uses "persistent async background agents" that "remain active throughout each session, rather than being spawned for individual tasks, helping avoid redundant information gathering," carrying out next steps on their own and choosing when to report back to the main agent (E10). VentureBeat also reported that every model call, tool run, approval, and edit is appended to a local event log before execution — a single source of truth that Meta says makes the runtime "replay-exact and restart-safe" (E11).
VentureBeat compiled benchmark results comparing Muse Spark 1.2 (running in Muse Code) against rival models:
| Benchmark | Muse Spark 1.2 | Rival scores |
|---|---|---|
| SWE-Bench Verified (original Muse Spark, April 2026) | 77.4 | Claude Opus 4.6: 80.8; Gemini 3.1 Pro: 80.6 (E12) |
| Terminal-Bench 2.1 | 82.9% | GPT-5.6 Terra (Codex): 81.8%; Grok 4.5 (Grok Build): 81.6%; Claude Opus 5 (max effort, Claude Code): 86.7% (E13) |
| DeepSWE 1.1 | 59.3% (3rd place) | Opus 5: 65.0%; GPT-5.6 Terra: 64.8% (E14) |
| Meta internal coding benchmark | 70.6% | GPT-5.6 Terra: 65.4%; Gemini 3.6 Flash: 63.9%; Opus 5: 79.4% (E15) |
Across every benchmark VentureBeat cited, Claude Opus 5 posted the top score. VentureBeat also reported that Muse Spark 1.2 improves on the prior 1.1 version by 6.7 points on Terminal-Bench and 6.3 points on DeepSWE (E16). Separately, Meta pointed Muse Spark 1.2 at GPU kernel optimization and let it run for more than 1,000 tool calls over up to 24 hours on NVIDIA Hopper hardware, according to VentureBeat (E17).
VentureBeat reported two distinct pricing tiers for Muse Spark 1.2:
| Tier | Input (per million tokens) | Output (per million tokens) | Cached input | Rate limit | Data use |
|---|---|---|---|---|---|
| Standard | $1.25 | $4.25 | $0.15 | 3,000 requests/min | Not used to train Meta's models (E18) |
| Contributor | $0.10 | $0.20 | $0.002 | 60 requests/min | Prompts and completions may be used to train future Meta models (E19, E21) |
VentureBeat noted the contributor tier is roughly 12x and 21x cheaper than standard on input and output tokens respectively, in exchange for explicit permission to use prompts and completions for training (E19). VentureBeat added that the contributor tier's much tighter rate limit — 60 requests per minute versus 3,000 for standard — signals it's aimed at individuals and small experiments rather than production workloads (E21).
Alexandr Wang, Meta's AI chief who leads Meta Superintelligence Labs, told the Wall Street Journal that Muse Code "can be an incredibly good option, especially from a cost perspective," for a lot of workflows and use cases, as reported by TechCrunch (E5). TechCrunch also reported that in June 2026, Meta expanded beyond its core advertising-focused AI use and entered the enterprise AI market with an agent aimed at customer service and support (E6).
According to VentureBeat, Meta's release of Muse Code in beta alongside Muse Spark 1.2 puts the company in direct competition with Anthropic's Claude Code, OpenAI's Codex, and the broader field of agentic coding harnesses (E7).
Asked on X whether Muse Code would be open-sourced, Zuckerberg replied "I'll have more to share on that soon," with VentureBeat noting he teased that open source may be in the cards for Muse Spark or another product entirely (E9).
VentureBeat reported that by early 2026, the Llama model family had been downloaded roughly 1.2 billion times, averaging about a million downloads a day (E22). By late 2025, however, Chinese open-weight models from DeepSeek, Alibaba, and Zhipu AI accounted for some 41% of downloads on Hugging Face, which VentureBeat said eroded "Llama's claim to leadership of the very movement it started" (E23).
VentureBeat dated the end of the Llama open-source era to April 8, 2026, when Meta Superintelligence Labs shipped the original Muse Spark — described by Wang as "the most powerful model that Meta has released" — as Meta's first proprietary model: cloud-only, with no downloadable weights and no self-hosting (E24). Asked at the time whether Llama development would continue, a Meta spokesperson told VentureBeat only that "our current Llama models will continue to be available as open source," staying silent on future ones (E25).
The evidence points to a company competing on price and workflow persistence rather than raw benchmark leadership: Wang frames Muse Code's appeal around cost (E5), and the contributor tier undercuts the standard tier by roughly 12x on input and 21x on output tokens (E19) — yet on every benchmark VentureBeat cited, Muse Spark 1.2 trails Claude Opus 5, from 82.9% versus 86.7% on Terminal-Bench 2.1 to 70.6% versus 79.4% on Meta's own internal benchmark (E13, E15). That cost-first posture coincides with a retreat from the open-weight strategy that built Llama's roughly 1.2-billion-download base (E22): Muse Spark and Muse Code are both proprietary (E8, E24), even as Chinese open-weight models took roughly 41% of Hugging Face downloads by late 2025 (E23) and Meta's spokesperson and CEO both left the door only narrowly open on future open-source releases (E9, E25).
Muse Code is a terminal-based AI coding agent Meta released in beta on August 5, 2026, for handling software engineering tasks — planning changes, writing code, and validating results — across large code repositories. It installs with a single command and is powered by Meta's Muse Spark model.
The standard tier costs $1.25 per million input tokens and $4.25 per million output tokens ($0.15 for cached input), with a 3,000 requests-per-minute limit, and prompts are not used for training. The contributor tier costs $0.10 per million input tokens and $0.20 per million output tokens ($0.002 cached), capped at 60 requests per minute, in exchange for allowing Meta to use the data for training.
Meta has not confirmed this. When asked on X, Zuckerberg said only "I'll have more to share on that soon," and a Meta spokesperson previously said only that current Llama models will remain open source, without addressing future models.
According to Intel's newsroom announcement, the company is investing €5 billion ($5.7 billion) to expand its Leixlip campus in Ireland, producing Intel Xeon 6 and next-generation Xeon processors on the Intel 3 node. Intel says the capital programme began earlier this year and builds on more than €30 billion invested in Ireland since 1989.
NVIDIA has moved Alpamayo 2 Super into commercial use, according to NVIDIA's official blog. The autonomous-driving reasoning model runs on Cosmos 3 Super Reasoner, holds three times the parameters of its 10-billion-parameter predecessors, and ships under the OpenMDW-1.1 license, which ITHome reports NVIDIA has extended across the entire Alpamayo model family for fine-tuning and commercial redistribution.
Asian equities sold off again on August 6, according to Liberty Times (自由時報): South Korea's KOSPI opened down 1.81% and later retested a low of 6,379.35, off 3.31%, while Japan's Nikkei 225 broke below the 66,000 level, falling as much as 1.64% to 65,208. The declines followed a mixed Wall Street close on August 5, when the Philadelphia Semiconductor Index (SOX) dropped 1.4%. According to Anue (鉅亨網), a similar episode occurred on June 14, when Trump's move to resume an Iran blockade drove oil prices up nearly 10% and briefly sent KOSPI down close to 3% and the Nikkei down more than 1,000 points before both indexes rebounded.