Meta (Muse Glimmer / Muse Spark / Llama): open weights still self-host; the frontier lane went closed

Meta is still the open-source name Hong Kong schools hear most often in "keep data on our network" conversations; the 2026 reality is dual-track — the strongest Muse Spark tier is closed and cannot be downloaded, so procurement that still treats Llama as the frontier picks the wrong generation.

For three years, many IT coordinators explained self-hosted large models to a principal by pointing at Llama on a school server. After Meta Superintelligence Labs formed in April 2026, the frontier line became closed-weight Muse Spark. What actually matters to schools is Muse Glimmer 30B, published on 9 August under Apache 2.0 — 30B parameters, runnable on a single 24GB card. That is the executable open plan today. Procurement documents that still say "Meta Llama frontier" without naming Glimmer or llama3.1 are already out of date.

What Meta is actually building

Meta's AI strategy in 2026 splits cleanly: a closed frontier lane (Muse Spark) for its own products and a limited API preview; an open deployment lane (Muse Glimmer, Llama 3.x) for institutions running models on their own hardware. Schools should care only about the second lane — however strong the frontier benchmarks, if weights cannot be downloaded, the school platform cannot use them.

Muse Glimmer 30B is distilled from Muse Spark and was explicitly designed to run agents on your own machine, with official support for Ollama, LM Studio, llama.cpp, ExecuTorch and MLX. A mature ecosystem means an IT team can find deployment documentation rather than inventing the stack from scratch.

Models a school procurement will encounter

Meta documents Muse Glimmer for devices with 24GB or 32GB of VRAM. Llama 3.1 8B sits lower and suits a pilot before upgrading to 30B.

Where the money actually goes

Self-hosted weights: Apache 2.0 download carries no licence fee and no user cap, so a sponsoring body can replicate one deployment across several schools. Hidden costs are hardware and staffing — one 24GB card is the start; cooling, power, backups and version upgrades follow. A school with no IT headroom may not beat a cloud subscription on total cost.

Cloud Muse Spark: $1.25 / $4.25 and not self-hostable, with data leaving the campus network — usually misaligned with a "data stays on site" procurement goal.

Meta has said it hopes to open-source future Muse versions, but there is no timetable. Build the self-hosting plan on what exists now, not on a bet that the next frontier weights will be published.

How it fits inside Edor.ai

Meta is not a native provider and arrives through local Ollama. Edor.ai natively supports OpenAI, Azure OpenAI, Anthropic, Poe and local Ollama. A school runs Muse Glimmer 30B or llama3.1 with Ollama on its own server, sets the provider to local Ollama in the admin panel, and conversations touch no external cloud while four-layer gating, teacher monitoring, ai_audit_logs and knowledge-base retrieval carry on unchanged, with full export available at any time.

We usually suggest a term on cloud hosting first to confirm teacher usage and gating needs before buying a graphics card. Going further, LoRA, QLoRA and INT4 quantisation make fine-tuning on school lesson plans feasible on a consumer GPU — the same techniques taught in the LLM Classroom.

Classroom fit and hard limits

Teachers: first drafts, worksheet variants, condensing a long circular for parents. A local 30B is enough and nothing with a pupil's name leaves the school. On a Thursday after school, a Primary 5 class teacher in the staff room said: "I let the local model draft three differentiated worksheets — I still write the marking criteria myself."

Students: routine practice questions and reading-comprehension hints, tutor mode only, with the platform holding to hints rather than answers and conversations a teacher can review.

IT coordinators: confirm card model and VRAM, then choose llama3.1 8B or a quantised Muse Glimmer 30B; run a one-month pilot on concurrent users and latency before adding machines.

Does not fit: expecting a local 30B to replace a cloud flagship for high-stakes comments; letting pupils download weights individually; assuming Muse Glimmer handles Cantonese speech; publishing AI output externally without teacher review.

Copy this into a teacher tool wired to a local Meta model if you want a concrete trial:

You are a Hong Kong primary Chinese-language teacher. Design a 20-minute reading comprehension worksheet on school bullying for Primary 5.
Output: one short passage under 300 words, five comprehension questions (three factual, two inferential), a separate answer section, and three guiding questions for the teacher.
Constraints: use no real school names or personal details; keep vocabulary at Primary 5 level; if anything touching law or statistics is uncertain, write "teacher to verify" rather than inventing it.

Local models slip on detail more often than cloud flagships — keep the verification line.

The trade-off for schools

• If data must stay entirely on the campus network and you have a 24GB card, start with Muse Glimmer 30B or llama3.1 — do not chase the unpublished Muse Spark frontier lane.
• If hardware and IT staffing are both tight, run a cloud term first and use actual usage data to argue for or against self-hosting capital, rather than buying machines before you know the workload.
• If you need high-quality Chinese comments or Cantonese oral practice, keep an OpenAI or Azure key for high-stakes output and let local Meta models handle everyday tasks only.