Meta (Muse Glimmer / Muse Spark / Llama): the main open-weight option for school self-hosting

Short answer: Meta now runs two tracks. The frontier Muse Spark line is closed and irrelevant to schools, while the model Hong Kong schools should actually look at is Muse Glimmer 30B, licensed under Apache 2.0 and small enough to run on one graphics card in a school server room.

What this is: the open-source standard-bearer that changed direction

For three years "open large language model" effectively meant Llama. When a school first heard that data could stay inside the campus network, the model in that conversation was Llama on a machine in the school's own server room. That picture needs updating in 2026.

In April 2026, after forming Meta Superintelligence Labs, Meta replaced Llama with Muse Spark as its frontier line, and Muse Spark is closed-weight, available only in Meta AI products and a limited private API preview. Meta's strongest model, in other words, can no longer be downloaded.

Meta has not abandoned open release altogether. Muse Glimmer 30B, published on 9 August 2026, is distilled from Muse Spark, ships under Apache 2.0, and was explicitly designed to run agents on your own machine. For a school, that single model is worth more than any Meta frontier release.

The current line-up: one closed lane, one open lane, and a legacy batch

• Muse Glimmer 30B (9 August 2026): a 30B dense multimodal model with a 131K-token context window, accepting text and image input, with official support for Ollama, LM Studio, llama.cpp, ExecuTorch and MLX. Quantised to 4-bit, the language model compresses to under 20GB and the documentation targets devices with 24GB or 32GB of VRAM.
• Muse Spark 1.1 / 1.2 (16 July and 5 August 2026): the closed frontier line with a 1M-token context window at $1.25 per 1M input tokens and $4.25 per 1M output tokens. This lane cannot be self-hosted.
• Llama 3.1 / 3.3 (from July 2024): an earlier generation, yet still the open model schools most often self-host on Ollama. Edor.ai's on-premises default is llama3.1.

The full comparison of versions, dates and prices is in the automatically generated series table further down this page.

Strengths: the four that matter most to a school

1. Data stays inside the campus network. Self-hosting means every sentence a pupil types is processed by the school's own machine. There is no cross-border transfer and no privacy assessment needed for one. That is the easiest point to explain at a parents' evening.
2. Clean licence terms. Apache 2.0 imposes no user cap, no revenue threshold and no application process, so a sponsoring body can replicate one deployment across several schools.
3. The hardware bar is genuinely within reach. A 24GB VRAM card is a high-end consumer part, not a data-centre purchase, which makes "buy one machine and pilot it for a year" a realistic school-based plan.
4. A mature ecosystem. Ollama, LM Studio and llama.cpp all have official support, so an IT coordinator can find plenty of documentation and troubleshooting material rather than working it out alone.

Limits: what to say plainly to a principal

• The frontier lane is shut. Meta has said it hopes to open-source future Muse versions, but there is no timetable. A school should not assume the next frontier weights will be published, and should build its self-hosting plan on what exists now.
• Quality still trails cloud flagships. A 30B model is clearly behind GPT-6 Astra or Claude Fable 5.1 on long-form reasoning, Chinese writing feedback and local Hong Kong knowledge. Treat it as a capable everyday assistant rather than the best brain available.
• Hardware and staffing are the hidden costs. One card is the start; cooling, power, backups and version upgrades follow. For a school with no IT staffing headroom the total cost may not beat a cloud subscription.
• No speech, and limited vision. Meta states plainly that Muse Glimmer does not support audio input or output, and video is processed only as individual frames. Cantonese oral practice needs a separate solution.
• The model card says it is not intended for download or use by under-18s. That does not stop a school using it, but it does mean the student-facing side must sit behind safety gating and teacher monitoring, with the school owning the risk assessment.

Where each tier fits

Teachers: first drafts of lesson plans, variants of worksheet questions, condensing a long circular into a message for parents. These ask little of the model, a local 30B is enough, and nothing with a pupil's name in it leaves the school.

Students: suitable for routine practice questions and reading-comprehension hints. Use it in tutor mode so the platform holds it to hints rather than answers, with conversations a teacher can review.

IT coordinators: check the card model and VRAM first, then decide between llama3.1 8B and a quantised Muse Glimmer 30B. Run a one-month pilot, measure concurrent users and response latency, and size the hardware from that.

Prompts worth testing

Paste these into a Edor.ai teacher tool or any interface connected to a local model. Note the verification line at the end of each — a local model is more likely than a cloud flagship to state something wrong, so that line is not optional.

You are a Hong Kong primary Chinese-language teacher. Design a 20-minute reading comprehension worksheet on the theme of 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 the vocabulary at Primary 5 level; if anything touching law or statistics is uncertain, write "teacher to verify" rather than inventing it.

Below is a draft marking rubric for a Secondary 3 school-based assessment project. Check whether it is clear and workable.
Identify three places where the wording is vague, give one more specific rewrite for each, then list two requirements students are most likely to misread.
Constraints: write in Hong Kong English conventions; do not rewrite the whole rubric; do not add new criteria; if you are unsure about normal Hong Kong school-based assessment practice, mark it "teacher to verify".
[paste rubric]

Availability inside Edor.ai

Meta is not a native provider. Edor.ai natively supports five options — OpenAI, Azure OpenAI, Anthropic, Poe and local Ollama — and Meta models arrive through the last of those. A school runs Muse Glimmer 30B or llama3.1 with Ollama on its own server and sets the provider to local Ollama in the admin panel.

This is precisely how a school keeps student data inside the campus network. Conversations touch no external cloud, while four-layer gating, teacher monitoring, the ai_audit_logs trail and knowledge-base retrieval all continue to work, with identical features to the cloud version. The data still belongs to the school and can be exported in full at any time.

Going one step further, a school can fine-tune open weights on its own lesson plans and model answers. LoRA, QLoRA and INT4 quantisation are what make that possible on a single consumer GPU, and they are the same techniques taught in the LLM Classroom. Our on-premises option and the course cover the same ground, so a teacher who finishes the lessons can read what the school's own system is doing.

Alternatives

• For stronger Chinese-language handling from open weights, see Alibaba Qwen, where Qwen3.8-27B is also Apache 2.0 and runs in 24GB of VRAM.
• For an open option framed around European data governance, see Mistral AI.
• To avoid managing hardware at all, see the native providers OpenAI and Anthropic Claude.
• To work out the cost and privacy trade-off before buying anything, read the open versus closed cost and privacy trade-off or start at the LLM Classroom.