Can a 30B local model replace our hosted frontier model?

For the whole workload, no, and Meta does not claim it can. Muse Glimmer's peer group is Gemma4-31B and Qwen3.6-27B, not Claude or GPT. What it can do is take over the high-volume, low-difficulty portion of an agent loop, which in most production agents is the majority of calls and a minority of the difficulty. We measure which steps move, then route per step behind one abstraction.

What changed with the Apache 2.0 licence?

Meta shipped open weights for years under its own Llama Community Licence, which carried an acceptable-use policy, a monthly-active-user threshold, and a naming clause. Every one of those was a question a legal team had to answer first. Apache 2.0 answers them in advance: commercial use, fine-tuning, distillation, and redistribution are permitted outright. For anyone who has had a licence review stall a deployment, that is the most consequential part of the release.

What about Muse Spark 1.2?

Zuckerberg committed to opening its weights, and as of the Glimmer launch there was no date, no licence, and no artifact list. We plan builds around weights that exist. If Spark 1.2 lands under a permissive licence it will matter a great deal, and because everything sits behind a vendor-neutral abstraction, adding it later is a config change plus an eval pass.

★ Meta (Muse + Llama)LLM Models8-week engagement

META MUSE · OPEN-WEIGHT, LOCAL-FIRST

Meta's open weights, integrated properly. An agent model that runs on the machine it works for.

Meta Superintelligence Labs shipped Muse Glimmer under Apache 2.0 in August 2026: a ~29.6B dense multimodal model built for agentic work that fits under 20 GB at 4-bit. That puts a tool-using, screenshot-reading agent on one consumer GPU with no data leaving the building. We integrate it directly, behind a vendor-neutral abstraction and an eval suite, and we are honest about which steps it should and should not run.

Open weightsApache 2.0Local inferenceAgentic codingEval pipelines

Start a conversation →All llm models →

Cycle

8 weeks · fixed price

Stack

Muse Glimmer, local or hosted

Output

Production code + eval suite

Handoff

Full source ownership

[THE SHORT VERSION]

The first local model worth putting in a production agent loop.

Local models have been good enough to demo and not reliable enough to trust with tool calls. Muse Glimmer changes that inside its size class: Meta trained it for end-to-end task completion, precise tool calls, failure recovery, and reading screenshots and documents, and it leads Gemma4-31B and Qwen3.6-27B across the agentic benchmarks. It is not a frontier model and Meta does not claim it is. What it is, is the first genuinely capable agent model most teams can run without sending a byte to anyone, under a licence that raises no questions.

When it fits

High-volume, low-difficulty agent steps: classification, extraction, routing, screenshot reading, first drafts
Regulated, air-gapped, or on-device work where the data cannot leave the machine
Products that ship weights inside a customer install, where Apache 2.0 removes the licence review

When it does not

Hard, long-horizon reasoning where a hosted frontier model still wins outright
Low-volume workloads, where operating local inference costs more than the API ever would

[HOW WE BUILD IT]

How we build with Meta Muse.

Size the local versus hosted split

We measure which steps of your agent loop a 30B local model handles at your quality bar, and which have to escalate. The output is a routing policy backed by numbers, not a preference.

Direct integration, thin abstraction

Glimmer sits behind the same small provider interface as Claude and GPT, through llama.cpp, MLX, vLLM, or a hosted partner. Swapping it in or out is a config change plus an eval pass.

Eval the quantization, not the model

You deploy the 4-bit artifact, so that is what we test. Quantization damage lands first on long tool chains and schema adherence, which is exactly this model's job.

Own the runtime, not a vendor

GPU sizing, context budgeting, speculative decoding with the DFlash drafter, and observability on every call. Your team gets the runbook and the dashboards.

[WHAT YOU GET]

What the engagement leaves behind.

Apache 2.0

Fine-tune, distil, redistribute

Per token, self-hosted

Under 20 GB

One GPU, or a Mac

Eval-gated

Tested on the artifact you ship

[THE FAMILY]

What Meta has actually shipped.

Muse Glimmer is the one you can download today, under Apache 2.0. Muse Spark 1.2 has an open-weight commitment and no date. The Llama line is still the right pick for plenty of self-host builds. We integrate all of it behind one vendor-neutral abstraction, then route by task, sensitivity, and cost.

LatestOpen weights

10 Aug 2026

Muse Glimmer 30B

meta-models/Muse-Glimmer-30B

Capacity

~29.6B dense (1.8B vision)

Runs on

One 24-32 GB GPU, or a Mac

Licence

Apache 2.0

Apache 2.0 weights, so commercial use, fine-tunes, and redistribution need nobody's permission
Built for agents rather than chat: tool calls, multi-step recovery, and multimodal input on a single consumer GPU
Quantized to roughly 4-bit it fits under 20 GB, with a DFlash drafter that gives 3.1x decode on an RTX 5090

Read the technical brief

AnnouncedOpen weights

Weights announced

Muse Spark 1.2

meta/muse-spark-1.2

Capacity

Frontier foundation model

Runs on

Datacenter, or hosted API

Licence

Not yet announced

Meta's most advanced model, built by the Superintelligence Labs team formed in 2025
Zuckerberg has committed to opening the weights, but no date, no licence, and no artifact list yet
Available on hosted routers today; treat the open release as a plan, not a shipped thing

Supported · brief not yet published

Current

Earlier in 2026

Muse Spark 1.1

meta/muse-spark-1.1

Capacity

Multimodal reasoning model

Runs on

Hosted API only

Licence

Hosted terms

The first major model out of Meta Superintelligence Labs: multimodal reasoning, tool use, multi-agent orchestration
Listed on OpenRouter around $1.25 input and $4.25 output per million tokens
The teacher model in the family. Glimmer is the small, local, open sibling

Supported · brief not yet published

PreviousOpen weights

2024 to 2025

Llama (3.x and 4 line)

meta-llama/*

Capacity

8B to 405B

Runs on

Laptop to GPU fleet

Licence

Llama Community Licence

The line that made open-weight LLMs a default option rather than a curiosity
Shipped under Meta's own community licence, not a standard OSI one, which the Muse line has now dropped
Still the right pick for many self-host builds. We cover the deployment side on the Llama page

Supported · brief not yet published

[METHODOLOGY · K-FRAMEWORK]

Integrated through the
K-Framework.

Every model we integrate runs through the same operating system. Three pillars, sixteen layers, one Compound Growth Loop. The methodology that keeps AI work from rotting after the first ship.

Read the K-Framework

Foundations

Direct API integration with the model. No LangChain, no orchestration vendor, no agent framework built on quicksand. Typed contracts, the same way we wire up Postgres.

Amplification

An eval suite built from your real tasks gates every prompt and model change. Quality is measured before it ships, not vibed in a demo.

Judgment

Governance, audit, and oversight wired in from day one. Who called what, with which prompt version, at what cost. Your auditors get answers, not screenshots.

[OBSERVABILITY]

Observability your team can read.

A model in production without observability is roulette. We instrument every integration so engineering and finance can see the same numbers, and so a regression at 3am surfaces before a customer opens a ticket.

Instrumented

Cost per call

Tokens in, tokens out, dollars spent. Sliced by feature, tenant, and route. Budgets enforced where it matters.

Instrumented

Latency p50 / p95 / p99

Real distributions, not averages. We know which routes are slow, and why.

Instrumented

Eval pass rates

The same eval suite that gates a release runs continuously in production. A regression on real traffic surfaces fast.

Instrumented

Prompt + completion logs

PII scrubbed at the proxy, shipped to your SIEM. Retention controls match your compliance window.

Dashboards your team owns, not ours. At handoff you get the queries, the alerts, and the runbook. We are not in the path to read your metrics.

[COMMON QUESTIONS]

Questions we get asked.

Can a 30B local model replace our hosted frontier model?: For the whole workload, no, and Meta does not claim it can. Muse Glimmer's peer group is Gemma4-31B and Qwen3.6-27B, not Claude or GPT. What it can do is take over the high-volume, low-difficulty portion of an agent loop, which in most production agents is the majority of calls and a minority of the difficulty. We measure which steps move, then route per step behind one abstraction.
What changed with the Apache 2.0 licence?: Meta shipped open weights for years under its own Llama Community Licence, which carried an acceptable-use policy, a monthly-active-user threshold, and a naming clause. Every one of those was a question a legal team had to answer first. Apache 2.0 answers them in advance: commercial use, fine-tuning, distillation, and redistribution are permitted outright. For anyone who has had a licence review stall a deployment, that is the most consequential part of the release.
What about Muse Spark 1.2?: Zuckerberg committed to opening its weights, and as of the Glimmer launch there was no date, no licence, and no artifact list. We plan builds around weights that exist. If Spark 1.2 lands under a permissive licence it will matter a great deal, and because everything sits behind a vendor-neutral abstraction, adding it later is a config change plus an eval pass.