Muse Glimmer 30B from Meta, Qwen3.8-27B from Alibaba and Gemma 4 31B from Google are three open models of the same size class, all under Apache 2.0 and so usable commercially. In 4 bit they take 18 to 20 GB in the Ollama library, which fits a GPU with 24 GB. For Muse Glimmer, Meta explicitly names 24 or 32 GB of GPU memory. According to the vendors, Qwen3.8-27B leads in agentic coding: on Terminal Bench 2.1, Qwen lists 73.0 for its own model and 51.7 for Muse Glimmer. Meta says Muse Glimmer is built for local agents with many tool calls, and it ships with a draft model for faster output. Gemma 4 is the broadest family, from phone-sized models up to the 31B version. Qwen has the longest context at 262,144 tokens natively. How the three perform on my DGX Spark is shown in my own test further down.
My test on the DGX Spark from November 3 and 4, 2026: of the models in this comparison, Gemma 4 26B-A4B was fastest at 68.0 tokens per second on German prose. Qwen3.8-27B reached 18.2 with its DFlash2 draft model and 50.1 on code, Muse Glimmer 12.2. Of the six German test tasks, Muse Glimmer, Gemma 4 31B and Qwen3.8-27B without a draft model passed five each.
I run my business with a team of fifteen AI employees built on Claude. Next to it, Qwen3.8-27B has been running on my DGX Spark as a local model since August, measured at 50.7 tokens per second in August 2026 and at 18.2 on German prose in November. Muse Glimmer is the obvious challenger for me: same size, same license, explicitly built for local agents. Everything I write about local AI is collected on my AI page.
The three models at a glance
Gemma 4 also comes as a mixture-of-experts variant, Gemma 4 26B-A4B, with 25.2 billion parameters of which 3.8 billion are active per token. Google writes that it runs "almost as fast as a 4B-parameter model". In 4 bit it takes 16 to 19 GB in Ollama. If speed matters more to you than quality, put it on the list too.
License: Apache 2.0 three times
According to their model cards, all three models come under Apache 2.0. That allows use in your own business and in products for clients, with no revenue cap and no duty to show a notice such as "Built with …". If you pass the model itself on, the license text and attribution notices go with it. For Muse Glimmer, Meta also points to its own usage policy in the repository. It lists prohibited uses, such as crime or collecting sensitive data about people without a legal basis, and states: "Muse Glimmer is not intended for individuals under the age of 18." I'm a developer, not a lawyer. This isn't legal advice, just the operator's perspective: my reading of the license texts.
Watch out for sibling models: not every Qwen model is under Apache 2.0. Qwen3.8-Flash-Next, for example, has its own license with special rules for providers of AI services. The differences in detail are covered in my article on open-weight licenses.
Hardware: what you need for the three
Meta writes on the Muse Glimmer card that the 4-bit version shrinks the language model to under 20 GB. That leaves enough room for context, vision encoder and draft model "within a 24 GB or 32 GB envelope". For full precision Meta lists 64 GB of GPU memory. Meta puts the quality loss of the 4-bit versions at 0.2 percent (32 GB version) and 1.0 percent (24 GB version), averaged over 15 benchmarks.
The file size is not the whole memory requirement. The context needs extra room, and according to its docs Ollama sets it based on GPU memory: 4k tokens below 24 GiB, 32k from 24 to 48 GiB and 256k from 48 GiB. So the 4-bit versions fit on a 24 GB card, but Ollama starts them there with a small default context. You can raise it as long as memory stays free.
There are two kinds of speed numbers. Meta measured Muse Glimmer in the 17 GB version at batch size 1: 74.9 tokens per second without and 233.4 with the DFlash draft model on an RTX 5090, 26.6 and 50.2 on an Apple M5 Max. On the Macs, Meta measured with ExecuTorch, on the RTX with llama.cpp, so not with Ollama. My own numbers for Qwen3.8-27B from August 2026: 50.7 tokens per second on the DGX Spark under SGLang in NVFP4, and 11 tokens per second on my MacBook Pro M3 Max with 64 GB, there in 8 bit through LM Studio. The two series use different machines and setups, so you can't read them directly against each other.
What the vendors measure
The values below come from the model cards. Each vendor measures with its own setup, and that makes a difference: for the same older model, Qwen3.6-27B, Meta's card lists 75.6 on OSWorld-Verified and Qwen's card 63.9. So read the table as a direction, not a ranking.
Qwen compares itself directly with Muse Glimmer on its card, and the Muse values there match Meta's own. Meta, on the other hand, compares Muse Glimmer with the predecessor Qwen3.6-27B, not with Qwen3.8. That is why there is no Qwen3.8 value for tool use on MCP Atlas, where Meta sees its model ahead.
Which model for which job, by vendor specs
Qwen3.8-27B has the highest values on the cards for coding and for operating user interfaces, and the longest context. Qwen names video as an explicit strength, up to hour-long recordings. On paper it is the first choice for coding agents on your own hardware.
Muse Glimmer 30B is, according to Meta, "purpose-built for autonomous agentic tasks on consumer hardware". Meta highlights tool calls with precise schemas and retrying after failed calls. Meta ships the DFlash draft model with it, and reasoning depth can be set in four levels from "low" to "xhigh".
Gemma 4 is the family with the widest choice, from E2B for phones to 31B for the workstation. According to Google, the 26B-A4B variant is the fast option. Only E2B, E4B and 12B take audio as input, not the 31B version.
My test on the DGX Spark
I measured all three models, plus Gemma 4 26B-A4B, on my DGX Spark on November 3 and 4, 2026, with the same setup as my test series for all candidates: 300 output tokens, one stream, median of three runs, thinking off or at the lowest level for all. Qwen ran under SGLang once with its DFlash2 draft model and once without, Muse Glimmer in Nvidia's NVFP4 version under vLLM without a draft model, the two Gemma models in 4 bit through Ollama. On top came six German test tasks from my daily work: a customer email from bullet points, a summary with check questions, tool calls with a deliberately broken response, a Python function with tests, an agent run in the terminal agent Pi, and a table from a screenshot. A measurement on the Mac is not part of this series. I describe the full setup and all individual results in Best LLMs for the DGX Spark, Measured.
¹ With my memory limit of 0.50, vLLM reserves about half of the machine. Muse Glimmer's weights take 23.4 GiB of that.
The runtimes differ: Gemma ran in 4 bit through Ollama, Qwen and Muse Glimmer in NVFP4. The numbers therefore apply to each whole setup, not as a pure model comparison.
How I read it: on coding, six tasks hardly separate the models. All four passed the Python task and the agent run in Pi, but Qwen with DFlash2 was fastest, 23 seconds against 152 for Muse Glimmer. Muse Glimmer kept Meta's promise of retrying after failed tool calls: after the broken response it tried again and only sent the email once the appointment had been created. The two Gemma models reported the error honestly but did not try again. On German, Gemma 4 31B came out ahead, with a passed customer email and all five check questions right. Muse Glimmer also answered all five questions correctly but invented an "order" in the email that did not exist. Both Qwen runs used the right calculation for one question and wrote down the wrong result. Muse Glimmer ran without its DFlash draft model because the time window on the machine was not long enough. How much faster it gets with it on the DGX Spark is not something this test shows.
Frequently asked questions
Is Muse Glimmer better than Qwen 3.8?
Not for coding, going by the vendor values: in every benchmark where Qwen compares both models, it lists higher values for Qwen3.8-27B, for example 73.0 against 51.7 on Terminal Bench 2.1. Meta presents Muse Glimmer as strong in tool use and agent tasks, but there it only compares against the predecessor Qwen3.6.
In my test on the DGX Spark, Muse Glimmer and Qwen3.8-27B without a draft model each passed five of the six German test tasks, Qwen with DFlash2 four.
Can I use Muse Glimmer commercially?
Yes, according to Meta, Muse Glimmer is under Apache 2.0. Meta's usage policy with prohibited uses applies on top, and you should read it once. This is not legal advice.
How much memory does Muse Glimmer need?
In 4 bit, according to Meta, under 20 GB for the language model, and 24 or 32 GB of GPU memory with everything around it. In Ollama the default tag is 18 GB, the 8-bit version 31 GB, full precision 57 GB.
Does Muse Glimmer run in Ollama?
Yes, under the name muse-glimmer. The library lists 18 GB, a 128K context and text and image as input for the default tag. The command is ollama run muse-glimmer.
Gemma 4 or Qwen 3.8?
For coding and long documents, the vendor values favor Qwen3.8-27B. For quick answers on little memory, the Gemma 4 variant 26B-A4B with only 3.8 billion active parameters makes the case.
On my DGX Spark, Gemma 4 26B-A4B reached 68.0 tokens per second on prose and 84.4 on code, Qwen3.8-27B with DFlash2 18.2 and 50.1.
Where to go from here
How a local Qwen holds up against Claude in daily work is covered in Qwen vs Claude: Local Qwen or Claude Code?. If you want to start on a Mac, read Local AI on a Mac: Mini as AI Server. My experience with the machine I test on is in DGX Spark Tested: Worth Buying?, and for an overview of all open models see my comparison of open source AI models. Which models in Ollama really run locally and which only in the cloud is covered in Ollama Models: Local or Cloud Only?.
