Models

Muse-Glimmer-30B

A dense 30B at 2.47× on 8-bit (8-bit quality at about 4-bit speed), or 1.7× on the 4-bit build in ~26 GB.

Made by: Meta
Size: 30B
Architecture: Dense multimodal, sliding + full attention
Features: Tool callingThinking

Run it (4-bit)

mlx-dspark serve --model mlx-community/Muse-Glimmer-30B-4bit

Serves an OpenAI + Anthropic API on http://127.0.0.1:8080. For a one-off answer, use mlx-dspark generate --model mlx-community/Muse-Glimmer-30B-4bit --prompt "…".

1.74× faster than plain decoding, mean of chat, code and math

22–27 tokens per second (plain: 14.0)

2.45 tokens accepted per round

~26 GB peak memory at chat length

Chat 1.57×

Code 1.70×

Math 1.94×

Dashed: plain decoding (1×). Solid: mlx-dspark on the same prompt, draft cap 2.

Drafter

DaoCloud/Muse-Glimmer-30B-DSpark resolves automatically with --mode auto (the default).

The drafter downloads with the model the first time you run it, and loads 4-bit quantized (acceptance doesn't depend on the drafter's precision). Lookup drafts are off by default for this pair (measured as a net loss); --lookup-drafts turns them on.

Measured 2026-08-12 on an M4 Pro, 48 GB with mlx 0.32.0. This predates the September 2026 verify kernels, which raised every 8-bit model re-measured since, so it is probably conservative.

Run it (8-bit)

mlx-dspark serve --model mlx-community/Muse-Glimmer-30B-8bit

Serves an OpenAI + Anthropic API on http://127.0.0.1:8080. For a one-off answer, use mlx-dspark generate --model mlx-community/Muse-Glimmer-30B-8bit --prompt "…". Resolves the same drafter as the 4-bit build.

2.47× faster than plain decoding, mean of chat, code and math

16–25 tokens per second (plain: 8.2)

3.31 tokens accepted per round

~40 GB peak memory at chat length

Chat 1.97×

Code 2.45×

Math 2.99×

Dashed: plain decoding (1×). Solid: mlx-dspark on the same prompt, draft cap 4.

Drafter

DaoCloud/Muse-Glimmer-30B-DSpark resolves automatically with --mode auto (the default).

The drafter downloads with the model the first time you run it, and loads 4-bit quantized (acceptance doesn't depend on the drafter's precision).

Measured 2026-08-12 on an M4 Pro, 48 GB with mlx 0.32.0. This predates the September 2026 verify kernels, which raised every 8-bit model re-measured since, so it is probably conservative.

Meta's Muse-Glimmer-30B is a dense, multimodal 30B with 3:1 sliding/full attention. It needs mlx-vlm ≥ 0.6.12, which a fresh install of mlx-dspark already requires. The community drafter by DaoCloud reuses both the target's embedding and its output head.

Which quant#

8-bit (~40 GB, fits a 48 GB Mac but tight) roughly doubles the 4-bit ratio: 2.47× mean at cap 4. It sits closer to the drafter's bf16 training verifier, and its verify curve stays flat to width 5. But 8-bit decode reads twice the bytes, so absolute speed is about the same as 4-bit on code. The better ratio buys 8-bit quality at about 4-bit speed.
4-bit (~26 GB) is the registry default for smaller Macs: 1.74× mean at cap 2.
bf16 (~60 GB) doesn't fit a 48 GB Mac.

Lookup drafts are off by default for this pair. Its output scaling and logit softcap make near-ties more frequent than on a typical model, so it diverges from single-row greedy at more positions. Every one of those divergences is a sub-ulp floating-point tie.

Numbers not matching your Mac? They shouldn't be identical: the draft cap is derived from your machine's measured curves. Run mlx-dspark benchmark --model mlx-community/Muse-Glimmer-30B-4bit --trials 3 to get your own. See how these are measured.