---
license: apache-2.0
language:
- en
library_name: gguf
pipeline_tag: text-generation
tags:
- gguf
- llama.cpp
- muse-glimmer
- turboquant
- tq3_4s
- vision
- image-text-to-text
base_model:
- unsloth/Muse-Glimmer-30B-GGUF
---

# Muse-Glimmer-30B-TQ3_4S

![Muse-Glimmer-30B-TQ3_4S](thumbnail.png)

## Required Runtime

This model uses the custom `TQ3_4S` tensor type. It requires
[turbo-tan/llama.cpp-tq3](https://github.com/turbo-tan/llama.cpp-tq3).
Stock `llama.cpp` builds without TurboQuant support cannot load it.

## Model Files

| File | Size | Purpose |
|---|---|---|
| `Muse-Glimmer-30B-TQ3_4S.gguf` | 13.78 GiB | Main model (4.25 bpw) |
| `mmproj-Muse-Glimmer-30B-Q8_0.gguf` | 2.0 GiB | Vision projection (image input) |

## Base Model

- Upstream parent: `unsloth/Muse-Glimmer-30B-GGUF` (from `meta-models/Muse-Glimmer-30B`, Apache-2.0)
- Quantization: TurboQuant `TQ3_4S` (four-scale turbo quant) with out6k recipe —
  output and embedding tensors preserved at q6_K precision
- Native context: 131,072 tokens

## Recommended Runtime

Text-only:

```bash
./build/bin/llama-server \
  -m Muse-Glimmer-30B-TQ3_4S.gguf \
  --host 127.0.0.1 --port 8080 \
  -c 32768 -np 1 -ngl 99 -fa on \
  --reasoning-format deepseek --jinja
```

With vision (mmproj):

```bash
./build/bin/llama-server \
  -m Muse-Glimmer-30B-TQ3_4S.gguf \
  --mmproj mmproj-Muse-Glimmer-30B-Q8_0.gguf \
  --host 127.0.0.1 --port 8080 \
  -c 32768 -np 1 -ngl 99 -fa on \
  --reasoning-format deepseek --jinja
```

Optional — DFlash speculative decoding (raises decode ~20%):

```bash
# separate drafter model required
--spec-type draft-dflash -md <drafter>.gguf --spec-draft-n-max 3
```

## Benchmarks

Measured on **NVIDIA RTX 3090 24 GB**, `turbo-tan/llama.cpp-tq3` build `f755f1ac1`,
thinking ON, temperature 0.

![Benchmark summary](benchmark.png)

### Evalplus (official scorer)

| Benchmark | pass@1 |
|---|---:|
| HumanEval | **93.3** |
| HumanEval+ | **89.0** |
| MBPP | **89.7** |
| MBPP+ | **74.6** |

### Hard86

| Benchmark | Result |
|---|---|
| Hard86 | **74/86** (86.0%) |

### Task suites (task breakdown)

| Suite | Score | Pass rate |
|---|---:|---|
| instructfollow | 96.7 | 14/15 |
| coding | 87.5 | 10/12 |
| dataextract | 82.8 | 9/15 |
| reasonmath | 80.0 | 12/15 |
| toolcall | 80.0 | 11/15 |
| speed | 70.8 | 9/9 |

### Speed

| Config | Result |
|---|---:|
| llama-bench pp2048 | 1,155 tok/s |
| llama-bench tg128 | 43.3 tok/s |
| Decode, 8K context, no drafter | 44.6 tok/s |
| Decode, 8K context, DFlash drafter (n_max=3) | 53.7 tok/s (+20%) |

Generation throughput during task suites: 47.9 tok/s.

## Validation

- Strict server smoke (`--reasoning off`): content exactly `ok` ✅
- Drafter signature verified in server logs: `block_size=16, mask_token_id=201818, n_extract=5`

## License

Apache-2.0. Use is also subject to the base model license and the license terms of the runtime.