Skip to content

Commit 129f834

Browse files
localai-org-maint-botlocalai-org-maint-bot
authored
feat(gallery): add Muse Glimmer 30B variants (#11446)
Add the official dynamic and 17 GB K-quant GGUF builds with the shared multimodal projector and optional DFlash acceleration. Assisted-by: Codex:gpt-5 Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com>
1 parent 0df63de commit 129f834

1 file changed

Lines changed: 205 additions & 0 deletions

File tree

‎gallery/index.yaml‎

Lines changed: 205 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,4 +1,209 @@
11
---
2+
- &muse-glimmer-30b
3+
name: "muse-glimmer-30b"
4+
variants:
5+
- model: muse-glimmer-30b-dflash
6+
- model: muse-glimmer-30b-17gb
7+
- model: muse-glimmer-30b-17gb-dflash
8+
url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
9+
urls:
10+
- https://huggingface.co/meta-models/Muse-Glimmer-30B
11+
- https://huggingface.co/meta-models/Muse-Glimmer-30B-GGUF
12+
description: |
13+
Muse Glimmer is Meta Superintelligence Labs' Apache-2.0 dense 30B model for
14+
autonomous agentic work, coding, tool use, long-horizon reasoning, and
15+
multimodal understanding. It supports more than 100 languages, interleaved
16+
text and image input through its 1.8B-parameter perception encoder, and a
17+
131K-token context window.
18+
19+
This entry uses the publisher's higher-quality dynamic K-quant GGUF and
20+
official quantized vision projector. Automatic variant selection can use
21+
the smaller 17 GB quantization or a DFlash-accelerated build when it fits.
22+
license: "apache-2.0"
23+
tags:
24+
- llm
25+
- gguf
26+
- cpu
27+
- gpu
28+
- vision
29+
- multimodal
30+
- reasoning
31+
- coding
32+
- agent
33+
- tools
34+
- multilingual
35+
- long-context
36+
last_checked: "2026-08-10"
37+
overrides:
38+
backend: llama-cpp
39+
function:
40+
automatic_tool_parsing_fallback: true
41+
grammar:
42+
disable: true
43+
known_usecases:
44+
- chat
45+
- vision
46+
mmproj: llama-cpp/mmproj/muse-glimmer-30b/mmproj-kquant.gguf
47+
options:
48+
- use_jinja:true
49+
parameters:
50+
model: llama-cpp/models/muse-glimmer-30b/muse-glimmer-30B-kquant-dynamic.gguf
51+
temperature: 1.0
52+
top_k: 64
53+
top_p: 0.95
54+
template:
55+
use_tokenizer_template: true
56+
files:
57+
- filename: llama-cpp/models/muse-glimmer-30b/muse-glimmer-30B-kquant-dynamic.gguf
58+
uri: huggingface://meta-models/Muse-Glimmer-30B-GGUF/muse-glimmer-30B-kquant-dynamic.gguf
59+
sha256: 513109c8319115f69eb09fb7b118c97c8167d15bc014fd7670d2e30489bf106c
60+
- filename: llama-cpp/mmproj/muse-glimmer-30b/mmproj-kquant.gguf
61+
uri: huggingface://meta-models/Muse-Glimmer-30B-GGUF/mmproj-kquant.gguf
62+
sha256: f48b452316f9b213758e8659444029b961a24a07f99a1abb2a9f88b06f7c00c6
63+
- !!merge <<: *muse-glimmer-30b
64+
name: "muse-glimmer-30b-dflash"
65+
variants: []
66+
description: |
67+
Muse Glimmer's higher-quality dynamic K-quant GGUF with the official
68+
quantized perception encoder and DFlash drafter. DFlash proposes blocks of
69+
up to 16 tokens for the target to verify in parallel, accelerating output
70+
without changing model quality. Flash attention is enabled for this path.
71+
tags:
72+
- llm
73+
- gguf
74+
- gpu
75+
- vision
76+
- multimodal
77+
- reasoning
78+
- coding
79+
- agent
80+
- tools
81+
- multilingual
82+
- long-context
83+
- dflash
84+
- speculative
85+
overrides:
86+
backend: llama-cpp
87+
flash_attention: "on"
88+
draft_model: llama-cpp/models/muse-glimmer-30b/dflash-kquant.gguf
89+
function:
90+
automatic_tool_parsing_fallback: true
91+
grammar:
92+
disable: true
93+
known_usecases:
94+
- chat
95+
- vision
96+
mmproj: llama-cpp/mmproj/muse-glimmer-30b/mmproj-kquant.gguf
97+
options:
98+
- use_jinja:true
99+
- spec_type:draft-dflash
100+
- spec_n_max:15
101+
parameters:
102+
model: llama-cpp/models/muse-glimmer-30b/muse-glimmer-30B-kquant-dynamic.gguf
103+
temperature: 1.0
104+
top_k: 64
105+
top_p: 0.95
106+
template:
107+
use_tokenizer_template: true
108+
files:
109+
- filename: llama-cpp/models/muse-glimmer-30b/muse-glimmer-30B-kquant-dynamic.gguf
110+
uri: huggingface://meta-models/Muse-Glimmer-30B-GGUF/muse-glimmer-30B-kquant-dynamic.gguf
111+
sha256: 513109c8319115f69eb09fb7b118c97c8167d15bc014fd7670d2e30489bf106c
112+
- filename: llama-cpp/mmproj/muse-glimmer-30b/mmproj-kquant.gguf
113+
uri: huggingface://meta-models/Muse-Glimmer-30B-GGUF/mmproj-kquant.gguf
114+
sha256: f48b452316f9b213758e8659444029b961a24a07f99a1abb2a9f88b06f7c00c6
115+
- filename: llama-cpp/models/muse-glimmer-30b/dflash-kquant.gguf
116+
uri: huggingface://meta-models/Muse-Glimmer-30B-GGUF/dflash-kquant.gguf
117+
sha256: 27d9a805fa29b943cfb6ad4843367cd4eaaaf06bd452d8cc3e00a2cd18a677bc
118+
- !!merge <<: *muse-glimmer-30b
119+
name: "muse-glimmer-30b-17gb"
120+
variants: []
121+
description: |
122+
Muse Glimmer's smaller 17 GB K-quant GGUF with the official quantized
123+
perception encoder. It preserves the model's agentic, coding, tool-use,
124+
multilingual, and image-understanding capabilities for hosts with less
125+
memory than the dynamic quantization requires.
126+
overrides:
127+
backend: llama-cpp
128+
function:
129+
automatic_tool_parsing_fallback: true
130+
grammar:
131+
disable: true
132+
known_usecases:
133+
- chat
134+
- vision
135+
mmproj: llama-cpp/mmproj/muse-glimmer-30b/mmproj-kquant.gguf
136+
options:
137+
- use_jinja:true
138+
parameters:
139+
model: llama-cpp/models/muse-glimmer-30b/muse-glimmer-30B-kquant-17gb.gguf
140+
temperature: 1.0
141+
top_k: 64
142+
top_p: 0.95
143+
template:
144+
use_tokenizer_template: true
145+
files:
146+
- filename: llama-cpp/models/muse-glimmer-30b/muse-glimmer-30B-kquant-17gb.gguf
147+
uri: huggingface://meta-models/Muse-Glimmer-30B-GGUF/muse-glimmer-30B-kquant-17gb.gguf
148+
sha256: 7e9b74b7c8875e9e265695df9613bf6290f2392e479ce740495a129019c488d8
149+
- filename: llama-cpp/mmproj/muse-glimmer-30b/mmproj-kquant.gguf
150+
uri: huggingface://meta-models/Muse-Glimmer-30B-GGUF/mmproj-kquant.gguf
151+
sha256: f48b452316f9b213758e8659444029b961a24a07f99a1abb2a9f88b06f7c00c6
152+
- !!merge <<: *muse-glimmer-30b
153+
name: "muse-glimmer-30b-17gb-dflash"
154+
variants: []
155+
description: |
156+
Muse Glimmer's smaller 17 GB K-quant GGUF with the official quantized
157+
perception encoder and DFlash drafter. This is the lowest-memory published
158+
build that retains image understanding and block-speculative decoding.
159+
Flash attention is enabled for the DFlash path.
160+
tags:
161+
- llm
162+
- gguf
163+
- gpu
164+
- vision
165+
- multimodal
166+
- reasoning
167+
- coding
168+
- agent
169+
- tools
170+
- multilingual
171+
- long-context
172+
- dflash
173+
- speculative
174+
overrides:
175+
backend: llama-cpp
176+
flash_attention: "on"
177+
draft_model: llama-cpp/models/muse-glimmer-30b/dflash-kquant.gguf
178+
function:
179+
automatic_tool_parsing_fallback: true
180+
grammar:
181+
disable: true
182+
known_usecases:
183+
- chat
184+
- vision
185+
mmproj: llama-cpp/mmproj/muse-glimmer-30b/mmproj-kquant.gguf
186+
options:
187+
- use_jinja:true
188+
- spec_type:draft-dflash
189+
- spec_n_max:15
190+
parameters:
191+
model: llama-cpp/models/muse-glimmer-30b/muse-glimmer-30B-kquant-17gb.gguf
192+
temperature: 1.0
193+
top_k: 64
194+
top_p: 0.95
195+
template:
196+
use_tokenizer_template: true
197+
files:
198+
- filename: llama-cpp/models/muse-glimmer-30b/muse-glimmer-30B-kquant-17gb.gguf
199+
uri: huggingface://meta-models/Muse-Glimmer-30B-GGUF/muse-glimmer-30B-kquant-17gb.gguf
200+
sha256: 7e9b74b7c8875e9e265695df9613bf6290f2392e479ce740495a129019c488d8
201+
- filename: llama-cpp/mmproj/muse-glimmer-30b/mmproj-kquant.gguf
202+
uri: huggingface://meta-models/Muse-Glimmer-30B-GGUF/mmproj-kquant.gguf
203+
sha256: f48b452316f9b213758e8659444029b961a24a07f99a1abb2a9f88b06f7c00c6
204+
- filename: llama-cpp/models/muse-glimmer-30b/dflash-kquant.gguf
205+
uri: huggingface://meta-models/Muse-Glimmer-30B-GGUF/dflash-kquant.gguf
206+
sha256: 27d9a805fa29b943cfb6ad4843367cd4eaaaf06bd452d8cc3e00a2cd18a677bc
2207
- &qwen3-5-9b-defiant-fable
3208
name: "qwen3.5-9b-defiant-fable-mtp"
4209
variants:

0 commit comments

Comments
 (0)