me/env-ansible
1
1 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

llm, refactor which models are still relevent to install.

Browse Source
This commit is contained in:
me committed 2026-11-01 05:47:15 -07:00
1 parent d3a5946232
commit 7efa7f1b91
1 file changed
+25 -31
Download Patch File Download Diff File Expand all files Collapse all files
tasks/workstation/linux/software/ai/llm.yml
+25 -31
View File
@@ -115,6 +115,16 @@
llm_removals:
##### QWEN #####
- { "purpose": ["chat"], "model": "qwen3:4b", "provider": "qwen_30_0040", "size_mb": 4000, "context": 123456 } # succeeded
### QWEN 3.0 Coder - Supposed to be better than 2.5 for machines that can handle it. Went okay, but better options now.
- { "purpose": ["code", "chat"], "model": "qwen3-coder:30b", "provider": "qwen_30_coder_0300_moe", "size_mb": "{{ llm_size_coef * 19000 }}", "context": 262144 }
- { "purpose": ["code", "chat"], "model": "qwen3-coder:420b", "provider": "qwen_30_coder_4200_moe", "size_mb": "{{ llm_size_coef * 290000 }}", "context": 262144 }
### QWEN 2.5 Coder - Ol'reliable! Did not seem to handle tools too well though. Defunitely not conversation.
- { "purpose": ["code"], "model": "qwen2.5-coder:0.5b", "provider": "qwen_25_coder_0005", "size_mb": "{{ llm_size_coef * 398 }}", "context": 32768 }
- { "purpose": ["code"], "model": "qwen2.5-coder:1.5b", "provider": "qwen_25_coder_0015", "size_mb": "{{ llm_size_coef * 986 }}", "context": 32768 }
- { "purpose": ["code"], "model": "qwen2.5-coder:3b", "provider": "qwen_25_coder_0030", "size_mb": "{{ llm_size_coef * 1900 }}", "context": 32768 }
- { "purpose": ["code"], "model": "qwen2.5-coder:7b", "provider": "qwen_25_coder_0070", "size_mb": "{{ llm_size_coef * 4700 }}", "context": 32768 }
- { "purpose": ["code"], "model": "qwen2.5-coder:14b", "provider": "qwen_25_coder_0140", "size_mb": "{{ llm_size_coef * 9000 }}", "context": 32768 }
- { "purpose": ["code"], "model": "qwen2.5-coder:32b", "provider": "qwen_25_coder_0320", "size_mb": "{{ llm_size_coef * 20000 }}", "context": 32768 }
##### Mistral #####
- { "purpose": ["code"], "model": "devstral:24b", "provider": "mistral_dev10_0240", "size_mb": "{{ llm_size_coef * 14000 }}", "context": 128000 } # succeeded
# https://ollama.com/library/mistral-small : Last Updated 2025-01-30, succeeded
@@ -122,11 +132,18 @@
- { "purpose": ["chat"], "model": "mistral-small:22b", "provider": "mistral_small30_0220", "size_mb": "{{ llm_size_coef * 13000 }}", "context": 128000 } # Yes, the smaller one has larger context.
# https://ollama.com/library/mistral-small3.1 : Last Update 2025-04-07, succeeded
- { "purpose": ["chat"], "model": "mistral-small3.1:24b", "provider": "mistral_small31_0240", "size_mb": "{{ llm_size_coef * 15000 }}", "context": 128000 }
# Supposedly ollama has it tagged at mistral-small with no params given. it's not on the site though. Let's find out!
# No, this is Mistral 3, as the website says. Ask Brave is wrooong.
- { "purpose": ["chat", "lang", "code"], "model": "mistral-small", "provider": "mistral_small40_1190", "size_mb": 60000 , "context": 256000 }
# https://ollama.com/library/mistral-nemo : Last Updated 2025-07-23
# Supposed to be a drop-in replacement for mistral:7b. Not really any more useful than the other options.
- { "purpose": ["chat", "lang"], "model": "mistral-nemo:12b", "provider": "mistral_nemo_0120", "size_mb": "{{ llm_size_coef * 7100 }}", "context": 128000 } # says 1000K in table but 128k in desc, going with desc.
llm_choices:
##### QWEN #####
### QWEN >3.5 - Need to try!
# Supposedly great for coding over any previous version.
# Prefer 27b (dense) over anything higher which likely uses MoE (essentially only uses 3b inside of a 35b, dumber but faster and potentially has access to more info).
# Supposedly great for coding over any previous version.
# Prefer 27b (dense) over anything higher which likely uses MoE (essentially only uses 3b inside of a 35b, dumber but faster and potentially has access to more info).
# Update: MoE actually runs really well on my workstation hardware, significantly faster than dense models, gives good results too.
# https://ollama.com/library/qwen3.8 : Last Updated 2026-09-25
# Known to overthink, suggested to prefer 3.6:27b unless setting temp to low/medium.
- { "purpose": ["code", "chat", "lang"], "model": "qwen3.8:27b", "provider": "qwen_38_0270", "size_mb": "{{ llm_size_coef * 18000 }}", "context": 262144 }
@@ -143,29 +160,17 @@
- { "purpose": ["chat", "code"], "model": "qwen3.5:27b", "provider": "qwen_35_0270", "size_mb": "{{ llm_size_coef * 17000 }}", "context": 262144 } # succeeded
- { "purpose": ["chat", "code"], "model": "qwen3.5:35b", "provider": "qwen_35_0350_moe", "size_mb": "{{ llm_size_coef * 22000 }}", "context": 262144 }
- { "purpose": ["chat", "code"], "model": "qwen3.5:122b", "provider": "qwen_35_1220_moe", "size_mb": "{{ llm_size_coef * 81000 }}", "context": 262144 }
### QWEN 3.0 Coder - Supposed to be better than 2.5 for machines that can handle it.
- { "purpose": ["code", "chat"], "model": "qwen3-coder:30b", "provider": "qwen_30_coder_0300_moe", "size_mb": "{{ llm_size_coef * 19000 }}", "context": 262144 }
- { "purpose": ["code", "chat"], "model": "qwen3-coder:420b", "provider": "qwen_30_coder_4200_moe", "size_mb": "{{ llm_size_coef * 290000 }}", "context": 262144 }
### QWEN 2.5 Coder - Ol'reliable!
- { "purpose": ["code"], "model": "qwen2.5-coder:0.5b", "provider": "qwen_25_coder_0005", "size_mb": "{{ llm_size_coef * 398 }}", "context": 32768 }
- { "purpose": ["code"], "model": "qwen2.5-coder:1.5b", "provider": "qwen_25_coder_0015", "size_mb": "{{ llm_size_coef * 986 }}", "context": 32768 }
- { "purpose": ["code"], "model": "qwen2.5-coder:3b", "provider": "qwen_25_coder_0030", "size_mb": "{{ llm_size_coef * 1900 }}", "context": 32768 }
- { "purpose": ["code"], "model": "qwen2.5-coder:7b", "provider": "qwen_25_coder_0070", "size_mb": "{{ llm_size_coef * 4700 }}", "context": 32768 }
- { "purpose": ["code"], "model": "qwen2.5-coder:14b", "provider": "qwen_25_coder_0140", "size_mb": "{{ llm_size_coef * 9000 }}", "context": 32768 }
- { "purpose": ["code"], "model": "qwen2.5-coder:32b", "provider": "qwen_25_coder_0320", "size_mb": "{{ llm_size_coef * 20000 }}", "context": 32768 }
##### Mistral, Ministral, and Devstral - By a French company. #####
# mistral-small-4
# https://huggingface.co/unsloth/Mistral-Small-4-119B-2603-GGUF
# Supposed to be better than Qwen equivalents, it's basically mistral, mixtral, magistral, and devstral all in one.
# unsloth version is supposed to be the best quantization. UD-Q3_K_XL is supposed to take 64GB RAM, lol, while Q4_K_M would need 80+ GB.
# needs done via 'ollama create', fails with pull. https://github.com/ollama/ollama/issues/5245
# https://huggingface.co/unsloth/Mistral-Small-4-119B-2603-GGUF
# Supposed to be better than Qwen equivalents, it's basically mistral, mixtral, magistral, and devstral all in one.
# unsloth version is supposed to be the best quantization. UD-Q3_K_XL is supposed to take 64GB RAM, lol, while Q4_K_M would need 80+ GB.
# needs done via 'ollama create', fails with pull. https://github.com/ollama/ollama/issues/5245
- { "purpose": ["chat", "lang", "code"], "model": "hf.co/unsloth/Mistral-Small-4-119B-2603-GGUF:UD-Q3_K_XL", "provider": "mistral_small40_1190_unsloth_q3", "size_mb": "{{ llm_size_coef * 54800 }}", "context": 256000 }
# https://huggingface.co/mistralai/Mistral-Small-4-119B-2603 : 242GB unquantized, lol.
# https://huggingface.co/mistralai/Mistral-Small-4-119B-2603 : 242GB unquantized, lol.
- { "purpose": ["chat", "lang", "code"], "model": "hf.co/mistralai/Mistral-Small-4-119B-2603", "provider": "mistral_small40_1190", "size_mb": "{{ llm_size_coef * 242000 }}", "context": 256000 }
# https://huggingface.co/unsloth/Mistral-Small-4-119B-2603 : 121GB, A6B MoE, hmmm!
# https://huggingface.co/unsloth/Mistral-Small-4-119B-2603 : 121GB, A6B MoE, hmmm!
- { "purpose": ["chat", "lang", "code"], "model": "hf.co/unsloth/Mistral-Small-4-119B-2603", "provider": "mistral_small40_1190_unsloth_a6b", "size_mb": "{{ llm_size_coef * 121000 }}", "context": 256000 }
# Supposedly ollama has it tagged at mistral-small with no params given. it's not on the site though. Let's find out!
- { "purpose": ["chat", "lang", "code"], "model": "mistral-small", "provider": "mistral_small40_1190", "size_mb": 60000 , "context": 256000 }
# https://ollama.com/library/devstral-small-2 : Last Updated 2025-12-13
# Most recommended in Mistral family for agentic coding.
- { "purpose": ["code"], "model": "devstral-small-2:24b", "provider": "mistral_dev20_0240_small", "size_mb": "{{ llm_size_coef * 15000 }}", "context": 393216 }
@@ -180,18 +185,7 @@
- { "purpose": ["chat", "lang"], "model": "ministral-3:3b", "provider": "mistral_mini30_0030", "size_mb": "{{ llm_size_coef * 3000 }}", "context": 256000 }
- { "purpose": ["chat", "lang"], "model": "ministral-3:8b", "provider": "mistral_mini30_0080", "size_mb": "{{ llm_size_coef * 6000 }}", "context": 256000 }
- { "purpose": ["chat", "lang"], "model": "ministral-3:14b", "provider": "mistral_mini30_0140", "size_mb": "{{ llm_size_coef * 9100 }}", "context": 256000 } # also rated super high for Croatian, gives more head room.
# https://ollama.com/library/mistral-nemo : Last Updated 2025-07-23
# Supposed to be a drop-in replacement for mistral:7b.
- { "purpose": ["chat", "lang"], "model": "mistral-nemo:12b", "provider": "mistral_nemo_0120", "size_mb": "{{ llm_size_coef * 7100 }}", "context": 128000 } # says 1000K in table but 128k in desc, going with desc.
##### TBD #####
### QWEN 2.5 - Older, but does not have "thinking", tempting!
#- qwen2.5:0.5b
#- qwen2.5:1.5b
#- qwen2.5:3b
#- qwen2.5:7b
#- qwen2.5:14b
#- qwen2.5:32b
#- qwen2.5:72b
### Z.ai
#- GLM-5.x
#- glm4-flash