From 40fc2f9b9bc7dd69d2bd8f0cc4e9f6f6e39833d2 Mon Sep 17 00:00:00 2001 From: Colin Date: Tue, 29 Jul 2025 00:12:57 +0000 Subject: [PATCH] ollama: document how good/bad more of these models are --- modules/services/ollama.nix | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/modules/services/ollama.nix b/modules/services/ollama.nix index 63a100b48..0daaf6cd1 100644 --- a/modules/services/ollama.nix +++ b/modules/services/ollama.nix @@ -41,14 +41,14 @@ let deepseek-r1-abliterated-32b deepseek-r1-abliterated-70b devstral-24b - dolphin3-8b + dolphin3-8b # gives incorrect RDMA RoCEv2 UDP port # dolphin-mistral-7b # UNCENSORED mistral; compliant # dolphin-mixtral-8x7b # about as fast as a 14b model, similar quality results. uncensored, but still preachy # falcon2-11b # code examples are lacking # gemma2-9b # fast, but not great for code # gemma2-27b # generates at 1word/sec, but decent coding results if you can wrangle it gemma3-12b - gemma3-27b + gemma3-27b # gives incorrect RDMA RoCEv2 UDP port gemma3n-e2b gemma3n-e4b # glm4-9b # it generates invalid code @@ -60,7 +60,7 @@ let llama3_2-uncensored-3b # llama3_3-70b # non-compliant; dodges iffy questions llama3_3-abliterated-70b # compliant, but slower and not as helpful as deepseek-r1-abliterated-70b - llama4-16x17b + llama4-16x17b # gives incorrect RDMA RoCEv2 UDP port magicoder-7b # it generates valid, if sparse, code magistral-24b marco-o1-7b # untested @@ -79,7 +79,7 @@ let # qwen2_5-7b # notably less quality than 32b (i.e. generates invalid code) # qwen2_5-14b # *almost* same quality to 32b variant, but faster qwen3-8b - qwen3-14b + qwen3-14b # gives correct RDMA RoCEv2 UDP port qwen3-30b # qwen2_5-32b-instruct-q2_K # lower-res version of default 32b (so, slightly faster, but generates invalid code where the full res generates valid code) qwen2_5-32b # generates 3~5 words/sec, but notably more accurate than coder-7b