feat: Qwen3.8-27B als Haupt-Coder verdrahtet & Mem0 endgueltig abgeloest
Ampel / ampel (push) Successful in 22s

This commit is contained in:
Hitonabi
2026-08-19 17:23:54 +02:00
parent 5b0909089f
commit bf145a3f97
2 changed files with 9 additions and 6 deletions
+1 -1
View File
@@ -8,7 +8,7 @@
"moe": false, "quant": "Q4_K_XL", "ctx": 65536, "tools": true, "vision": true
},
{
"role": "fast", "name": "Qwen3.8-27B", "repo": "unsloth/Qwen3.8-27B-GGUF",
"role": "coder", "name": "Qwen3.8-27B", "repo": "unsloth/Qwen3.8-27B-GGUF",
"family": "qwen", "generation": 3.8, "total_params_b": 27, "active_params_b": 27,
"moe": false, "quant": "Q4_K_M", "ctx": 65536, "tools": true, "vision": true
},
+8 -5
View File
@@ -24,11 +24,13 @@ models:
Qwen3.8-27B:
# Qwen 3.8 27B: Dichtes 27B-Modell mit hybrider Linear-Attention (48/64 Schichten linear),
# nativem Multimodal-Support (mmproj-BF16) und Tool-Calling via --jinja.
# DFlash-2 Drafter (/srv/models/Qwen3.8-27B-DFlash2-GGUF/Qwen3.8-27B-DFlash2-Q8_0.gguf)
# liegt bereit und wird aktiviert, sobald upstream llama.cpp 81-Tensor-DFlash2 freigibt.
# Unser neuer Haupt-Coder mit 262k Kontext, Thinking Mode und extrem sauberer Code-Qualität.
cmd: |
llama-server -m /srv/models/Qwen3.8-27B-GGUF/Qwen3.8-27B-Q4_K_M.gguf --host 127.0.0.1 --port ${PORT} -c 131072 -ngl 999 -fa on --no-mmap --mmproj /srv/models/Qwen3.8-27B-GGUF/mmproj-BF16.gguf --jinja --parallel 2 -cram 16384 -ctk q8_0 -ctv q8_0
ttl: 300
ttl: 5400
aliases:
- coder
- dense-planer
capabilities:
in: [text, image]
out: [text]
@@ -49,11 +51,12 @@ models:
tools: true
context: 65536
Qwen3-Coder-Next:
# MoE-Coder Fallback (80B MoE, 3B aktiv).
cmd: |
llama-server -m /srv/models/Qwen3-Coder-Next-GGUF/Qwen3-Coder-Next-Q4_K_M.gguf --host 127.0.0.1 --port ${PORT} -c 131072 -ngl 999 -fa on --no-mmap --jinja --cache-reuse 256 -cram 16384
ttl: 5400
ttl: 1800
aliases:
- coder
- coder-moe
capabilities:
in: [text]
out: [text]