Feedback - Agentic coding
First feedback based on my personal experience using OpenCode:
OrionLLM/GRM-2.6-Plus is easy to steer and it uses skills and mcp better but struggles with harder problems and makes simple mistakes sometimes.OrionLLM/GRM-2.6-Plus-0628 is harder to steer (sometimes it just tries to resolve the issue mentioned in context completely ignoring my current message). More common loops (in general prefers to use the same tools used before) in comparison to previous version and sometimes it overthink simple problems. Definitely better for harder problems.
I use OrionLLM/GRM-2.6-Plus and OrionLLM/GRM-2.6-Plus-0628 in q8_0 with config:
temp=0.6
top-p=0.95
top-k=20
min-p=0.00
Thanks for the feedback! GRM-2.6-Plus-0628 was specifically trained to improve performance on difficult problems, simple errors, and long-horizon tasks, making it more practical for everyday programming use. In our upcoming models, we’ll focus on addressing the issues you mentioned! If possible, could you let me know which model you would use the GRM-2.6-Plus or the GRM-2.6-Plus-0628?
I use for planning GRM-2.6-Plus-0628 (temperature: 0.8) and for execution of the plan(+ steering for some fixes) GRM-2.6-Plus (temperature: 0.6). If I should choose one for both(plan/execute) then GRM-2.6-Plus since its follow instructions better in general for precise changes.
Generated q4 quant: https://huggingface.co/eugene-kamenev/GRM-2.6-Plus-0628-Q4_K_M-GGUF
Here are the args:
["/app/llama-server","--cache-reuse","256","--host","127.0.0.1","--jinja","--min-p","0.00","--no-mmproj-offload","--port","55101","--presence-penalty","0","--repeat-last-n","512","--repeat-penalty","1.05","--temperature","0.4","--top-k","20","--top-p","0.8","--no-warmup","--alias","grm-2.6-plus-0628","--batch-size","4096","--cache-type-k","q4_0","--cache-type-v","q4_0","--flash-attn","on","--fit-ctx","160000","--kv-unified","--model","/models/models/GRM-2.6-Plus-0628/grm-2.6-plus-0628-q4_k_m.gguf","--n-predict","-1","--n-cpu-moe","0","--n-gpu-layers","all","--parallel","1","--reasoning","on","--ubatch-size","4096"]
Best model on my plate now. Thank you for this.