Tiel-Coder-35B-A3B-MTP-UD-Q4_K_S.gguf
Something about quality of Tiel-Coder-35B-A3B-MTP-UD-Q4_K_S.gguf
Requested html snake game, result is broken unplayable game and this:
The basic Qwen 3.6 35B A3B never did such things in the same environment and with the same prompt, none of the other models I've tested (excluding fine-tunes) did this either.
Thank you for the bug report! :) Can you tell me what backend you use to serve, at what parameters, and what coding harness you use? These factor in greatly. I have only tested with the optimal parameters, on llama.cpp and oMLX, with Pi.dev coding agent. So I'd love to learn what breaks this model and how!
llama.cpp + opencode.
-reasoning-budget 4096
-cache-type-k q4_0 --cache-type-v q4_0
-spec-type draft-mtp,ngram-mod --spec-draft-n-max 4 --spec-draft-p-min 0.8
Nothing special - using same settigns with another models and no problems.
Prompt is "Create a snake clone in a single html page, use whatever libraries you want, research for graphics as you need them, you can download from any source you want."
Used 26k tokens.
Thanks! I recommend using reasoning budget 8192 or 16384, so you let it cook when it needs to :) This model will not overthink when it's not needed, in my experience, which makes it safe and valuable to increase the budget. Also I recommend using q8_0 for kv, not q4, because this model uses extremely small amounts of RAM for KV cache either way compared to the 27b dense models, and q8 gives better speed and quality. At q8 kv a 262k context takes 2.7GB RAM. Also make sure you set the context limit high enough to fit all your work, whether that's 64k or 128k or 262k tokens. If I was you I would rather use a smaller quant for the model weights to give the model a large enough context window at q8 kv, but that's just me :)
These parameters are optimized for agentic coding, and what I used for the benchmarks:
temperature 0.6, top_p 0.95, top_k 20
If you use wrong params on any of these it will cause the model to degrade too.
Hope these suggestions help!