peculiar-ragdoll

/

Tiel-Coder-35B-A3B-GGUF-MTP

like 62

Thanks a lot for the support, glad to hear from others it's working! :)

yeah this model is just crazy good

i tested this on my own tasks (reverse engineering, modding games (coding, scripting), agentic stuff in general, finance stuff etc) and yeah the q4 kl model holds up pretty well

it beats the base model qwen 3.6 a3b 35b q5 quant (i ran the apex quality one which ig is comparable to q5), ornith 1.5 a3b 35b, even the qwen 3.8 27b 2bit qat version which surprised me a bit. the q4 of qwen 3.8 27b is not runnable on my pc (eh i consider 5 tok/s to be unsuable category) so i didnt compare that

the accuracy of this model in those tasks as well as efficiency is very good compared to its peers.
i have tested so many fine tunes and this one is actually worth using for my usecase.
huge work man, i love this model
keep working on stuff

i dont really do formal benchmarks, so heres a summary of the efficiency if anyone wanna dig into it

The 13 tasks

#	Task	What it tested
1	re_1_asm	Reading raw x86 disassembly and explaining what it computes
2	re_2_patch	Calculating exact binary patch bytes to alter game behavior (had a planted error in the prompt — miscounted instruction length by 1 byte)
3	mod_1_scm	Writing a C++ handler for a custom game-script opcode VM
4	mod_2_rtxremix	Domain knowledge on ray-tracing mod tooling (texture/mesh hash stability, dynamic buffers, draw-call batching) for an old fixed-function-pipeline game
5	code_1_algo	O(n log n) longest-increasing-subsequence implementation
6	code_2_bugfix	Spotting a ZeroDivisionError root cause
7	code_3_refactor	Cleaning up a nested-ternary/if-else JS function
8	tool_1_call	Single structured tool call
9	tool_2_multistep	Planning a multi-step tool-call sequence
10	agent_1_plan	Agentic diagnose-and-fix plan for a broken build
11	agent_2_selfcorrect	Recovering from a self-introduced test failure
12	fin_1_options	Options payoff/P&L math
13	fin_2_portfolio	Portfolio expected return + variance math

Where each model excelled or died

Tiel-Coder — clean sweep. All 13 correct. On task #2 (the planted error), it's the only model that explicitly caught and told me the prompt's byte count was wrong before answering — the most trustworthy handling of that task. Also the fastest and most token-efficient of everything tested.

Ornith (Tiel-Coder's base model) — also a clean sweep, all 13 correct. On task #2 it didn't call out my error directly, but its patch was self-consistent and it explicitly reasoned through why the leftover byte doesn't matter — genuinely careful, just slower and more roundabout getting there than Tiel-Coder.

Qwen3.6 (older base) — functionally correct on all 13, but weaker on task #2: it noticed the same discrepancy internally, explicitly decided in its own reasoning to ignore it, and gave an answer without flagging anything to me. Not wrong on its own terms, but the least trustworthy of the three "healthy" models — and by far the slowest/most verbose (~2x the tokens of Tiel-Coder for the same battery).

The 2-bit quant (27B dense) — this is where a model actually died. Task #2 came back factually broken: it mislabeled a conditional-jump opcode as an unconditional jump (would produce a non-functional patch if actually used) plus a separate arithmetic error in the offset. Task #4 came back empty — cut off mid-thought with zero final answer despite no token cap. It was also dramatically the slowest (roughly 4x Tiel-Coder's wall-clock time for the same battery) with two individual tasks each taking 17–28 minutes alone.

Net picture: 3 of 4 models are solid daily-use quality; the 2-bit quant is not — it's both the slowest and the only one that produced a genuinely wrong, unusable answer.

in my quantum physics tests,this model though about 12k~ tokens and scored 2.1 out of 5,ornith 1.5 35b a3b scored 4 out of 5 but though about 60k~ tokens,in my setup that model is more usable than others.

Thanks for sharing, AsThirtyThree! That results totally checks out. As stated in the model card, this is a model specifically engineered to maximize coding ability, while sacrificing a bit of graduate level knowledge and reasoning in other fields to achieve it :) I do not necessarily recommend TielCoder as a general purpose science or research model: it's a dedicated software engineering worker, implementer and debugger. It's not bad at non-software stuff by any means, it's just not the best around. My favourite 35b-a3b model for more general knowledge and reasoning stuff is Nail-35b-a3b. On my tests it beats Ornith1.5 cleanly on MMLU-Pro, on both speed and correct answer.