TielCoder 35B-A3B: open coding model reportedly matches far larger rivals in 22GB

What the benchmark round-up claims

A post on dev.to published on August 25 walks through a community benchmark of 35B-A3B models: sparse mixture-of-experts models with roughly 35 billion total parameters but only around 3 billion active per token. The author reports testing several of them on genuine problems from real codebases and ranks TielCoder 35B-A3B as the strongest and most consistent of the group, on both correctness and speed.

The headline claim is that TielCoder's 22GB 4-bit quantization performs at the level of Opus 4.6 medium on these tasks and finishes ahead of other 35B-A3B coding models named KAT-Coder and Nail. The stated motivation for the project is that Qwen 3.8 27B, while excellent, is slow — so a lean MoE coder that handles real problems quickly fills a genuine gap.

Three changes under the hood

According to the post, TielCoder differs from a stock Qwen 35B model in three ways:

• It is a finetune built on Ornith, a base from the Qwen 35B family.
• It applies a code-focused dynamic imatrix quantization, which reportedly matters most at the Q4 level.
• It ships an optimized chat template called Qwen-Sharp, which the author says lets the model reach its conclusions in fewer tokens — a meaningful advantage in agentic coding workflows.

The combined effect, the author claims, is a model that is clearly better than stock Qwen 35B at coding while also running faster.

Speed versus raw accuracy

Asked in the comments why Qwen 3.8 27B was missing from the comparison, the user quoted as peculiar-ragdoll framed it as a trade-off: roughly a sixfold speedup over Qwen 3.8 27B at medium effort, in exchange for solving fewer tasks overall. Another commenter, Anaeijon, put the result in context: Qwen 3.8 beats all of the A3B models, but that is expected, since the A3B models are built to run far faster on much less hardware.

A benchmark designed to resist training

The results rest on SWE Live, described in the post as an agentic software-engineering benchmark that regularly adds new tasks, which makes it impossible to train on the test set. The author calls it the only coding benchmark they consider trustworthy.

Where to get it

The weights are published on Hugging Face under the peculiar-ragdoll account in three variants: standard GGUF files, GGUF builds with MTP, and an MLX build in the oQ4e format aimed at Apple Silicon.

Why it matters

If the claims hold up, a 22GB footprint puts a frontier-adjacent coding assistant on ordinary consumer GPUs with 24GB of VRAM, with no cloud subscription required. The post also makes a subtler point worth noting: two of the three big improvements came not from the finetune itself but from quantization choices and chat-template tuning. For anyone running models locally, that is a reminder that inference-side engineering can matter as much as training.

Two caveats deserve stating plainly. Everything here comes from a single community-authored post with self-reported results, and matching Opus 4.6 medium is the author's assessment on one benchmark rather than an independent evaluation. Meanwhile, the post notes, the community is still waiting on Qwen3-Coder-Next, a native MoE coder from Qwen that could reset this entire category.