ARustyCoder93
A 93-of-256 routed-expert cut of DeepSeek-V4-Flash in GGUF, made for local
agentic coding on one 96 GB GPU. This is not a DeepSeek release. It is a
derived model produced by expert pruning; DeepSeek did not make or endorse
it. It inherits DeepSeek-V4-Flash's MIT license.
Full recipe, scripts, llama.cpp patch and every checksum:
https://github.com/apchat-agent/arustycoder | |
|---|
| file | arustycoder93.gguf — 62,213,283,424 bytes |
| sha256 | 90a4c8a8a32fc664d822f1df093d5e228111c94eebf9398ae5c5737391ecb058 |
| source | antirez/deepseek-v4-gguf → DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf (MXFP4 experts, Q8 attention/shared/output, F16 compressor/indexer); kept tensors are byte-identical to it |
| experts | 93 per layer kept out of 256 (top-6 routing unchanged); 43 MoE layers |
| runs on | stock llama.cpp at or after commit 22b8e31 (2026-08-15). Our serving build for the 2026-08-16 smoke tests was that commit plus the census patch from the recipe repo, which is inert for this file (no full-router tensors). llama-server -m arustycoder93.gguf -ngl 99 -c 262144 -fa on |
| footprint | ~67 GB resident at 256K context on an RTX PRO 6000; ~46 tok/s decode, ~470 tok/s prompt |
How the 93 experts were chosen
Two "grow-from-nothing" censuses (start from a random 8 experts per layer
with the full 256-wide router kept as side tensors, record which experts the
router wants, rebuild at 16, 32, 64) — one on a code/systems corpus, one on
a general corpus. ARustyCoder93 is the union of the two keep-64 sets:
78..93 experts per layer (mean 87.3), padded to 93 in the smaller layers by
merged rank because a GGUF needs a single expert count. Each keep-64 alone
predicts held-out text well but cannot finish generating; the union does.
Method, numbers and the merge rule are in the repository README.
What it does and does not do
- 6/6 coding asks (write Rust, fix Python, explain C, refactor JS, a tool
call, a multi-file plan) finished cleanly, 86-1993 completion tokens, with
the server settings above. Prompts and raw responses are in the recipe
repo under
results/.
- Known limit: a harder single-shot prompt (a generic LRU cache) fell into a
verbatim repetition loop in its reasoning to a 12k-token cap. Expect this
on some hard reasoning; a DRY sampler stops the loop but corrupts
identifiers in code.
- No standard benchmark has been run on this cut. It is an experiment that
turned out useful for interactive coding. Use the unpruned model when
correctness matters more than footprint.
Files
arustycoder93.gguf — the modelconfig.txt, prune.log — build record of this exact filemerged-wants.tsv, coder-census-64.wants.tsv, general-census-64.wants.tsv — the expert rankings it was pruned fromPROVENANCE.md — checksums, versions, commitsLICENSE-DeepSeek-V4-Flash — the upstream MIT notice (DeepSeek); LICENSE — MIT for the recipe
HF metadata note: base_model points at DeepSeek-V4-Flash; there is no
"pruned" relation tag on the Hub, so none is set.
Provenance / credit
Base model: DeepSeek-V4-Flash by DeepSeek (MIT). GGUF source and its
quantization: antirez/deepseek-v4-gguf (tagged MIT on the Hub). Pruning
method, patch and scripts: this project (MIT), 2026-08-15/17.
