DeepSeek V4.1 Flash Coding

We introduce DeepSeek V4.1 Flash Coding from Nace.AI: the 763B DeepSeek V4.1 Flash 1017 in a 71.8 GB GGUF.

Overview

DeepSeek V4.1 Flash Coding is a Nace Specialized Quant for coding and agentic tasks. It is built and tuned on real code and agent work, and it runs on a single 48 GB Mac with llama.cpp.

  • 71.8 GB for a 763B-parameter model
  • 37 of 38 LiveCodeBench v6 problems solved

Results

Held-out code KL

LiveCodeBench problems solved

  • 71.8 GB for a 763B-parameter model.
  • 0.140 KL to a near-lossless Q8 reference on real code, against 0.161 for EXL3 at 124 GB.
  • 88.4% of next tokens match the Q8 reference.
  • 37 of 38 LiveCodeBench v6 problems solved, level with ISTA Coder on the same two seeds.

How it is built

Coding-first calibration

Built from real coding work: agents fixing real repositories, contest solutions and code commits. Every choice in the file follows how DeepSeek routes code.

Code calibration

Code-aware expert pruning

DeepSeek V4.1 Flash has 384 experts in each of its 40 layers. We keep the 200 per layer that code uses most. They carry 90% of the model's code routing.

Kept experts per layer

Mixed precision per layer

Every expert is scored by how often it fires, how much the router relies on it and how much rounding costs it. The 16 layers where code needs it most get larger, higher-precision experts.

Bits per layer

Nace Weight Space Quantization

An agent that edits the weights themselves.

Nace Weight Space Quantization puts an AI agent (Claude Fable 5.2) to work on a finished quantized file. With reading, navigation, quantization and checking tools, it reshapes how weights are stored, rebalances precision toward the layers coding leans on, and prunes and swaps experts toward what coding work uses.

Nace Weight Space Quantization

Every edit is measured against a near-lossless Q8 reference, and only edits that score better stay. The file in this repo went through all four steps: the agent's edits cut held-out code KL from 0.142 to 0.140 at the same 71.8 GB.

Run it

The download is the 71.8 GB main weights plus the 110.6 GB Engram table in engram/.

hf download nace-ai/DeepSeek-V4.1-Flash-Coding-GGUF --local-dir coding
llama-server -m coding/DeepSeek-V4.1-Flash-1017-Coding-IQ1_S-200E.gguf \
  --ple coding/engram/DeepSeek-V4.1-Flash-1017-engram-IQ4_NL.gguf_file -c 32768

Use temperature 1.0 and top_p 1.0.

License

MIT, the same license as DeepSeek V4.1 Flash 1017. The full license text is in LICENSE.

Downloads last month
4,321
GGUF
Model size
291B params
Architecture
deepseek41
Hardware compatibility
Log In to add your hardware

1-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 1 Ask for provider support

Model tree for nace-ai/DeepSeek-V4.1-Flash-Coding-GGUF

Quantized
(230)
this model