DeepSeek-V4.1-Flash-1017 Coder: a Compressed Coding Model for Consumer Workstations
In a previous article we looked at how to launch DeepSeek-V4.1-Flash on a consumer workstation by combining VRAM, system RAM and NVMe storage. Today we return with a second option aimed squarely at software development: DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF, published by ISTA-DASLab. This release takes the same base model and removes a large part of its sparse experts, shrinking the weight files enough that a complete coding assistant can live entirely in graphics memory and system memory, with no NVMe offloading at all.
What the coder variant changes
The new build is not a simple re-quantization. It is a capability-targeted release: the compression pipeline was designed around coding and agentic workloads, so the parts of the network that matter most for programming are preserved with higher precision while the rest is aggressively reduced. The result is a model that behaves like a specialist rather than a generalist, and that fits into hardware most developers already own.
GSQ and RCO: the two compression pillars
The first pillar is Gumbel-Softmax Quantization (GSQ). Instead of rounding weights to a fixed grid, GSQ learns per-coordinate grid assignments together with per-group scales, which produces far more accurate low-bit scalar quantization. The second pillar is Riemannian Constrained Optimization (RCO), which treats the bit budget of each tensor as a hard constraint during joint optimization. In practice RCO decides where the available bits should go, allocating more precision to sensitive layers and less to layers that tolerate heavy compression. Together the two techniques keep the model coherent at bit rates that would normally destroy quality.
Expert pruning: from 384 to 192 routed experts
The most visible change is structural. Half of the sparse mixture-of-experts routed layers are removed entirely: out of the original 384 routed experts per layer, exactly 192 are kept, selected for their relevance to coding tasks. The surviving weights keep the per-tensor precision of the IQ2_XXS build. Because whole experts disappear instead of being merely quantized, the saving is much larger than a bit-width reduction alone would give.
Memory footprint
The transformer weights now occupy 72.1 GB, compared with 138.2 GB for the unpruned IQ2_XXS build. Measured against the original model's transformer parameters this works out to roughly one effective bit per parameter, although the surviving parameters average 2.06 bits because the dense tensors are kept at higher precision. The model still uses the shared 110.6 GB Engram lookup table from the unpruned release; it reads one row per token and can stay on disk or in system memory, so it does not have to be resident in VRAM.

Hardware configurations
The practical consequence is that a dual-GPU workstation with a generous amount of system memory can host the whole thing without touching the NVMe drive. Two RTX 3090 cards, or two AMD R9700 cards, combined with 192 GB of RAM, are enough to keep both the transformer weights and the Engram table in fast memory. If you prefer to spend less, you can still offload part of the Engram table to an NVMe drive; even then the total footprint stays below 50 GB compared with the original 2.2 bpw quantization of the base model.
| Configuration | GPU memory | System RAM | NVMe usage |
|---|---|---|---|
| Dual RTX 3090 | 48 GB | 192 GB | none |
| Dual AMD R9700 | 64 GB | 192 GB | none |
| Single GPU with offload | 24 GB | 128 GB | Engram rows on disk |
Benchmark results
Aggressive pruning usually costs accuracy, but the published evaluations show that most of the coding ability survives. On LiveCodeBench v6 the coder build scores 88.57 against 89.71 for the base model, retaining 98.7 percent of the original capability. On SWE-bench Verified it reaches 77.00, preserving 94.8 percent of baseline performance. In other words, a multi-billion parameter coding model that once required a server now runs on a workstation while losing only a few points on realistic software engineering tasks.
Getting started
The GGUF files, the quantization configuration and the evaluation plots are all available in the ISTA-DASLab repository on Hugging Face. Download the variant that matches your memory budget, load it with your preferred inference engine, and point it at your codebase. If you are coming from the previous DeepSeek-V4.1-Flash guide, the workflow is identical; only the model files and the memory plan change.
Let's work together
Do you need more info, help with your project, or to develop an idea?
Whether it's an easy question, a quick doubt, or just a 5-minute chat, send me a message—it costs nothing and I'm always ready to help. I love discussing a problem to understand it, getting creative with solutions, and focusing on simple, reliable, and straightforward ideas that we can actuate quickly.
Contact me →