GSQ-RCO

GSQ-RCO is a model-compression workflow that combines Gumbel-Softmax Quantization (GSQ) with Riemannian Constrained Optimization (RCO). GSQ improves the low-precision representation of individual weight tensors; RCO chooses their precision assignments under a model-size budget. The DASLab releases package the selected tensors in standard GGUF files rather than introducing a new inference format.^[1]

The methods address different parts of quantization. A better low-bit tensor does not determine how much precision each part of a model should receive, while a good allocation cannot recover information absent from its candidate tensors. GSQ-RCO combines those two stages; the name also appears on released quantizations of Qwen3.8-27B and Qwen3.8-Flash-Next.^[1][4][5]

Research origins

GSQ was introduced by Alireza Dadgarnia and colleagues. Michael Helcig and Dan Alistarh introduced RCO. The two reference implementations are separate Apache-2.0 repositories; a compressed checkpoint retains the licensing conditions of its underlying model.^{[2][3][4][5][6]}

How the methods work together

GSQ: improving candidate tensors

GSQ is a post-training method: it starts with pretrained weights rather than training a low-bit language model from scratch. Its reference workflow initializes from GPTQ, then optimizes trainable logits over discrete weight choices alongside quantization scales. Gumbel-Softmax sampling provides differentiable choices during optimization; the final checkpoint contains hard assignments and scales, not those auxiliary training logits.^[4]

The method uses small scalar grids. At higher bit widths, a local-shift formulation considers nearby grid positions instead of learning a distribution over every possible value. The implementation uses the Lion optimizer and can offload completed layers before model reassembly. Refining an existing GGUF checkpoint preserves its deployment representation.^[4]

RCO: allocating the size budget

RCO assigns one option to each group, such as a bit width for a tensor or a keep/prune decision for an expert. It optimizes a loss that depends on the compressed model jointly, rather than assuming independent tensor sensitivities add up to the complete model's error.^[5]

Softmax probabilities describe the candidate assignments. Their expected cost defines a constraint surface. RCO projects gradients onto its tangent space, takes an Adam step, retracts the logits back to the budget, and transports optimizer momentum to the new tangent space. Gumbel straight-through estimation supplies gradients through discrete choices, while a budget-constrained solver selects feasible assignments.^[5]

"Exact budget" refers to the continuous expected-cost constraint, within numerical precision. Decoding onto the solver's discrete cost grid can introduce small deviations: the paper gives 2.251 bits for a 2.25-bit target. Budget feasibility does not prove that the selected compression is globally optimal. RCO also retains optimization hyperparameters, including learning rate, sampling count, and temperature; it removes constraint-penalty coefficients, not all tuning.^[3]

Building a GGUF release

The documented release process first builds a database of GSQ-quantized tensor candidates. RCO searches for an allocation at the chosen budget, and assembly combines those candidates into a GGUF checkpoint. Releases include tensor-allocation text files so the selected quantization type for each tensor can be inspected.^[6]

For the Flash-Next mixture-of-experts releases, the format constrains allocation granularity: the card describes layer-level choices for fused routed-expert tensors, rather than an independently selectable GGUF type for every expert.^[7]

Released checkpoints

The official collection includes the following builds. Sizes below are the authors' published download sizes, not total runtime memory requirements.^[1][6][7][8]

Flash-Next's reported bit widths exclude its fixed-precision n-gram table. That table occupies a separate 28.8 GB shard at 4.5 bits per weight. Both shards must be downloaded. The documented memory-mapped loading path can keep the table on disk; KV-cache and optional vision-projector memory must still be allowed for.^[7]

Coder combines quantization with mixture-of-experts expert pruning. Its effective 1.89-bit figure is measured against the original transformer parameters and excludes the fixed n-gram shard. It does not mean that each surviving weight has a 1.89-bit representation. RCO enforces equal retained expert counts per layer for GGUF export. The card warns that the code-targeted compression sacrifices capabilities outside its calibration domains.^[8]

Evaluation and interpretation

The GSQ paper reports that refining Qwen3-8B's Q2_K checkpoint increased the mean of AIME25, GPQA Diamond, and MMLU-Pro from 50.03 to 56.28 under its evaluation protocol. This is a result for that checkpoint and test suite, not a prediction for every low-bit model.^[2]

For Qwen3.8-27B, the release card reports that the 11.8 GB IQ3_S build matched the BF16 base on AIME25 (100.00) and LiveCodeBench v6 (85.71), while scoring 89.39 against 89.90 on GPQA-Diamond. Its "task-lossless" description concerns those measured tasks; it does not establish identical outputs or preservation of every capability.^[6]

For Flash-Next IQ3_S, the card reports a three-task mean of 93.26 against the base's 93.12. It explicitly attributes small above-baseline results to benchmark variance rather than evidence that quantization made a better model.^[7]

After its October 17 update, the Coder card reports 79.20 on SWE-bench Verified versus 82.80 for the BF16 base, and 87.43 on LiveCodeBench v6 for both, at xhigh reasoning effort. These are author-reported measurements of the released compressed model.^[8]

GSQ's paper also reports 374.0 output tokens per second per GPU for its 2.13-bit Llama-3.1-70B configuration versus 60.3 for BF16, using vLLM, Humming kernels, NVIDIA L40s GPUs, and a ShareGPT workload. This setup is separate from the later GGUF releases and does not establish their speed on consumer hardware.^[2]

Deployment and limitations

The released GGUFs are documented for llama.cpp, Ollama, and LM Studio. Multimodal use requires the corresponding vision projector. A filename suffix identifies a quantization class, not a promise that every tensor uses an identical representation.^[6][7]

Strata is a separate inference project that supports these compressed Flash-Next builds. Its documented execution options distribute storage and computation across available hardware; that deployment layer is distinct from the GSQ-RCO compression search. Small GPU memory alone does not describe the RAM, disk, and context requirements of the complete system.^[9]

GSQ requires temporary optimization memory for auxiliary logits and depends on its initialization. The paper does not claim that scalar quantization universally matches more expressive vector or trellis methods.^[2] RCO's result quality depends on calibration data, and its discrete search is heuristic rather than an optimality guarantee. Its inexpensive budget handling assumes additive costs; non-additive costs, such as fused-operator latency, need different constraint handling.^[3]

References

1. ^1 ^2 ^3ISTA-DASLab. GSQ-RCO-GGUF collection. Hugging Face. Accessed October 21, 2026.
2. ^1 ^2 ^3 ^4 ^5Alireza Dadgarnia et al. GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling. arXiv:2604.18556v2, May 15, 2026; first submitted April 20, 2026.
3. ^1 ^2 ^3 ^4Michael Helcig and Dan Alistarh. Model Compression with Exact Budget Constraints via Riemannian Manifolds. arXiv:2605.00649v3, September 27, 2026; first submitted May 1, 2026.
4. ^1 ^2 ^3 ^4IST-DASLab. GSQ reference implementation and documentation. GitHub. Accessed October 21, 2026.
5. ^1 ^2 ^3 ^4IST-DASLab. RCO reference implementation and documentation. GitHub. Accessed October 21, 2026.
6. ^1 ^2 ^3 ^4 ^5 ^6ISTA-DASLab. Qwen3.8-27B-GSQ-RCO-GGUF model card. Hugging Face. Accessed October 21, 2026.
7. ^1 ^2 ^3 ^4 ^5 ^6ISTA-DASLab. Qwen3.8-Flash-Next-GSQ-RCO-GGUF model card. Hugging Face. Accessed October 21, 2026.
8. ^1 ^2 ^3 ^4ISTA-DASLab. Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF model card. Hugging Face. Accessed October 21, 2026.
9. ^Niko1221. Strata documentation. GitHub. Accessed October 21, 2026.