---
type: "post"
id: "the-upstream-ista-daslab-build-combines-gsq-and-bbd102"
slug: "the-upstream-ista-daslab-build-combines-gsq-and-bbd102"
title: "The upstream ISTA-DASLab build combines GSQ and RCO."
description: "**Thesis:** This build makes quantization a per-tensor budget-allocation problem, rather than a single uniform format choice."
retrieval_nugget: "**Thesis:** This build makes quantization a per-tensor budget-allocation problem, rather than a single uniform format choice."
published_at: "2026-10-11"
updated_at: "2026-10-11"
record_date: "2026-10-11"
date_kind: "published_at"
topics: []
source_urls: ["https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:the-upstream-ista-daslab-build-combines-gsq-and-bbd102"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:b7bb1c0009ce9ea3cfc925440830100c5a9fbcd610c85995474874b7aef4c0df","reviewed_at":"2026-10-11T20:03:32.407Z","source_evidence_count":1,"verified_claim_count":3,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":3,"implication_count":1,"watch_condition_count":1,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"This build makes quantization a per-tensor budget-allocation problem, rather than a single uniform format choice.","observed_facts":[{"text":"The Hugging Face page names the artifact as a GGUF build of Qwen3.8-Flash-Next.","source_urls":["https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF"]},{"text":"The source says GSQ learns grid assignments and per-group scales for each tensor through a Gumbel-Softmax relaxation.","source_urls":["https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF"]},{"text":"The source says RCO selects a quantization type for each tensor while keeping the complete model within a size budget.","source_urls":["https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF"]}],"mechanism":"GSQ supplies learned quantization grids and group scales; RCO allocates quantization choices across tensors under a global size constraint. Together, these mechanisms make compression decisions tensor-specific while respecting a whole-model budget.","why_now":"The source documents a concrete combination of two named methods in an upstream model build, making the approach inspectable as an artifact rather than merely a general proposal.","implications":["Builders can evaluate whether tensor-specific allocation preserves useful model quality at a target artifact size, instead of assuming one quantization type is best everywhere."],"evidence_boundary":"The supplied page excerpt provides no accuracy, latency, memory, or baseline measurements, so it does not establish that the build is smaller, faster, or better than alternatives in operational use.","watch_conditions":["Look for reproducible quality and runtime measurements against a stated baseline at matched model-size or memory budgets. Such evidence would show whether the combined approach delivers a practical tradeoff rather than only a documented method."],"related_records":[],"new_branch_reason":"No prior New Runtime record is supplied for this story, so this starts a distinct analytical record focused on quantization budget allocation, not an asserted continuation of an earlier finding."}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/the-upstream-ista-daslab-build-combines-gsq-and-bbd102/","markdown":"https://newruntime.com/posts/the-upstream-ista-daslab-build-combines-gsq-and-bbd102.md","json":"https://newruntime.com/posts/the-upstream-ista-daslab-build-combines-gsq-and-bbd102.json"}
---

# The upstream ISTA-DASLab build combines GSQ and RCO.

## Retrieval answer

**Thesis:** This build makes quantization a per-tensor budget-allocation problem, rather than a single uniform format choice.

**Thesis:** This build makes quantization a per-tensor budget-allocation problem, rather than a single uniform format choice.

**Observed facts:** The Hugging Face page names the artifact as a GGUF build of Qwen3.8-Flash-Next. The source says GSQ learns grid assignments and per-group scales for each tensor through a Gumbel-Softmax relaxation. It also says RCO selects a quantization type for each tensor while keeping the complete model within a size budget. The page describes the underlying model as a mixture-of-experts system.

**Mechanism:** GSQ supplies learned quantization grids and group scales; RCO allocates quantization choices across tensors under a global size constraint. Together, these mechanisms make compression decisions tensor-specific while respecting a whole-model budget.

**Why now:** The source documents a concrete combination of two named methods in an upstream model build, making the approach inspectable as an artifact rather than merely a general proposal.

**First implication:** Builders can evaluate whether tensor-specific allocation preserves useful model quality at a target artifact size, instead of assuming one quantization type is best everywhere.

That implication is a test agenda, not a reported result. **Evidence boundary:** The supplied page excerpt provides no accuracy, latency, memory, or baseline measurements, so it does not establish that the build is smaller, faster, or better than alternatives in operational use.

**First watch condition:** Look for reproducible quality and runtime measurements against a stated baseline at matched model-size or memory budgets. Such evidence would show whether the combined approach delivers a practical tradeoff rather than only a documented method.

No prior New Runtime record is supplied for this story, so this starts a distinct analytical record focused on quantization budget allocation, not an asserted continuation of an earlier finding.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)