Hardcore Reviews
Hardcore Reviews

Sparse Attention Architecture Compared: QSA Picks Tokens, GDN Compresses History, DSA Reuses Indexes -- Five Flagships and the Layer That Decides Whether 1M Context Is Affordable

Every lab cuts the cost of long-context attention differently, so this review classifies open-weight flagships by architecture route rather than by parameter count or price. The lead subject, Qwen3.8-Flash-Next, takes a hybrid route - GDN compresses history while QSA picks important context at micro-block granularity (125B main model plus 51B of N-gram embeddings, 6B active per token, native 262,144 tokens extensible to 1M with YaRN). It is set against Hy4 preview's Gated DSA with cross-layer IndexCache reuse, GLM-5.3-Flash's sparse-plus-linear hybrid, and DeepSeek's DSA line. The routes collapse into three families: sparse token selection, linear or recurrent compression, and hybrids of both; parameter count, active ratio, context, license and API price snapshots serve as supporting columns. Division of labor with two earlier reviews on this site: those ran the API price math at the 320B tier and the deployment-threshold math at 700B-2.8T, while this one runs only the architecture math. Five scenario verdicts close it out, with one caveat repeated: active ratio saves compute, but the attention mechanism decides whether long context is affordable at all - and for models with unconfirmed licensing, check the model page before commercial use.

Published August 30, 20269 min read
<!-- sparse-attention-architecture-comparison-review | review | Sparse Attention Architecture Compared: QSA Picks Tokens, GDN Compresses History, DSA Reuses Indexes -- Five Flagships and the Layer That Decides Whether 1M Context Is Affordable -->

On August 26, 2026, Qwen shipped two things on the same day: Qwen3.8-Flash-Next, a 125B multimodal MoE model, and FlashQLA, a kernel library written specifically for GDN. Put them side by side and the main battlefield of this year's architecture race comes into view. It is not parameters. It is not leaderboard scores. It is the attention cost of long context.

By August 2026, "supports 1M context" is table stakes for a flagship: Hy4 preview, Kimi K3, GLM-5.3-Flash, the DeepSeek V4 line, and the production Qwen3.8-Flash all claim 1M. But supporting 1M and affording 1M are two different things. Under dense attention, compute grows with sequence length and the KV cache piles up linearly with it -- running 1M used to double inference cost. This year the ways vendors compress that curve have split into three clear routes: picking tokens (sparse), compressing history (linear/recurrent), and reusing indexes (sharing sparse selection across layers). Which route a model takes decides the real VRAM and billing curve of your long-context workload.

Scope first: this is a representative comparison, not hands-on benchmarking. Architecture descriptions, parameters, and cost-reduction figures come from official blogs, READMEs, model cards, and public reporting, with the sourcing basis labeled per item (official release / media citation / our earlier fact-check / converted / engineering estimate), as of 2026-08-30. Derived figures such as activation ratios are converted or engineering estimates and exclude KV cache and runtime overhead. Prices and repo stats are snapshots; defer to official pricing pages and vendor sites. Any field the vendor has not disclosed is marked "Not collected" -- we do not infer or extrapolate. Not investment advice.

Division of labor first, so you do not mix up the ledgers: our lite flagship API price comparison covers the API unit-price and cache math of 320B-class models; our open-source flagship comparison covers the parameter, activation-ratio, and deployment-threshold math of 700B-2.8T models. This article covers the third ledger -- the attention architecture ledger: how each flagship drives down long-context attention cost, where the routes diverge, and what to look at when choosing. Three independent ledgers; read all three.

1. The Route Table: Five Flagships, Three Routes

ModelAttention routeMechanismDisclosed cost reductionBasis
Qwen3.8-Flash-Next (Alibaba)Hybrid: linear compression + sparse selectionGDN compresses history; QSA uses a compressed lightweight indexer to pick important context at micro-block granularity, running full attention only on the selected partTraining cost about 1/9 of Qwen3.7-Plus (officially attributed to hybrid attention, Gated Residual, and Muon combined)Official blog / README
Tencent Hy4 previewSparse selection + cross-layer index reuseGated DSA (64 heads, sparse index top-k 2048) plus IndexCache for cross-layer sparse index reuse; the README states it is inspired by DeepSeek and GLMNot collectedOfficial README / press release
Zhipu GLM-5.3-FlashHybrid: sparse + linear attentionFirst open frontier model to adopt a hybrid sparse and linear attention architecture, paired with mHC manifold-constrained hyper-connectionsAttention compute down 3.01x, KV cache down 4.44x (vs. GLM-5.3)Official release / media citation
Kimi K3 (Moonshot)Hybrid linear attentionKDA hybrid linear attention plus attention residualsNot collectedTencent Cloud Developer Community citation
DeepSeek V4-FlashNot collectedStructure identical to the April Preview (284B / 13B MoE); this round reworked post-training onlyNot collectedOfficial changelog

Three first-glance takeaways:

  1. Pure sparse is down to one vendor; hybrids are the mainstream. Three models run explicit hybrids (Qwen's GDN+QSA, Zhipu's sparse+linear, Kimi's KDA), only Hy4 preview is pure sparse selection, and one has not disclosed. The reason is simple: sparse answers "which positions to compute," linear answers "how to store history," and the two cost curves do not overlap -- stacking them compresses both ends more easily than picking either alone.
  2. Only one vendor puts numbers on the savings. GLM-5.3-Flash reports 3.01x lower attention compute and 4.44x smaller KV cache (against its own GLM-5.3), the only verifiable quantitative basis among the five. Qwen's 1/9 is a training-side figure and the combined result of hybrid attention, Gated Residual, and Muon -- it does not mean "attention cost is 9x lower." Hy4 and Kimi K3 do not disclose attention savings. None of the numbers you can see is directly comparable across vendors.
  3. One architecture is still undisclosed. The DeepSeek V4-Flash changelog says only that structure and size are unchanged and post-training was reworked; the attention mechanism is not disclosed. That is a genuine transparency gap at selection time -- its price is very competitive (section 3), but you do not know which layer buys you that efficiency.

2. Mechanism Breakdown: Picking Tokens, Compressing History, Reusing Indexes

Route one: picking tokens (sparse attention) -- QSA and Gated DSA. Accept that not every historical token deserves the current step's full attention, use a lightweight indexer to pick the important ones, and run full attention only on those. Hy4 preview's Gated DSA uses a sparse index of top-k 2048 -- each step looks carefully at 2048 positions rather than all of 1M -- and IndexCache reuses the selection across layers so the layer below consumes what the layer above already picked. Qwen's QSA follows the same logic but pushes granularity down to micro-blocks and stresses that the indexer itself is compressed -- the picking step cannot be expensive either, or it eats back the attention it saved.

Route two: compressing history (linear/recurrent) -- GDN and KDA. Stop keeping the full KV sequence; compress the context that has already happened into a compact state that later steps read directly. That is what Gated DeltaNet (GDN) does. The cost is information loss, and quality depends heavily on the gating design; the benefit is that cost stops scaling with sequence length -- the only route that fundamentally sidesteps linear KV cache growth. Kimi K3's KDA belongs to the same family and adds attention residuals as compensation.

Route three: hybrids -- GDN+QSA, sparse+linear. Use the linear half to compress most of the history and keep a small slice on sparse or full attention to protect precision. This is the dominant engineering compromise today. Zhipu's numbers show why: compute down 3.01x, KV cache down 4.44x -- the KV cache drops more than compute does, which is the "compress history" half at work.

One sentence for all three routes: the activation ratio saves compute; the attention mechanism decides whether long context is affordable at all. The former sets the FLOPs per token; the latter sets how that cost grows with length.

3. The Supporting Columns: Parameters, Activation Ratio, Context, License, Price

Architecture sets the curve, but these five columns decide whether you can actually use it (snapshot 2026-08-30; activation ratios are converted):

ModelTotal / active per tokenActivation ratio (converted)ContextLicenseAPI price (per 1M tokens, snapshot)
Qwen3.8-Flash-Next125B (plus 51B N-gram embeddings) / 6B~4.8% (6B/125B); ~3.4% if the 51B N-gram table counts toward totalNative 262,144, extended to 1,000,000 via YaRNUnconfirmed: GitHub API license field is None, no LICENSE file in the repo; check HF / ModelScope model pages for the license shipped with the weightsProduction Qwen3.8-Flash: $0.16/$0.47 (official X account); $0.15/$0.47 (third-party aggregator, 08-28); 1 RMB / 3 RMB (RMB basis). Defer to the official QwenCloud site
Tencent Hy4 preview770B / 49B~6.4%1MApache 2.0$0.834/$2.501, cache read $0.042 (third-party OpenRouter snapshot 2026-08-28, not an official rate card)
Zhipu GLM-5.3-Flash320B / 18B~5.6%1M (max output 128K)MIT0.8 RMB / 2.8 RMB, cache hit 0.23 RMB (official pricing, launched 2026-08-26)
Kimi K32.8T / Not collectedNot collected1MCommercial use allowed (terms per official channels, our earlier reporting)Output 100 RMB/M (official pricing via media)
DeepSeek V4-Flash284B / 13B~4.6%1MMIT$0.14/$0.28, cache hit $0.0028 (official pricing page); domestic pricing also uses peak/off-peak tiers

Three things worth calling out:

  1. Every activation ratio lands in the 5%-6.5% band. Converted: Qwen ~4.8% (~3.4% counting the 51B N-gram table), DeepSeek ~4.6%, Zhipu ~5.6%, Hy4 ~6.4%. On this dimension the five cannot differentiate anymore -- it is table stakes, not a differentiator. The real split returns to attention architecture: at the same ~5% activation ratio, dense attention and GDN+QSA produce two entirely different cost curves at 1M.
  2. Qwen's 51B N-gram embeddings are the design outlier. They do not participate in per-token compute (activation is only 6B) and can be offloaded to host memory, scheduled via asynchronous prefetch overlapped with compute -- decoupling capacity from VRAM. The cost is a new PCIe bandwidth dependency; the benefit is that a 125B-class multimodal model becomes feasible on consumer GPUs.
  3. The license is the only red-flagged cell in this article. The Qwen3.8-Flash-Next repo has no LICENSE file, the GitHub API license field returns None, and the community is already asking. We do not assert that it is or is not Apache 2.0 -- we mark it Unconfirmed. For contrast: Hy4 preview is Apache 2.0; GLM-5.3-Flash and DeepSeek V4-Flash are MIT. Open weights do not equal an open-source license -- the former lets you download, run, and modify; the latter decides whether you can ship it in a commercial product. Read the license text on the model page yourself before you ship.

4. Engineering Reality: Working in a Paper Does Not Mean Fast on a GPU

This is the section most often skipped, and the one most likely to bite.

DimensionQwen3.8-Flash-NextHy4 previewGLM-5.3-FlashKimi K3DeepSeek V4-Flash
Kernel / supporting engineeringFlashQLA open-sourced the same day (GDN kernel library, MIT); install requirements SM90 and up, CUDA 12.8+, PyTorch 2.8+Official vLLM / SGLang images, TP8 minimum; native MTP speculative decoding layer, 3 draft tokens by defaultIn-house SGLang inference engine: W8A8 quantization + INT8/FP8/BF16 mixed cache quantization + EPD three-stage disaggregated scheduling; 3x end-to-end (official basis)MXFP4 quantized weights (4 bits/param)Not collected
Kernel speedup2-3x forward, 2x backward (vs. FLA Triton kernel; vendor self-reported, not third-party reproduced)Not collectedNot collectedNot collectedNot collected
Context basisNative 262,144, extrapolated to 1,000,000 via YaRNNative 1MNative 1MNative 1MNative 1M
Weights / quantizationNot collectedBF16 + FP8 dual formatNot collectedMXFP4Not collected

Four ways to read it:

  1. "Native" and "extrapolated" are not the same thing. Only Qwen3.8-Flash-Next reaches 1M by extrapolating from a native 262,144 window with YaRN; the other four list native 1M. A native window is a length seen during training; extrapolation is positional-encoding interpolation and usually behaves differently in the tail of long documents. This does not mean extrapolation is unusable -- it means that before running a full 1M, do your own retrieval test instead of capacity planning on the spec number.
  2. Someone has to write the kernel to pay off the new architecture's speed debt. Qwen shipping the model and the GDN kernel library the same day is itself the answer: an architecture that holds in a paper is a math problem, and making it fast on a GPU is an engineering problem. But hold the 2-3x figure to its reference frame -- the baseline is the FLA Triton kernel, not every backend and not a hand-written CUDA version, and the vendor names pretraining and on-device agentic inference as the scenes where it pays off.
  3. The kernel floor can be higher than the model floor. FlashQLA requires SM90 and up, meaning a Hopper (H100/H800 generation) or Blackwell card; A100 (SM80) and consumer RTX 30/40 series (SM86/SM89) are not on the support list, and CUDA 12.8 plus PyTorch 2.8 are hard requirements. To get GDN running at the vendor's stated speed, the hardware bar is tougher than the bar for simply loading the model.
  4. Hy4 and GLM take the other path: official prebuilt images. Hy4 ships vLLM / SGLang images with deployment docs written at cluster scale; GLM ships its own SGLang engine with quantization and scheduling included. For teams with no kernel engineers, having an official image affects time-to-production more than how advanced the architecture is.

5. Pick by Scenario: Five Verdicts

  1. Want the most controllable long-context cost curve and can live with a pending license: Qwen3.8-Flash-Next. GDN+QSA holds down both ends -- compressing history and picking tokens -- and 6B activation (~4.8%) plus a 51B N-gram table that offloads to host memory make it the most self-hosting-friendly design here; FlashQLA shipping the same day shows the kernel is not an IOU. But the license is unconfirmed -- check the HF / ModelScope model pages before commercial use. For the hands-on path, see our Qwen3.8-Flash-Next local deployment SOP (same batch).
  2. Want a clean license plus quantified savings: Zhipu GLM-5.3-Flash. MIT, 1M context, 3.01x lower compute and 4.44x smaller KV cache (official basis, vs. GLM-5.3), 320B/18B. It is the only one of the five that reports architecture savings as a checkable number, and that saving is what funds the 0.8 RMB / 2.8 RMB pricing -- the low price is architecture, not a burn-rate subsidy.
  3. Want a clean license, complete official images, and the pure-sparse route: Tencent Hy4 preview. Apache 2.0, BF16/FP8 dual weights, Gated DSA + IndexCache, native MTP speculative decoding, vLLM/SGLang images. The representative of the token-picking route, suited to teams that want the sparse dividend without writing kernels; the price is 770B of resident weights, so self-hosting is a cluster-scale threshold.
  4. Want maximum scale, with attention savings as a secondary concern: Kimi K3. 2.8T total, KDA hybrid linear attention, MXFP4 weights to compress VRAM, 1M context. It solves the capability ceiling; output at 100 RMB/M is a premium tier, not a value pick.
  5. For every scenario: when choosing a 2026 flagship, treat attention architecture as a first-class citizen before parameters. Activation ratios have converged into a 5%-6.5% band and no longer differentiate; what actually drives your long-context bill is the attention route, whether the context is native or extrapolated, and whether kernels and official images ship with it. Read all three ledgers -- parameters and deployment thresholds and API price and cache math cover the other two. And the last point: for any model with an unconfirmed license, check the model page before you ship commercially.

One line to close: when activation ratios have converged around 5% and every flagship claims 1M context, flagships are no longer separated by parameter count but by whose attention makes long context genuinely affordable -- and the page that says whether you may ship it commercially deserves to be opened before any 3.01x figure.

FAQ

Q1: What is the actual difference between sparse and linear attention? A1: The difference is which line item each one saves. Sparse attention (QSA, Gated DSA) is "picking tokens": it keeps the full KV sequence but runs full attention only on positions the indexer selects, saving compute. Linear/recurrent attention (GDN, KDA) is "compressing history": it compresses past context into a compact state, saving KV cache and the overhead that grows with length, at the price of compression loss. One optimizes how much you compute; the other optimizes how much you store. Hybrids save both.

Q2: Does "supports 1M context" mean I can run a full 1M? A2: Not necessarily -- first check whether it is native or extrapolated. Hy4 preview, GLM-5.3-Flash, Kimi K3, and DeepSeek V4-Flash all list native 1M. Qwen3.8-Flash-Next has a native window of 262,144 and reaches 1,000,000 through YaRN interpolation (the production Qwen3.8-Flash defaults to 1M; defer to the official site). Extrapolation usually behaves differently from a native window in the tail of long documents, so run your own retrieval test before committing to a full 1M.

Q3: Which affects cost more, the activation ratio or the attention architecture? A3: It depends on how long your context is. The activation ratio sets the compute burned per token, but it has converged -- converted across these five it sits in a roughly 4.6%-6.4% band (~3.4% for Qwen when the 51B N-gram table is counted), too tight to differentiate. The attention architecture sets how that cost grows with sequence length: at short context (tens of thousands of tokens) the activation ratio dominates; at long context (hundreds of thousands to 1M) the attention architecture immediately becomes the primary variable. For agent loops, whole-repo code, and long research documents, ask about the route before asking about parameters.

Q4: Can Qwen3.8-Flash-Next actually be used commercially? A4: This article marks it Unconfirmed and will not decide for you. The checkable facts: the GitHub API license field returns None; the repo root contains only README.md and tech_report.pdf; the README states that the license file ships with the weights on Hugging Face Hub or ModelScope; and there is already a discussion thread on HF asking about it. Open weights (download, run, modify) and an open-source license (commercial use, redistribution) are two different things. Read the license text shipped with the weights on the model page yourself before commercial use.

Q5: For self-hosting, is the architecture enough to go on? A5: No -- you also need to check the kernel, the hardware floor, and official images. Someone has to pay off the new architecture's speed debt: Qwen pays it with FlashQLA, shipped the same day, but that requires SM90 and up, CUDA 12.8+, and PyTorch 2.8+, with A100 and consumer RTX 30/40 series off the support list. Hy4 preview and GLM-5.3-Flash take the other path, shipping vLLM/SGLang images and an in-house inference engine. And remember that total parameters set resident VRAM: 770B and 2.8T are both cluster-scale thresholds.


References

This is a representative comparison compiled from public reporting and official releases, snapshot 2026-08-30, not hands-on benchmarking and not investment advice. Architecture descriptions and cost-reduction figures defer to each vendor's official basis; activation ratios and other derived figures are converted or engineering estimates. The Qwen3.8-Flash-Next license status is unconfirmed -- verify against the license file shipped with the weights on the HF / ModelScope model pages before commercial use. Prices and repo data may change at any time; verify against official pricing pages before purchasing.

This article is AI-assisted and human-edited. Last updated: 2026-08-30

FAQ

What is the actual difference between sparse and linear attention?
The difference is which line item each one saves. Sparse attention (QSA, Gated DSA) is "picking tokens": it keeps the full KV sequence but runs full attention only on positions the indexer selects, saving **compute**. Linear/recurrent attention (GDN, KDA) is "compressing history": it compresses past context into a compact state, saving **KV cache and the overhead that grows with length**, at the price of compression loss. One optimizes how much you compute; the other optimizes how much you store. Hybrids save both.
Does "supports 1M context" mean I can run a full 1M?
Not necessarily -- first check whether it is native or extrapolated. Hy4 preview, GLM-5.3-Flash, Kimi K3, and DeepSeek V4-Flash all list native 1M. Qwen3.8-Flash-Next has a native window of 262,144 and reaches 1,000,000 through YaRN interpolation (the production Qwen3.8-Flash defaults to 1M; defer to the official site). Extrapolation usually behaves differently from a native window in the tail of long documents, so run your own retrieval test before committing to a full 1M.
Which affects cost more, the activation ratio or the attention architecture?
It depends on how long your context is. The activation ratio sets the compute burned per token, but it has converged -- converted across these five it sits in a roughly 4.6%-6.4% band (~3.4% for Qwen when the 51B N-gram table is counted), too tight to differentiate. The attention architecture sets how that cost grows with sequence length: at short context (tens of thousands of tokens) the activation ratio dominates; at long context (hundreds of thousands to 1M) the attention architecture immediately becomes the primary variable. For agent loops, whole-repo code, and long research documents, ask about the route before asking about parameters.
Can Qwen3.8-Flash-Next actually be used commercially?
This article marks it **Unconfirmed** and will not decide for you. The checkable facts: the GitHub API license field returns None; the repo root contains only `README.md` and `tech_report.pdf`; the README states that the license file ships with the weights on Hugging Face Hub or ModelScope; and there is already a discussion thread on HF asking about it. Open weights (download, run, modify) and an open-source license (commercial use, redistribution) are two different things. Read the license text shipped with the weights on the model page yourself before commercial use.
For self-hosting, is the architecture enough to go on?
No -- you also need to check the kernel, the hardware floor, and official images. Someone has to pay off the new architecture's speed debt: Qwen pays it with FlashQLA, shipped the same day, but that requires SM90 and up, CUDA 12.8+, and PyTorch 2.8+, with A100 and consumer RTX 30/40 series off the support list. Hy4 preview and GLM-5.3-Flash take the other path, shipping vLLM/SGLang images and an in-house inference engine. And remember that total parameters set resident VRAM: 770B and 2.8T are both cluster-scale thresholds.

Related

Hardcore Reviews

Closed API vs Open Weights: What Does One Image Really Cost

With ChatGPT Images 2.5 and Ant's open-source LLaDA-Image landing in the same week, text-to-image has split into closed APIs versus self-hosted open weights. This review ignores image quality and runs the cost-and-control numbers instead: five routes - closed APIs, self-hosted open weights, per-second third-party inference platforms, local consumer hardware, and domestic cloud APIs - with per-image cost projected at two volumes (100 and 10,000 images per day), plus a comparison table and scenario-based selection (hobby use, e-commerce batch, data-sensitive industries, brand-style fine-tuning, maximum quality). It flags four traps: undeclared licenses, cold starts on per-second billing, Chinese text rendering, and cross-border data transfer. Explicitly scoped apart from our 8-26 capability review of reasoning image models. Representative comparison, not hands-on benchmarking; pricing per official sites.

Sep 9, 20269 min read
Hardcore Reviews

5 Model Hosting Platforms Compared After Nvidia's HF Deal

After NVIDIA's Hugging Face acquisition, "where do open models live and run" became a must-answer question. This review compares five model hosting and distribution platforms: Hugging Face (Hub+Spaces+Inference Providers), ModelScope (domestic compliance and download advantage in China), Replicate (per-second billed, one-click API), fal.ai (strong at generative inference), and OpenRouter (multi-model aggregate routing). Includes official 2026-09 snapshot pricing (HF PRO \$9/mo, Replicate T4 \$0.000225/s, fal Serverless H100 from \$1.89/h and more), a full comparison table and scenario-based selection; also clarifies the division of labor with our earlier API-gateway review. Representative comparison, not hands-on benchmarking.

Sep 8, 20269 min read
Hardcore Reviews

CodeArena: Fable 5.1 Leads, Qwen Near at 1/8 Price

CodeArena, run by LMArena, is a frontend-coding leaderboard (end-to-end web-app generation, human-preference Elo). As of 2026-09-03: Claude Fable 5.1 leads at 1765 ($40/M), Qwen3.8-Max-0902 hit 1691 on day one and now ~1688 ($5/M, reaching the front rank at one-eighth the price), Gemini 3.8 Flash sits at 1567 (cheap variant, #18), Kimi K3 ~1674; GPT-6 Astra just launched 9/3 and its coding score is pending. Takeaway: Elo measures preference not accuracy — weigh price-performance and your own needs.

Sep 5, 20269 min read