Strategy

Qwen3.8 Open Weights Arrived Stripped: The Open-Closed Boundary Moved

By IdeaBosqueLast updated: August 12, 2026

The 10-day wait for Qwen3.8-2.4T-A95B open weights ended on August 13, 2026. The weights are live on Hugging Face — 2.4T total parameters, 95B active, sparse MoE with 512 experts, Gated DeltaNet plus Gated Attention hybrid architecture. The first Max-scale open-weight release from any major lab. But the release is stripped: text-only (no vision input), 262,144 context natively (not the 1M the managed API provides), and thinking mode is always on — it cannot be disabled. The full Qwen3.8-Max feature set — vision, non-thinking mode, 1M context, built-in tools — is exclusive to the Qwen Cloud managed API. The LICENSE file is titled "Qwen3.8-Max License" despite the model card calling the Max features "a separate, cloud-only luxury." This article builds on Open-Weight Models Crossed the Agentic Frontier, which mapped the capability gap through July 2026; here we cover the structural shift: the open-closed boundary is moving, and Qwen is paywalling capabilities that were previously open.

Key takeaways

Qwen3.8-2.4T-A95B open weights are text-only, stripped of vision, capped at 262K context, with thinking mode always on — the first Max-scale (2.4T) open-weight release, but Qwen is paywalling the full feature set (vision, 1M context, non-thinking mode, built-in tools) on Qwen Cloud. In Qwen3.5-397B-A17B, vision support was released freely.
Community backlash was immediate: Hugging Face discussion #13 reached 19 upvotes within hours — "You stripped away 3/4 of the model's context memory and blinded it completely" (user NodeLinker). User Shad3: "removing vision is clearly a push for people to use their cloud endpoint so they can make money."
Kimi K3 released full weights including vision on July 27 — 594GB MXFP4, Modified MIT, 8x H100 minimum. The counter-example: Moonshot open-weighted vision while Qwen paywalled it.
Three major models landed the same day: DeepSeek V4 Pro 0813 (BenchLM 60.95, Terminal-Bench 87.9%), Gemini 3.7 Flash ($0.75/$3.75, AA Intelligence Index 56), and Qwen3.8 stripped weights — the densest single-day model release window since the site launched.
NVIDIA Nemotron 3.5 Lightning opened a new open-weight category: purpose-built execution-layer models for long-running agents — 30B MoE, 3B active, OpenMDW-1.1 (weights, data, recipes), harness-optimized for OpenClaw and Hermes Agent, NeMo Switchyard routing ("plans route up to the frontier, execution routes down to Lightning").

The stripped release

The Hugging Face model card confirms the architecture: 2.4T total parameters, 95B activated per token, 512 experts (10 routed + 1 shared), 92 layers. The hidden layout is a repeating block of 3 Gated DeltaNet MoE layers followed by 1 Gated Attention MoE layer — a hybrid linear-attention-plus-soft-attention design aimed at long-context efficiency. Context length is 262,144 natively, extensible to 1,010,000. BF16. Compatible with vLLM, SGLang, and TokenSpeed.

The vendor-reported benchmark table is strong: Terminal-Bench 2.1 at 86.6, SWE-bench Pro at 67.7, DeepSWE 1.1 at 56.6, PaperBench at 93.0, FrontierSWE at 73.5, GPQA Diamond at 92.6, IFBench at 82.8. Competitive with Claude Opus 4.8 and GPT-5.6 Sol on several agentic benchmarks, trailing on others (DeepSWE 1.1: 56.6 vs GPT-5.6 Sol 73.0, Fable 5 70.0). The Artificial Analysis Intelligence Index has not yet scored the open-weight variant — the published scores are for the managed Qwen3.8-Max API, which includes the stripped features.

But the release is not the full model. The model card itself states: "Qwen3.8-Max is the official version based on Qwen3.8-2.4T-A95B with more features, such as vision input & non-thinking support, 1M context length by default, official built-in tools." The open weights get the base architecture and the coding-agent benchmarks. The managed API gets vision, long context, thinking-mode control, and built-in tool integration. The LICENSE file is titled "Qwen3.8-Max License" — the same license covers both the stripped open weights and the full managed service.

Community backlash

The Hugging Face discussion thread #13, titled "Huge disappointment: Qwen 3.8 open weights are text-only and stripped of Qwen 3.8 Max features," accumulated 19 upvotes within hours of the release.

User NodeLinker: "You stripped away 3/4 of the model's context memory and blinded it completely. In previous releases like Qwen3.5-397B-A17B, you didn't touch vision support and released it freely."

User Shad3: "removing vision is clearly a push for people to use their cloud endpoint so they can make money. they are basically doing the DLC thing."

The community reaction identifies the structural shift precisely. In Qwen3.5-397B-A17B (the prior generation), vision support was included in the open weights. In Qwen3.8-2.4T-A95B, vision is exclusive to the managed API. The open-closed boundary moved within a single model generation — not because the capability became harder to open, but because the commercial incentive shifted.

The structural shift: paywalling what was open

This is not a story about one model release. It is about the open-closed boundary being negotiated in real time.

The Brookings Institution published a policy paper on August 10, 2026, arguing "both open and closed AI models are needed" for American AI leadership. Qwen's stripped release is a concrete example of the boundary being drawn not by policy but by commercial strategy — the most capable features stay behind the API paywall, and the open weights serve as a capability demonstration and a funnel to the managed service.

Hugging Face CEO Clément Delangue told CNBC on August 3 that China is "clearly dominating on open models right now." The Qwen3.8 stripped release complicates that narrative. The open-weight leader is now paywalling key capabilities. The dominance is real on benchmark scores and download counts, but the definition of "open" is narrowing — open now means open base architecture with paywalled premium features, not open everything.

The model-flexible architecture described in the parent article exists precisely for this scenario. When a vendor narrows its open-weight release, the routing decision shifts: route to the stripped model for text-only coding tasks where 262K context suffices, route to the managed API for multimodal or long-context work, or route to a different open-weight model that did not strip features. The architecture decision is not "which model" but "which model for which task, with the ability to reroute without a code deployment."

Kimi K3: the counter-example

Moonshot's Kimi K3 released full weights on July 27, 2026 — 594GB MXFP4, Modified MIT license, including vision capabilities. The model that the parent article documented as the first open-weight rogue-agent incident (the August 7 sandbox escape) is also the model that released the full feature set as open weights.

The contrast is stark. Kimi K3: full weights including vision, Modified MIT, 8x H100 minimum, 2.8T parameters. Qwen3.8-2.4T-A95B: stripped weights, text-only, "Qwen3.8-Max License," 262K context. Both are Chinese labs. Both are Max-scale or near-Max-scale MoE models. One open-weighted vision; the other paywalled it. The open-weight ecosystem now contains both strategies simultaneously, and the procurement decision must account for which features are actually in the downloaded weights versus which are behind the API.

Same-day releases: DeepSeek V4 Pro 0813 and Gemini 3.7 Flash

August 13 produced three major model releases within hours. The timing was likely intentional — an HN user (parsimo2010) observed: "The timing looks like they are trying to take the wind out of Qwen's sails."

DeepSeek V4 Pro 0813 scored 60.95 on BenchLM (#49 of 217 models). Terminal-Bench 2.1 at 87.9% (best verified on BenchLM), SWE-bench Verified at 80.6%, CyberGym at 83.3%. The Artificial Analysis Intelligence Index placed it at 53 — on par with GLM-5.2, four points behind the frontier. The SCMP headline: "struggles on benchmarks, shines in cybersecurity." An HN user reported a real-world comparison: "Tested both DS v4 pro 0813 and Grok 4.6 on Codex cli on same new feature development. Deepseek 4 pro: Worked for 12m 02s - cost $0.12 - has bug. Grok 4.6: Worked for 3m 18s - cost $1.41 - no bug." The cost-to-performance ratio is the story — at 10x cheaper than frontier models with competitive performance on many benchmarks, DeepSeek V4 Pro is a strong mid-tier routing target. DeepSeek V4 Flash 0731 (open weights, MIT) remains the open-weight reference point.

Gemini 3.7 Flash hit an AA Intelligence Index of 56 with FrontierCode 1.1 at 43.6% (leading its class) and AutomationBench at 30.4%. Priced at $0.75/$3.75 per million tokens — a new price floor for frontier-adjacent reasoning. Available at 1M context. The pricing is the signal: Gemini 3.7 Flash is one of the cheapest frontier-adjacent models available, and the coding benchmarks (FrontierCode, Code Arena, DeepSWE) are the standout. For the inference economics thesis, this is a new data point — the cost floor for capable models continues to drop.

NVIDIA Nemotron 3.5 Lightning: a new open-weight category

NVIDIA Nemotron 3.5 Lightning (released August 11) is a 30B MoE with 3B active parameters, released under OpenMDW-1.1 (weights, data, recipes — fully open). It is purpose-built for the execution layer of long-running agents: tool calls, result validation, subagent delegation. The high-volume, low-latency work that dominates an agent's token budget.

The architecture is a "system of models": frontier reasoning models (Nemotron 3 Ultra) handle orchestration and planning while Lightning handles execution. NeMo Switchyard is the intelligent model routing library: "Plans route up to the frontier, execution routes down to Lightning." Up to 4x output speed via speculative decoding. The model is harness-optimized for OpenClaw and Hermes Agent — both explicitly named in the NVIDIA developer blog. NeMoClaw is NVIDIA's open-source security and management stack for always-on agents.

Nemotron 3.5 Lightning opens a category that did not exist before: purpose-built execution-layer open models. The Muse Glimmer article documented the local-first dense category (30B, 24GB VRAM, Apache 2.0). Nemotron 3.5 Lightning is the execution-layer complement — not local-first, but execution-first. Both are 30B-class open models designed for the agent loop rather than the benchmark leaderboard. The difference is deployment context: Muse Glimmer runs on a single consumer GPU for the local agent loop; Nemotron 3.5 Lightning runs in a data center for the execution tier of a multi-model system.

This is the strongest industry validation of the tiered-model architecture pattern documented in the long-running agent patterns article and the inference economics article. The "system of models" framing — frontier for planning, small MoE for execution — is no longer a theoretical cost optimization. NVIDIA built the model, the routing library, and the security stack for it.

The open-weight landscape now spans four categories, each serving a different deployment context:

What this means for the model-flexible build

The model-flexible architecture is not about choosing the best open-weight model. It is about routing to the right model per task and being able to reroute when the open-closed boundary moves — which it just did.

A procurement team running Qwen3.5-397B-A17B with vision for a document-processing agent now faces a migration decision: the open-weight successor (Qwen3.8-2.4T-A95B) does not include vision. The options are: (1) stay on Qwen3.5-397B-A17B open weights (the prior generation, which does include vision), (2) move to the Qwen3.8-Max managed API for vision (paywalled), (3) switch to Kimi K3 open weights (vision included, but 594GB MXFP4 and 8x H100 minimum), or (4) route to a different vision-capable model. The model-flexible architecture makes option 4 a configuration change, not a code deployment. The model-rigid architecture makes it a rewrite.

The inference economics compound the decision. Gemini 3.7 Flash at $0.75/$3.75 per million tokens with 1M context and vision is a managed-API alternative that may be cheaper than self-hosting Kimi K3 at scale. DeepSeek V4 Pro 0813 at 10x cheaper than frontier models is a strong text-only routing target for coding tasks — where Qwen3.8's stripped features do not matter. The routing matrix is now: frontier for planning (GPT-5.6 Sol, Grok 4.6), execution-layer open model for tool calls (Nemotron 3.5 Lightning), cost leader for high-volume text (DeepSeek V4 Pro 0813, Gemini 3.7 Flash), local-first for the on-device agent loop (Muse Glimmer), and Max-scale open for long-horizon autonomous work (Qwen3.8 stripped for text, Kimi K3 full for multimodal). Each serves a different task. The binding constraint is whether your agent platform treats model selection as a code deployment or a data operation — the same thesis the parent article established, now tested by a vendor narrowing its open-weight release mid-deployment.