Qwen 3.6 27B vs Laguna XS 2.1 Vision: which local LLM is better?

Until now, Qwen 3.6 27B was the easy answer. It is dense, natively multimodal, realistically quantizable for an RTX 5090, and already served successfully in our local planner test. Laguna XS 2.1 was faster and cheaper in our newer hosted comparison, but the upstream model was text-only. The new Laguna XS 2.1 Vision NVFP4 changes the shape of that decision: Laguna is now a real local multimodal contender, although its vision package is still experimental and not yet a drop-in serving release.

The short answer

If you want one proven local model for text, screenshots, and video today, choose Qwen 3.6 27B. It has native multimodal integration, mature serving instructions, and a practical NVFP4 path for a 32GB RTX 5090.

If you care most about speed, agentic efficiency, and the upside of an American open-weight coding model that can now be given sight, Laguna XS 2.1 Vision NVFP4 is suddenly a strong contender. Its MoE architecture is a major advantage: the model stores 33B parameters but activates only 3B per token, so it can use far less per-token compute than dense Qwen. Our same-payload text benchmark already put Laguna ahead on exact-action peer consensus, median latency, and hosted replay cost.

But this is not a new vision benchmark. Our measured Laguna row was non-vision: the tested OpenRouter route accepted text only, and the suite sent no images to any model. The new checkpoint changes Laguna's capability surface, not the historical scores. End-to-end image inference and NVFP4 equivalence for the vision package are also still pending.

Our verdict is therefore:

• Best ready-to-use local multimodal model: Qwen 3.6 27B.
• Best architecture for raw speed potential: Laguna XS 2.1 Vision.
• Potential winner after serving and visual-agent validation: genuinely open.

Two very different architectures

These models occupy a similar local-deployment conversation, but they do not spend compute in the same way.

The Qwen model card describes a dense 27B language model with 64 layers. Its repeating pattern is three Gated DeltaNet layers followed by one gated full-attention layer. That hybrid cuts the cost of long-sequence processing without routing tokens through different feed-forward experts. Every token still uses the dense 27B backbone.

Qwen's multimodality is part of the released model architecture. Its 27-layer vision encoder uses 1,152-dimensional hidden states, 16-pixel patches, spatial 2x2 merging, and a 5,120-dimensional output that matches the language backbone. Image and video token IDs, processor behavior, and serving examples are already wired into the checkpoint.

Poolside's Laguna XS 2.1 attacks efficiency from the other direction. It stores 33B total parameters but activates only 3B per token. Thirty of its 40 layers use a 512-token sliding window; ten use global attention. Its 256 routed experts give the model much more stored specialization than its active compute suggests.

That MoE design is a plus, especially for speed. Qwen must run its dense 27B backbone for every token; Laguna routes each token through a small subset of experts and activates only 3B parameters. With a serving engine that has efficient MoE routing and kernels, Laguna should need substantially less language-model compute per generated token and can deliver lower latency or higher throughput. Our hosted measurement is consistent with that architectural advantage: Laguna reached a 1.15-second median versus Qwen's 2.23 seconds, with a much tighter p95.

Those latency numbers came from cloud routes, not the same physical GPU. In our view, the difference could become even larger on a single RTX 5090 once Laguna has an optimized local runtime: Qwen has to execute a dense 27B language model for every generated token, while Laguna's router activates roughly 3B. A local head-to-head would also remove cloud-provider scheduling, batching, network, and route differences. That is an architectural expectation, not a result we have measured yet; the Laguna Vision package still needs its end-to-end 5090 serving path.

The new WebBrain vision checkpoint leaves that Laguna backbone frozen. It also freezes a 27-layer MoonViT tower from Moonshot AI's Kimi K2.6. The only newly trained component is a 30.68M-parameter projector: layer normalization, a 2x2 patch merge, then two linear transformations with GELU to turn MoonViT features into Laguna tokens. The projector was trained on 100,000 examples, with up to 512 merged image tokens inside 2,048-token training sequences.

That distinction matters. Qwen is an integrated multimodal release. Laguna Vision is a modular graft: an efficient American text backbone, a Chinese vision tower, and a compact WebBrain-trained bridge between them.

What the WebBrain benchmark actually says

Our American-Chinese open-model frontier benchmark sent the same 100 WebBrain first-action cases to thirteen OpenRouter routes. The primary score compared each model's normalized action with all twelve peers; no single reference model acted as judge.

Here are the directly comparable Qwen and Laguna rows:

This is a split decision.

Laguna was faster, cheaper, more schema-reliable, and three places higher in exact-action consensus: seventh versus ninth. Qwen chose the benchmark's ideal tool more often and produced more exact ideal actions. Qwen also had nine schema-invalid calls, mostly malformed accessibility filters, while Laguna had one invalid press_keys argument.

The table may actually understate Laguna's local speed advantage. It compares two independently hosted cloud routes, so it cannot isolate architecture from provider infrastructure. On identical RTX 5090 hardware with equally mature NVFP4 kernels, we would expect Laguna's 3B-active MoE to separate further from Qwen's dense 27B on token-generation speed. The necessary qualification is equally important: Qwen's local path already works, while that Laguna measurement has not been run.

Our earlier frozen Sonnet-reference Laguna test was more negative: 65% all-case Sonnet alignment, six exact first actions, and three malformed tool names. That verdict was fair for that route and harness, but it was never a vision result. The upstream Laguna model was text-to-text, the endpoint exposed no image input, and the first-action test did not measure screenshot understanding. The newer thirteen-model rerun also remained text-only for Laguna, but its peer-consensus result showed that the model was more competitive than the original one-reference framing suggested.

So the fair reading is not “Laguna beat Qwen.” It is this: Laguna's text backbone already had enough planner quality and efficiency to belong in the conversation, while Qwen retained the more mature and more precise multimodal package. Adding a vision path removes Laguna's biggest structural negative.

Why the vision checkpoint changes everything

A browser agent cannot live on text alone. Accessibility trees and extracted page text miss canvases, charts, visual error states, selected controls, layout relationships, and interfaces with poor semantics. A text-only Laguna route could be an efficient planner, but it could not be WebBrain's only local model.

Laguna XS 2.1 Vision NVFP4 changes that product equation. The repository now contains:

• the exact pinned Poolside Laguna XS 2.1 NVFP4 backbone;
• a frozen 416.9M-parameter MoonViT vision tower;
• the final 30.68M-parameter PatchMerger projector trained on 100,000 examples;
• fingerprints and machine-readable provenance for the packaged artifacts.

In architectural terms, the missing bridge now exists. Laguna is no longer disqualified from the multimodal local-model shortlist simply because the upstream checkpoint cannot see.

That does not mean the work is finished. The model card intentionally does not publish a serving command yet. Multimodal configuration, processor and serving integration, NVFP4 equivalence, single-GPU loading, and final image-inference validation remain open. The target serving recipe is currently a Blackwell RTX PRO 6000 96GB configuration, not the already demonstrated 32GB Qwen workstation path. Fine-grained OCR, small controls, GUI grounding, and hallucination calibration also need broader evaluation.

“Changes everything” therefore means Laguna has moved from structurally incomplete to technically plausible as a full browser model. It does not mean an unvalidated development checkpoint has already defeated Qwen.

Which one should you run?

The hardware distinction is important. A 3B-active MoE does not mean only 3B parameters need to be stored. Laguna still carries a 33B backbone, and the vision tower, projector, KV cache, serving engine, and context length all consume memory. But once the package is resident, the low active-parameter count is a genuine compute and speed advantage. Quantization and MoE-aware runtime design determine how much of that theoretical advantage appears on a particular machine.

American and Chinese open weights are almost head-to-head

The geopolitical picture has changed quickly. In our thirteen-model benchmark, Chinese open-weight models still held the first two positions and six of the top nine. Qwen remained our best single-RTX-5090 recommendation because it combined ownership, multimodality, and practical deployment better than anything else in the set.

Laguna XS 2.1 narrowed the American side of the gap even before vision: a 33B/3B-active model placed seventh, responded in 1.15 seconds at the median, and cost seven cents for the replay. Giving that backbone a credible path to visual understanding makes Poolside's model a real competitor rather than a clever text-only specialist.

American and Chinese open-weight models are now almost head-to-head in the local-agent conversation: Qwen leads on integrated multimodal maturity and consumer-GPU readiness; Laguna challenges on active-compute efficiency, latency, hosted cost, and agentic specialization. If the Laguna vision package passes its remaining serving and visual-quality gates, the choice will come down to workload and hardware instead of a categorical capability gap.

There is also a useful complication: Laguna Vision itself crosses the national boundary. Poolside supplies the American language backbone, Moonshot supplies the Chinese MoonViT tower, and WebBrain supplies the projector and packaging. Open-weight progress is becoming competitive and collaborative at the same time.

Bottom line

Qwen 3.6 27B is still better for most people today. It is the model to choose when you want a tested, integrated local VLM on a 32GB-class GPU and cannot spend time finishing a serving stack.

Laguna XS 2.1 is now a genuinely strong contender. Its MoE architecture is not merely a size trick: activating 3B rather than 27B language parameters gives it a real path to beating Qwen on latency and throughput. Its same-payload planner row was already faster and cheaper than Qwen's, and the new MoonViT projector removes the biggest reason we previously could not recommend it as a complete browser model.

The next decisive test is no longer another text-only first-action replay. It is a controlled, end-to-end visual-agent benchmark on the finished NVFP4 package: screenshots, OCR, charts, rich editors, spatial grounding, tool-call validity, latency, and memory on real local hardware. Until then, Qwen wins the deployment decision—but Laguna has made the race real.

Tags: #Qwen36 #LagunaXS #Poolside #MoonViT #NVFP4 #LocalLLM #LocalVLM #OpenWeights #BrowserAgent #WebBrain