EvalSignal 003 / Independent model evaluation
American open weights are competitive again
On July 2, 2026, Poolside released Laguna XS 2.1, a compact coding model with 33B total parameters and only 3B active per token.
On July 15, Thinking Machines introduced Inkling and previewed Inkling Small, a 276B-total, 12B-active model with text, image, and audio input.
They are very different releases, but together they make a credible case: American open-weight models are competitive again.
A fixed WebBrain evaluation gave 13 models the same 100 browser states, user instructions, tool definitions, and output limits. It compared 1,300 first-action planner calls through peer consensus rather than relying on a single model as judge.
Next in EvalSignal: what is changing at China's model frontier, from DeepSeek and Tencent's inexpensive text planners to Kimi and GLM's broader frontier challenge.
The most useful American result may not be the model that finished first.
It is Poolside Laguna XS 2.1: a 33B-total, 3B-active model that ranked seventh among 13 models, responded in 1.15 seconds at the median, and cost $0.073 for the complete 100-case replay.
| Model | Exact | Valid | Median | Cost |
|---|---|---|---|---|
| Laguna XS 2.1 33B / 3B active, text |
41.3% | 88/89 | 1.15s | $0.073 |
| Inkling Small 276B / 12B active, text/image/audio |
34.8% | 84/84 | 1.05s | $0.243 |
| Qwen 3.6 27B 27B dense, text/image/video |
38.1% | 83/92 | 2.23s | $0.670 |
Laguna XS changes the US open-weight discussion
Laguna XS is not a universal browser model. Its tested route was text-only, it selected the expected tool on 31 cases, and its argument precision remained weaker than the leaders.
But it is compact, fast, downloadable, locally plausible, and almost free to operate. It placed between MiniMax M3 and Kimi K3 on exact peer consensus. That makes it a credible planner for text-state agents, coding workflows, and systems that route visual work to a separate model.
Inkling chooses capability breadth
Thinking Machines' preview of Inkling Small ranked lower, but supports text, image, and audio with a 12B-active footprint. Its 1.05-second median was the fastest in the benchmark.
Laguna offers stronger first-action consensus. Inkling offers a broader input surface. They represent two viable American open-weight strategies rather than one obvious winner.
Qwen remains our browser-focused local pick
For a private workstation, Qwen 3.6 27B remains our preferred browser-agent model for a single RTX 5090. It supports image and video input, performs slightly better on browser screenshot understanding, and has an advantage in coding and tool-oriented work that matters to browser automation.
The warning is tool reliability. Nine calls in this hosted run failed schema validation, mostly because of malformed accessibility filters. Production deployments should validate arguments and retry malformed calls.
Gemma 4 31B is the strongest local alternative.
Qwen has a modest advantage in vision, coding, and browser-agent work. Gemma offers broader multilingual capability and is particularly attractive for translated pages, cross-language research, and language-heavy analysis. Gemma was not part of the thirteen-model planner run; this comparison comes from a separate local browser-vision evaluation.
The American open-weight gap is narrowing, but it has not closed. Chinese models still occupy the first two consensus positions and provide the stronger open-weight multimodal frontier. The meaningful change is that developers now have credible American downloadable models at compact operating points.
Related: Qwen and Gemma browser-vision comparison
Supported by WebBrain
WebBrain is an open-source browser agent for Chrome, Firefox, and Edge. Run models such as Qwen and Gemma locally, or connect to a hosted provider.