Gemma. Laguna. Beam.
The essentials, without the noise.
Beam for the strongest agentic capability of the three. Laguna for efficient coding. Gemma for compact versatility. None of these three models demonstrates world-leading absolute capability.
Which team would we back?
This study assesses the Gemma team solely through Gemma's published results. These are judgments about demonstrated strengths, not a controlled ranking of training teams. None of the three has demonstrated absolute frontier leadership through the models studied here.
| Model studied | Our assessment | 2026 release |
|---|---|---|
| Gemma 4 Gemma team · Google | Strong compact-model versatility and multimodality. The published Gemma results do not demonstrate absolute frontier leadership. | April 2 31B / 26B-A4B |
| Laguna-S 2.1 Poolside | An efficient-coding bet: 118B total, 8B active parameters. Absolute frontier capability remains unproven. | July 21 |
| Beam Reflection | Our provisional pick for raw agentic capability within this trio. A credible challenger, not the global leader. | October 5 Scores revised October 8 |
Gemma and Beam launched roughly six months apart: these are different generations. Sources: Google · Poolside · Reflection.
The actual frontier? Still a gap.
Long, difficult tasks reveal more than near-saturated benchmarks. Even Reflection's own comparison places Beam behind leading open models.
| Benchmark · % | Beam | Kimi K3 | DeepSeek V4.1 Flash |
|---|---|---|---|
| DeepSWE v1.1 · long-horizon engineering | 44.4 | 68.0 | 74.2 |
| Terminal-Bench 2.1 | 80.1 | 88.3 | 90.6 |
| AutomationBench public | 37.0 | 46.7 | 54.8 |
| HLE without tools | 36.2 | 46.9 | 39.1 |
Indicative, not controlled: results compiled by Reflection, without guaranteed matching agents, budgets or settings. The gaps point in the same direction, but do not independently measure training-team quality.
Independent reference: the DeepSWE leaderboard reports roughly 74% for GPT-6 Astra and Claude Opus 5, both using mini-swe-agent. These are external reference points, not evidence about any of the three teams studied. Beam and Laguna are absent from the table consulted, so this is not a controlled head-to-head with them.
The practical decision
- For maximum capability: do not restrict testing to this trio; include global leaders.
- Within the trio: test Beam first for raw capability, Laguna-S for coding versus cost, and Gemma for multimodality and deployment constraints.
- For a frontier-challenger bet: Reflection ahead of Poolside, with moderate confidence. Poolside for a specialized strategy.
A great model at a given cost, large-scale training, and a world record are three different achievements. Public results do not identify the best team at equal training budgets.
Additional benchmarks — coding and reasoning
Published scores, not our own evaluations. Higher is better. A shared benchmark name does not guarantee the same agent, inference budget or settings. “—” means no aligned score included here, not zero.
Coding and agents
| Benchmark · % | Gemma 4 31B | Laguna-S 2.1 | Beam |
|---|---|---|---|
| SWE-bench Verified | 52.0 | — | 80.9 |
| SWE-bench Multilingual | 51.7 | 78.5 | 78.0 |
| Terminal-Bench 2.0 | 42.9 | — | — |
| Terminal-Bench 2.1 | — | 70.2 | 80.1 |
| DeepSWE v1.1 | — | 40.4 | 44.4 |
| SWE Atlas Codebase QnA | — | 46.2 | 34.6 |
Indicative cross-source comparison. Gemma results come from Qwen's evaluations; Laguna from Poolside; Beam from Reflection. Laguna's DeepSWE result uses its “pool” agent. Do not treat the small DeepSWE gap as a controlled win, or compare Terminal-Bench 2.0 with 2.1.
Reasoning and knowledge
| Benchmark · % | Gemma 4 31B | Beam |
|---|---|---|
| AIME 2026 | 89.2 | 97.8 |
| GPQA Diamond | 84.3 | 90.5 |
| HLE without tools | 19.5 | 36.2 |
| SciCode | 43.0 | 49.7 |
| IFBench | 76.0 | 79.7 |
Vendor-reported results from Google, its technical report, and Reflection. Not a matched-budget evaluation. No corresponding Laguna scores included.
Comparison caveats
- Laguna-S reports 40.4 on DeepSWE v1.1 using its “pool” agent, versus mini-swe-agent on the independent leaderboard: no direct ranking. Poolside methodology.
- Do not mix Terminal-Bench versions or base and post-trained models.
- Active parameters and token counts do not establish actual serving costs. Efficiency hypotheses need validation on the intended tasks and infrastructure.
- A product score reveals neither the recipe nor the full training budget. Conclusions about the Gemma team use Gemma evidence only; results from other Google teams are outside this study's scope.