Skip to content
CrucibleMark

Laguna S 2.1

Laguna S 2.1 has been Poolside’s third consecutive open coding model since July 2026, a MoE with 117.6 billion total and only 8.5 billion active parameters per token. The Linux Foundation’s OpenMDW-1.1 license permits free commercial use and modification. The model supports 1,024,000 tokens of context, interleaved thinking with tool calls within the chain of thought, and native tool calling for long-horizon agentic workflows.

Poolside Version 2.1 Commercial use permitted MoE 117.6 B (8.5 B active) 1024 K Context 11/2025 $0.1 / $0.2 per 1M

  • Open Weights
  • Server
  • vLLM
  • Text
  • Long Context
  • Instruction-Tuned
  • Batch

Sovereign Risk: LOW Poolside is a US-based company. Publication under the open OpenMDW-1.1 license, hosted by the Linux Foundation, enables independent auditing, modification, and redistribution of the weights for both commercial and non-commercial purposes. With fully local deployment, any dependency on US cloud infrastructure is eliminated, resulting in a low risk rating.

Tool-use profile: 6 assets in detail

Comparison of asset performance (P1/P2/Combined) against the fleet average

Asset performance (radar)

Score Breakdown vs. Fleet Average

Tool-use details

Asset performance, reliability, and runtime profile

Name P1 P2 Combined
EU License Research 40 15 25
HTTP Fetch & Extract 80 15 47.5
Tool Failure Handling (404) 80 80 80
Web Search & Tool Selection 95 80 88.3
URL Construction & Fetch 80 80 80
Multilingual Search & Synthesis 95 15 55

CrucibleMark evaluates tool use across 6 independent tests. Click a test name for details.

Reliability

  • Tool Call Valid: No
  • Retry: Not required
  • Hallucination: Detected

Reliability measures how consistently a model actually executes tool calls: Tool Call Valid schema and format accepted, Retry Required successful only after retry, Hallucination Flag fabricated tools or parameters detected. All three green means production-ready.

Runtime profile

Call 1
35.92
First Request
MCP
1.11
Protocol Latency
Synthesis
56.54
Response Generation
Total
561.38
Sum of All Phases
Token
18308
Input + Output
Cost
$0
Cost per Run

The runtime profile shows the latency and cost metrics for the model run: Call 1 First Request, MCP Protocol Latency, Synthesis Response Generation, Total sum of all phases, supplemented by Token input and output and Cost cost per run.

Tool-use review

· Long Context · Instruction-Tuned

Deployment Verdict

Conditional deploy: tool usage works operationally in many cases, but a detected hallucination and an invalid tool call mean the model does not have a sufficiently reliable trust profile for unsupervised MCP pipelines.

Tool Execution Profile

Laguna S 2.1 demonstrates genuine tool selection rather than pure pattern-following. On the Web Search & Tool Selection test — which evaluates the choice between search and direct retrieval without an explicit hint — it identifies the need for web_search with high confidence. This points to workable planning in open retrieval flows. On the URL Construction & Fetch test, which measures the derivation of a target URL from prior knowledge and the subsequent fetch, it remains usable but not deterministic enough. The gap between very strong tool selection and only solid URL construction reveals that the model frequently understands the next step correctly but does not always produce protocol-clean execution. This is confirmed by the tool_call_valid=false finding. No retry was required, so the issue lies in first-pass execution rather than a pure formatting error.

Synthesis Fidelity

How well does it compress tool results? Poorly. P2 performance at 33.33 is the clear bottleneck. HTTP Fetch & Extract — designed to pull precise facts from real page content — drops to 15. Multilingual Search & Synthesis also falls well short of production requirements at the synthesis stage despite successful retrieval. The model frequently finds material but fails to compress it reliably into a clean response.

Does it stay grounded in tool output or fall back on training data? Only partially trustworthy. On the Honeypot EU License Research test, which checks whether current licensing restrictions are actually retrieved from web sources, it does not hallucinate overtly, but at P2=20 it is too weak substantively to qualify as reliable compliance synthesis. In addition, one hallucination was detected globally. This is not a cosmetic flaw but a security risk: once a model presents fabricated facts as tool output, it undermines the entire pipeline.

Error Resilience

Acceptable for production. On the 404 test, which measures transparent behavior when a tool call fails, Laguna S 2.1 communicates the error cleanly and does not fabricate page content. This is the minimum requirement for robust tool pipelines, and the model meets it here.

Operational Profile

Call 1: 4.85s. Call 2: 31.33s. MCP latency: 0.80s. Total: 221.90s. Slow relative to the response quality achieved. Cost/run: local. Inexpensive to operate, but time costs are high.

Conclusion & Recommendation

Suitable for assisted agent runs with human review, exploratory web research, and workflows where tool selection matters more than final synthesis. Not suitable for compliance, policy, or executive summary pipelines where responses must remain strictly grounded in tool evidence. If you deploy it, do so only with hard response validation, tool call schema checks, and a downstream verifier for every synthesized claim.

This analysis was generated automatically based on benchmark data. Model used: GPT-5.4 by OpenAI. The raw data and full methodology are documented in the GitHub project.