Gemma 4 E2B (Unsloth)

Gemma 4 E2B is the most compact Gemma 4 model and the entry point into the new family under the Apache 2.0 license. PLE architecture with 5.1B total and 2.3B effective parameters, 128,000 tokens of context, multimodal input for text, image, and audio, plus built-in reasoning and function calling. Runs locally as Unsloth GGUF on Edge- and phone-class hardware.

Google Version 4 Commercial use permitted Dense 5.1 B (2.3 B active) 128 K Context 01/2025 locally tested

Open Weights
Nano
llama.cpp
Text
Vision
Audio
Instruction-Tuned
Real-Time

Sovereign Risk: LOW TODO

Tool-use details

Asset performance, reliability, and runtime profile

Name	P1	P2	Combined
EU License Research	35	0	14
HTTP Fetch & Extract	80	60	70
Tool Failure Handling (404)	80	60	73
Web Search & Tool Selection	100	80	91
URL Construction & Fetch	80	80	80
Multilingual Search & Synthesis	35	20	27.5

Name

Combined

EU License Research

HTTP Fetch & Extract

Tool Failure Handling (404)

Web Search & Tool Selection

100

URL Construction & Fetch

Multilingual Search & Synthesis

27.5

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

Reliability

Tool Call Valid: No

Retry: Not required

Hallucination: Not 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: 1.03; First Request

MCP: 0.47; Protocol Latency

Synthesis: 4.41; Response Generation

Total: 35.46; Sum of All Phases

Token: 10477; 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.

Deployment Verdict

Conditional deploy: The model shows usable tool orientation, but is not reliable enough for autonomous MCP pipelines due to invalid tool calls and weak synthesis fidelity.

Tool Execution Profile

Gemma 4 E2B recognizes tool requirements not merely schematically, but with genuine situational adaptation. In the Web Search & Tool Selection test — which checks whether search is chosen over fetch without any explicit hint — it makes the correct tool selection reliably. In the URL Construction test, which measures the independent derivation of a target URL followed by a fetch, it performs fundamentally correctly, though not precisely enough for strictly deterministic workflows. The overall picture is therefore better than the raw score suggests: the model understands when search is necessary and when direct retrieval suffices.

The operational catch lies not in selection but in protocol adherence. Tool call valid is false. For production use, this means: the planning logic is usable, but the handoff to MCP infrastructure is not stable enough. Since no retry was required, this points more toward unreliable execution output than a mere formatting issue on the first attempt.

Synthesis Fidelity

How well does it consolidate tool results? Only to a limited degree. P2 performance is the primary argument against broad deployment. Solid scores on HTTP Fetch & Extract and URL Construction & Fetch show that the model can still consolidate individual tool results reasonably well. As soon as the task demands more selection, edge cases, or multilingual content, consolidation quality visibly deteriorates.

Does it stay within the tool result or fall back on training data? The trust signal here is mixed. In the Honeypot EU License Research test — which checks whether current license restrictions are actually retrieved from web sources — the model scores P2=0. On the positive side, no hallucination was flagged. The actual production signal is negative: it delivers no reliable, source-bound consolidation precisely where compliance-relevant currency is critical.

Error Resilience

The model responds acceptably to tool failures. In the 404 test, which checks for transparent error communication versus fabricated replacement content, it does not hallucinate page content. This matters for production. A failed call therefore does not automatically compromise the factual basis of the entire response. Error communication is not strong, but sufficiently safe.

Sovereignty Profile

Locally deployable with Apache 2.0 weights and therefore sovereign in operation. At a combined score of 59.25, it sits 8.50 points below the fleet average of 67.75. The advantage is control over runtime and data path — not fleet-competitive tool performance.

Conclusion & Recommendation

Suitable for local, data-sensitive pipelines with a human in the loop, clearly defined tool paths, and low compliance burden. Not suitable for autonomous research workflows, multilingual synthesis, license or policy checks, or any pipeline in which MCP calls must be strictly valid and content reliably consolidated. If you deploy it, treat it as a low-cost local executor operating under a tight guardrail and post-validation layer.

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.