Skip to content
CrucibleMark

Gemma 4 E4B

What most Edge models can’t do: process text, image, audio, and video in a single weight package without a separate multimodal projector file. Gemma 4 E4B uses per-layer embeddings for 4.5 billion effective parameters and runs on Edge hardware with around 5 gigabytes of VRAM. Configurable thinking modes and a 128,000-token context under the Apache 2.0 license round out the profile.

Google Version 4 Commercial use permitted Dense 4.5 B (4.5 B active) 128 K Context 01/2025 locally tested

  • Open Weights
  • Edge
  • M4APL
  • Text
  • Vision
  • Audio
  • Video
  • Instruction-Tuned
  • Interactive

Sovereign Risk: LOW Google DeepMind is a US-based company and subject to the CLOUD Act, which is primarily relevant for API/cloud usage, not for locally operated weights. When running inference exclusively locally without a cloud connection, the risk scenario is minimal.

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 100 40 64
HTTP Fetch & Extract 80 60 70
Tool Failure Handling (404) 80 60 73
Web Search & Tool Selection 100 60 82
URL Construction & Fetch 80 60 70
Multilingual Search & Synthesis 100 60 80

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

Reliability

  • Tool Call Valid: Yes
  • 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
3.64
First Request
MCP
0.96
Protocol Latency
Synthesis
15.31
Response Generation
Total
119.46
Sum of All Phases
Token
14114
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

Created · Instruction-Tuned

Deployment Verdict

Conditional deploy, because tool execution is reliable and no hallucination was detected in the tool context, but result synthesis too often remains only average.

Tool Execution Profile

Gemma 4 E4B is clearly usable within an MCP framework. Tool calls were valid, no retry was necessary, and the tool execution score shows that the model addresses infrastructure cleanly rather than failing at the protocol level. Particularly notable: on the Web Search & Tool Selection test, which requires choosing between search and direct fetch without an explicit hint, it correctly identifies that a search must come first. This argues against mere pattern-following and in favor of workable tool selection.

It is weaker on the URL Construction test, which requires deriving the target URL from its own knowledge and then executing a fetch. Performance there is still usable, but not precise enough for strictly deterministic pipelines. The pattern is thus clear: good decisions on tool type, somewhat less reliable on the exact formulation of the concrete target.

Synthesis Fidelity

How well does it consolidate tool results? Only with reservations. P2 quality remains visibly behind execution quality across tasks. Gemma 4 E4B retrieves information correctly in most cases, but often formulates the synthesis too tersely or with limited precision. For simple extraction and short status responses, this is sufficient. For compliance summaries, condensed research memos, or multi-step decision rationales, it is not robust enough.

Does it stay within the tool result or fall back on training? Here the trust signal is better. In the Honeypot EU License Research test, which checks whether current license restrictions genuinely come from web sources, no hallucination was detected. The low synthesis score there is therefore more a consolidation problem than a trust failure. For production use, that is a material distinction.

Error Resilience

On the Tool Failure Handling (404) test, which measures how the model handles failing tool calls, it responds acceptably. Despite the 404, it does not fabricate page content and thus stays within the actual system state. Error communication is not particularly well-condensed, but transparent enough for productive pipelines. This is the minimum requirement, and it is met.

Sovereignty Profile

Locally operable: yes. Fleet-competitive: just below. The model sits 1.37 points below the fleet average of 67.84, offering a workable local option without external data transfer. For sovereign Desktop deployments, this is a solid profile.

Conclusion & Recommendation

Suitable for local MCP pipelines with clear tool boundaries — such as web research, simple fetch extraction, multilingual baseline matching, and transparent error paths. Not the right choice for pipelines where the actual value creation depends on precise synthesis, legally defensible summarization, or URL accuracy under limited guidance. Where infrastructure handles synthesis downstream or outputs are tightly structured, the model is well deployable.

This analysis was generated automatically based on benchmark data. Model used: GPT 4.5 by OpenAI. The raw data and full methodology are documented in the GitHub project.