Gemma 4 E4B

Nemotron 3.5 Bolt E4B 4GB is a derivative model obtained through post-training on top of Gemma 4 E4B, not a new architecture trained from scratch.

Perplexity currently offers three local orchestration models: Gemma 4 E4B (4-bit quantized, 6.6GB, requires 16GB of memory), Qwen 3.5 35B (requires 32GB of memory) and Perplexity's in-house 19GB model (requires 32GB of memory).

NVIDIA fine-tuned Nemotron 3.5 Bolt E4B based on Gemma 4 E4B, surpassing the original Gemma 4 E4B on 16 of 21 NVIDIA tests.

Nemotron 3.5 Bolt E4B scores 63.2 on Tau2-Bench, while the original Gemma 4 E4B scores 38.7.

The problem of DPO training previously crashing for Gemma 4 E2B and E4B (including Nemotron versions) has been fixed.

Pricor's processing pipeline is: first a rule analyzer finds the header row, row types, column types and number formats, then the fine-tuned multilingual model Laya assigns each column to a closed set of roles, columns Laya is unsure about go to Gemma 4 E4B running locally via Ollama, and anything still uncertain is handed to the user to answer.

On the same machine, the author compared three Gemma variants on 76 synthetic cases: Gemma 4 E4B reached 97%, Gemma 3 4B reached 78% with 5 false positives, and Gemma 4 12B reached 80% with 32 call timeouts.

On a laptop with an Intel Core Ultra 5, 16GB of RAM, and no discrete GPU, setting OLLAMA_IGPU_ENABLE=1 raised Gemma 4 E4B's prompt processing speed from about 31 tokens/s to about 315 tokens/s.

With integrated GPU acceleration enabled, reviewing a full game takes about 90 seconds, of which Stockfish analysis takes about 20 seconds and each comment takes about 12 seconds.

Deduplicating the Q/K/V weights saved 488 MiB of VRAM, and keeping the text embedding table on the CPU saved 594 MiB, enabling the two models Gemma 4 E4B and Qwen3-TTS 1.7B to be loaded simultaneously, with the graphics card showing 7,448 MiB used out of 8,188 MiB.

This local voice assistant system is fully self-hosted, with voice detection and recognition running on the CPU, the LLM (Gemma 4 E4B, QAT q4_0 quantization, based on llama.cpp) responsible for generating answers, and the TTS (Qwen3-TTS 1.7B) synthesizing speech sentence by sentence.

DecisionShield adopts a separated frontend-backend architecture: the frontend uses React+Vite+Tailwind CSS+Recharts+Framer Motion, the backend uses Python+FastAPI+Pydantic+Uvicorn, the AI layer uses 4-bit quantized local inference of Gemma 4 E4B, and the MVP stage uses in-memory storage.

DecisionShield uses the Gemma 4 E4B model for semantic reasoning and classification; Gemma generates structured JSON, which after Pydantic validation is handed to a Python decision engine to compute the score.

Gemma 4 E4B runs locally via Ollama on the author's RTX 5060 laptop, analyzing photos one by one and answering seating-related questions in a fixed JSON format, with photos whose format does not conform recorded directly as failures.

In the author's manual verification of 40 Mainz photos, when identifying seats Gemma 4 E4B found 2 true seats with 1 false positive, while the map (using the criterion of a bench within 15 meters of the camera) found only 1 with 4 false positives.

In identifying steps and shade, Gemma performed poorly: E4B falsely reported steps in 16 photos containing no steps, and falsely reported shade in 26 photos containing no shade, so the route planner uses only the seating answer.

In terms of inference speed, Gemma 4 E4B takes approximately 0.9 seconds per photo and E2B approximately 0.6 seconds, completing 165 photos within three minutes.

During development, the model server crashed twice (error 0xc0000409) when E4B was first loaded, because only half of the model could fit into available VRAM, and the first photo took 77 seconds before stabilizing at around 1 second.

In a robot dog test using Gemma 4 E4B, VicOne LAB R7 found that text on a poster was treated as an instruction and changed the robot dog's movement.