8B total / 4B effective (dense Gemma 4 E4B)
65,536 tokens (64K)
OpenMDW License Agreement, version 1.1
Open-weight compact text-generation model further post-trained from Gemma 4 E4B IT for fast local conversational agents, voice and device assistants, interactive characters, and robotic, industrial, and embedded interfaces. Despite its E4B designation, the checkpoint contains approximately 8.0B parameters; NVIDIA describes it as 4B effective and markets approximately 4GB as a target runtime GPU-memory footprint rather than its parameter count or download size. Its 42-layer text backbone uses a 2,560-dimensional hidden state, 10,240-dimensional feed-forward layers, eight query heads, two key/value heads, 256-dimensional attention heads, tied embeddings, a 262,144-token vocabulary, and a hybrid pattern of five 512-token sliding-attention layers followed by one full-attention layer, producing seven full-attention layers across the network. The published configuration supports 65,536 tokens and shares key/value projections across 18 layers. Although the Gemma 4 conditional-generation configuration retains vision and audio configuration blocks, NVIDIA specifies text input and text output and disables image, video, and audio inputs in its serving recipe. Post-training used multi-environment RLVR and RLHF through NeMo-RL and NeMo-Gym without an SFT stage, covering conversational tool use, tool schemas, instruction following and hierarchy, structured output, and formatting. Training used a global batch of 256, eight rollouts per prompt, maximum generation lengths of 4,096 tokens for reasoning samples and 1,024 for non-reasoning samples, and a deliberate 2:1 non-reasoning-to-reasoning ratio. The model supports thinking on and off, runtime thinking-budget control, structured multi-turn tool calls, and an optional conversational mode that encourages concise, speech-ready answers. NVIDIA reports BF16 results including 53.10 on MultiChallenge, 67.05 on RoleBench Custom, 68.51 on BFCL-v3, 55.10 on Tau2-Bench, 84.79 on IHEval, 63.79 on IFBench, 86.99 on RULER at 64K, 72.10 on LongMemEval, and 63.16 on RelyToolBench; it wins 15 of the technical report's 21 comparisons against Gemma 4 E4B IT, Qwen3.5-4B, and Nemotron 3 Nano 4B. NVIDIA says its average score is roughly 40% above Nemotron 3 Nano 4B and approximately 10% above evaluated peers while producing up to 80% fewer tokens in reasoning-enabled comparisons. The canonical BF16 target occupies approximately 15.88 GB plus a 110 MB frozen shared-KV file. NVIDIA's Q4_K_M GGUF occupies approximately 6.18 GB and uses importance-matrix calibration over roughly one million tokens of RL rollouts; measured peak GPU memory ranges from 3.5 GB for a short RTX 5070 workload to approximately 5.0 GB for a 32K-cache Jetson AGX Orin workload. Companion DFlash draft models propose token blocks in parallel without changing the target distribution; NVIDIA reports up to 2.74 times per-user decode speed versus inference without speculative decoding and 1.21 times the batch-one generation throughput of Gemma 4 E4B IT with its original MTP drafter. The official DFlash variants include a 736.8 MB BF16 GGUF and a 580.3 MB NVFP4 safetensors checkpoint. ggml-org additionally publishes automatic BF16 and Q8_0 GGUF target conversions and an NVFP4 DFlash conversion. The ggml-org conversion metadata labels the repository Apache-2.0, but the underlying NVIDIA target and draft checkpoints are governed by OpenMDW 1.1, which is therefore the authoritative model license.