Google Launches Gemma 4, a 26‑B Multimodal LLM with Multi‑Step Planning and No‑GPU

Google’s Gemma 4 arrives as a Mixture‑of‑Experts (MoE) transformer that activates only a fraction of its total parameters per token, a design choice that underpins its “no‑GPU” claim. The 26‑billion‑parameter model is partitioned into 4‑billion‑parameter expert subnetworks, with the routing layer selecting the most relevant expert for each token during inference. According to the opencode write‑up, this selective activation limits active compute to roughly 4 B parameters, allowing the model to run on consumer‑grade CPUs and even on a MacBook without a discrete GPU. The same report notes that the model can be quantized to 4‑bit or 8‑bit precision, requiring 16–18 GB of RAM for the 4‑bit variant and 28–30 GB for 8‑bit, which fits within the unified memory of recent MacBook Pros.

Beyond its hardware efficiency, Gemma 4 adds a “thinking” mode that triggers multi‑step planning and deep logical reasoning. The Zoom Bangla News article describes this mode as an optional reasoning pass where the model generates intermediate latent states before producing a final answer, effectively allowing it to decompose complex queries into sub‑tasks. In practice, the model demonstrated tool‑calling capabilities, natively supporting function calls without external API scaffolding. The ummid.com report highlights a demonstration where Gemma 4 performed Linux‑version lookups, searched documentation, and returned accurate results, all while staying offline. This native tool usage is a direct extension of the multi‑step planning pipeline, enabling the model to orchestrate external utilities as part of its reasoning chain.

Gemma 4 is offered in four configurations: E2B, E4B, 26B‑A4B, and 31B. The 26B‑A4B variant is positioned as the sweet spot for local deployment, according to the opencode piece, because it balances model capacity with the reduced active‑parameter footprint. The same source reports that the model supports a context window of up to 256 K tokens, a substantial increase over typical 8 K‑16 K windows in earlier LLMs. This extended context is crucial for tasks that require long‑form reasoning or the ingestion of large codebases, and it dovetails with the model’s tool‑calling ability by allowing it to retain more state across multiple function invocations.

Benchmark data in the opencode article shows Gemma 4 performing competitively on reasoning and coding suites. While the piece does not provide raw scores, it states that the model “scores well” on standard benchmarks, suggesting that the MoE architecture does not sacrifice accuracy for speed. The report also mentions that the model’s performance scales with precision: the 4‑bit quantized version runs faster but may incur a modest drop in fidelity, whereas the 8‑bit version offers a middle ground. This flexibility gives developers the option to prioritize latency or precision based on their hardware constraints.

The release of Gemma 4 signals Google’s broader push into multimodal, locally runnable AI. By combining MoE efficiency, extended context, and built‑in tool calling, the model aims to bridge the gap between cloud‑only LLM services and on‑device intelligence. As the opencode write‑up observes, the ability to run a 26‑billion‑parameter model on a laptop without a GPU could democratize access to advanced reasoning capabilities for developers, researchers, and hobbyists who lack cloud budgets. However, the model’s reliance on CPU‑heavy routing and the memory demands of 16‑30 GB still set a floor that excludes lower‑end hardware, a limitation that Google will need to address in future iterations if it hopes to achieve truly universal local AI deployment.