# Sm · Small / Local Model — element 3 of 58

> Intelligence that fits in your pocket. Turns edge cases into on-device answers.

- **Group:** 1 · Intelligence
- **Necessity:** Optional
- **Price band:** Free
- **Maturity:** Experimental
- **Edition:** v2026.Q3 · verified 2026-09-13

## Leading tools (v2026.Q3)

- **Ollama** — local runtime of choice
- **Gemma 4 (E2B / E4B)** — the on-device workhorse
- **LM Studio** — local llms, real gui
- **Qwen3.5 Small** — tiny multimodal agents
- **llama.cpp** — the engine underneath everything

## Our take

The quiet trend: small models keep eating tasks that used to need frontier calls. Watch your inference bill for the signal.

## Combines with

Ow, Rt


## The top 5 — deep dossier (verified 2026-09-13)

Ollama, if you can only hold one — 8.9M monthly developers, presence in 85% of the Fortune 500, and a $65M Series B (Jul 2026) make it the de-facto local runtime, and local inference stays free and unlimited. Gemma 4 E-series is the model to load first: Apache-2.0, multimodal, and built for 2–4GB footprints. LM Studio wins when you want a GUI, Mac-optimized MLX speed, and free commercial use; llama.cpp when inference ships inside your own product; Qwen3.5 Small when your edge agent needs eyes and ears at under 1GB.

1. **Ollama** (Ollama Inc.) — Local: free, unlimited · Cloud free tier · Pro $20/mo · Max $100/mo (signups paused). Best for: The default way to pull and run any open model locally — one command, OpenAI-compatible API, and every agent framework already speaks it. Why: The category's Docker moment: 8.9M monthly developers, ~176k GitHub stars, used inside 85% of the Fortune 500, all built by a 14-person team that raised a $65M Series B in July 2026 ($88M total). In 2026 it re-platformed Apple Silicon inference onto MLX — up to 90% faster for coding agents (Jun 2026) — while keeping GGUF/llama.cpp compatibility. Watch: The cloud pivot ($20–100/mo tiers, datacenter models) is where monetization pressure lives — local stays free today, but the incentive gradient now points at hosted usage. Power users note the new app and MLX default reduced the old single-binary simplicity. [https://ollama.com](https://ollama.com)
2. **Gemma 4 (E2B / E4B)** (Google) — Free — Apache 2.0 weights (license upgraded from the restrictive Gemma terms). Best for: The first model to load on any laptop or phone — multimodal (text, image, audio), 128K context, 2–4GB effective footprint. Why: The on-device workhorse of 2026: released Apr 2, 2026 under Apache 2.0 — dropping the old Gemma license restrictions — with E2B (2.3B effective) and E4B (4.5B) variants that run fully offline on mobile and IoT hardware. The family has passed 400M downloads with 100k+ community variants, and E2B hits 60% MMLU-Pro — numbers that needed 30B+ models two years ago. Watch: Google's release cadence makes any Gemma pick obsolete in ~12 months, and the 26B/31B variants pull you out of small-model territory (see Ow). Audio input is limited to the smaller variants; long-context quality degrades past 32K on E2B in community testing. [https://ai.google.dev/gemma](https://ai.google.dev/gemma)
3. **LM Studio** (Element Labs) — Free incl. commercial use · optional cloud inference $0.13–3.00/M tokens · Enterprise custom. Best for: Teams that want local models with a real GUI — model discovery, chat, RAG, plus Python/TypeScript SDKs and a headless server mode underneath. Why: The most polished way to run GGUF and MLX models on a desktop, free even for commercial use. 2026 closed its two gaps with Ollama: llmster headless server mode (Jan 2026) for deployments, and the Bionic agent app (Jul 16, 2026) that puts an agentic layer over the same local runtime. Dual-engine (llama.cpp + MLX) with KV-cache checkpointing tuned for long-context agent loops. Watch: The core app is closed-source — you are trusting a VC-backed company's roadmap, and the Bionic launch signals focus shifting toward a cloud-billed agent product ('Bionic Pass' pricing still unannounced Aug 2026). CLI repo has ~5.1k stars vs Ollama's 176k — a fraction of the ecosystem gravity. [https://lmstudio.ai](https://lmstudio.ai)
4. **Qwen3.5 Small** (Alibaba) — Free weights on Hugging Face / ModelScope (Instruct + Base). Best for: Multimodal edge agents — the 0.8B and 2B run on phones and IoT chips with native image/video input; the 4B is the strongest small agent base of 2026. Why: Released Mar 2, 2026 as a purpose-built small family (0.8B / 2B / 4B / 9B) on the same native-multimodal Qwen3.5 foundation — a 0.8B model that processes video is the clearest marker yet of the edge-AI era. The 4B adds 262K context and on-demand thinking across 201 languages; the 9B closes on models 5–10x its size via scaled RL. Watch: License terms for the Small series weren't confirmed on a vendor page at review time (the Qwen3 line was Apache 2.0 — verify per model card). China provenance still blocks adoption in some Western enterprises regardless of open weights, and benchmark claims are mostly self-reported. [https://qwen.ai](https://qwen.ai)
5. **llama.cpp** (ggml-org (Georgi Gerganov)) — Free, MIT — vendor-neutral C/C++, no strings. Best for: Shipping inference inside your own product — the MIT-licensed engine that runs GGUF models on effectively every chip made. Why: The engine underneath the whole category: ~122.7k stars, 21k forks, 10k+ commits, and the GGUF quantization format it defined (1.5-bit to 8-bit) is the lingua franca of local AI. CUDA, Metal, Vulkan, HIP, x86, Apple Silicon, even RISC-V — plus a built-in OpenAI-compatible server and VLM support, with zero runtime dependencies to license or trust. Watch: It's a toolkit, not a product — model management, updates, and UX are on you (that's why Ollama and LM Studio exist). On Apple Silicon, MLX now decodes 1.4–1.8x faster, which pushed even Ollama to route safetensors to MLX in 2026. [https://github.com/ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp)

### How to choose
- If You just want open models running on your machine today → Ollama — one command, OpenAI-compatible endpoint, and every framework integrates with it. Ignore the cloud tiers until you need datacenter-size models.
- If You're choosing the model, not the runtime → Gemma 4 E4B as default (Apache 2.0, multimodal, 128K); Qwen3.5-4B when the agent needs image/video input; Phi-4-mini when you want MIT license and math/reasoning at 3.8B.
- If Inference ships inside your product → llama.cpp (MIT, every platform) — or MLX if you're Apple-only, where it decodes 1.4–1.8x faster and M5 neural accelerators cut time-to-first-token 4x.
- If You're building a mobile app → Use the platform stack, not a runtime port: Apple Foundation Models (free ~3B on-device, open to third-party models since WWDC 2026) or Google AI Edge/LiteRT with Gemma 4 E2B; Nexa SDK for Qualcomm NPUs.
- If Your frontier-model bill has line items for classification, extraction, or summarization → Move them to a small local model — NVIDIA's position paper argues most agentic subtasks are 'repetitive, scoped, non-conversational' and SLM-sized; this is the inference-bill signal the element's tagline is about.

### The field (18 more)

Phi-4 family, MLX, gpt-oss-20b, SmolLM3-3B, LFM2 / Liquid Nanos, Foundry Local, Apple Foundation Models, Google AI Edge / LiteRT-LM, Nexa SDK (acquired), Docker Model Runner, Jan, llamafile, LocalAI, GPT4All (fading), ExecuTorch, Ministral 3, Open WebUI, Ollama Cloud Max tier (sunsetting)

Full dossier data: https://elems.ai/e/sm.json

---
Source: [elems.ai](https://elems.ai/e/sm.html) — the periodic table of the AI-led startup. Data: https://elems.ai/elements.json (CC BY 4.0, cite elems.ai).