Most of the people I teach and work with are trying to get a capable model onto a phone, a laptop, or a classroom machine that still works when the network drops. They care about memory bandwidth, licenses they can ship, and whether a student can reproduce the result. The models that matter for that work sit in the 2 to 12 billion parameter range, run locally, and escalate the hard queries to a larger cloud model when they have to.
If you care about licenses, memory budgets, and what “open” actually means on a model card, it is worth getting specific.
Gemma 4, and why the license change matters
Google DeepMind shipped Gemma 4 on April 2, 2026 under Apache 2.0. That is a real shift from earlier Gemma terms, and it is the reason I would put this family on a course syllabus. The sizes span phones through workstations: E2B at 2.3B effective parameters (5.1B with embeddings), E4B at 4.5B effective / 8B, a 26B A4B MoE with 3.8B active / 25.2B total, and a 31B dense model.
On June 3 they added Gemma 4 12B Unified, an encoder-free multimodal model that Google says runs locally on 16GB of VRAM or unified memory, with native audio. The model card is worth reading in full. Google claims 150 million Gemma 4 downloads by the 12B launch. Those are vendor figures, so treat them as a signal of adoption rather than an audited count. The more useful signal is the product shape: on-device, multimodal, and wired for agent skills through AI Edge, Android AICore, and LiteRT-LM. Gemma 4 E2B can sit under 1.5GB on some devices with 2-bit or 4-bit weights and memory-mapped PLE. Google’s AI Edge numbers (133 prefill / 7.6 decode tok/s on a Raspberry Pi 5 CPU, much higher on a Qualcomm NPU) are the kind of measurements I want students to reproduce.
The rest of the permissive edge
Alibaba shipped Qwen3.5 on February 16, 2026, also Apache 2.0 and natively multimodal, with language coverage moving from 119 to 201. The dense SKUs that fit local work landed in March: 9B, 4B, 2B, and 0.8B. Microsoft’s Phi-4-mini-reasoning (April 2025, 3.8B, MIT, 128K context) is still the model I reach for when VRAM is tight and the task is math or short-form reasoning. It was distilled with synthetic traces from DeepSeek-R1, which is a useful reminder that small reasoners often inherit their habits from a much larger teacher.
Mistral Small 4, announced March 16, 2026, needs a careful read. It is Apache 2.0, and Hugging Face lists Mistral-Small-4-119B-2603 with a 256K window, but the name is doing a lot of work. This is a 119B MoE with roughly 6B active parameters. That is a workstation or multi-GPU box, not a phone. Sparse activation is a beautiful engineering trick, the same family of idea Apple used streaming routed experts from NAND in AFM 3. It does not make those models phone-class.
Llama 4 Scout and Maverick
Meta released Llama 4 Scout and Maverick on April 5, 2025. Scout is 17B active / 109B total, 16 experts, with a claimed 10 million token context and Int4 on a single H100. Maverick is 17B active / 400B total, 128 experts. They are natively multimodal MoE models distilled from still-unreleased Behemoth. They are open-weight. They are released under the Llama Community License, which still has EU domicile limits and requires Meta’s permission above 700 million monthly active users.
If you are teaching or shipping, read that license before you treat Llama as the default local stack. Apache 2.0 and MIT families are simply easier to put in a student’s repo. “Open-weight” and “OSI-open” are different sentences.
SmolLM3, because a full recipe is rare
Hugging Face’s SmolLM3 (July 8, 2025) is a 3B Apache-2.0 model trained on 11 trillion tokens, with dual /think and /no_think modes and 128K context (64K train plus YaRN). HF reports it beating Llama 3.2 3B and Qwen2.5 3B, and with thinking enabled they publish AIME 2025 at 36.7% versus 9.3% without, LiveCodeBench at 30.0% versus 15.2%, and GPQA Diamond at 41.7% versus 35.7%.
What still makes it essential is the artifact: architecture, data mixture, alignment, and merge are public. Most “open” releases hand you a checkpoint. This one hands you a path you can reproduce, which is exactly what a classroom or a research intern needs.
What the benchmarks still say
Phone-class models remain weak on the long tail. On Google’s own card, Gemma 4 E2B scores 24.5% on Tau2 against 76.9% for the 31B, and LiveCodeBench v6 is 44.0% versus 80.0%. Even the 31B is at 19.5% on HLE with no tools. An April 2026 bake-off also showed how prompt-fragile small reasoners can be: Phi-4-reasoning on GSM8K dropped from 0.67 under chain-of-thought to 0.11 under few-shot CoT.
So the architecture that is actually shipping is hybrid. 2026 systems papers such as SWARM-LLM keep a swarm of edge SLMs as the default path and call a 70B-class cloud model for roughly a quarter of queries. That is a design you can explain on a whiteboard: local first, escalate when the router says the task is hard, tool-heavy, or safety-sensitive. I would keep a log that shows students why a query left the device.
Time to dig deeper
If you want a product default, start with Gemma 4 E4B or 12B, or Qwen3.5 4B or 9B. If you want a teaching default, start with SmolLM3 and actually walk the recipe. If you want a math specialist in a tight memory budget, Phi-4-mini is still the one I would load first.
Three primary sources will take you further than this post can: the Gemma 4 launch and model card, the SmolLM3 writeup, and The Llama 4 herd. Download one of the 2 to 12B checkpoints. Run the same prompt with /think and without, on-device and in the cloud. Write down where it fails. That exercise will teach you more about 2026 Small AI than another parameter-count headline.