The Picks at a Glance
Six models make the 8GB cut in August 2026. Four arrived in the July-August on-device wave; two are proven incumbents. Every speed figure below carries its source.
| Model | Params | Footprint | Speed | License |
|---|---|---|---|---|
| LFM2.5-2.6B | 2.6B dense | <2.5 GB RAM | 220 tok/s, M5 Max (vendor) | LFM Open |
| Maple Preview 20B-A1B | 20B-A1B ternary | Fits 8GB | 281 tok/s, M5 Pro (vendor) | MIT |
| Bonsai 27B | 27B (1-bit) | 3.9-5.9 GB | Community figures pending | Apache 2.0 |
| Nanbeige4.2-3B | 3B looped | Fits 8GB easily | Community figures pending | Apache 2.0 |
| Phi-4 Mini | 3.8B dense | ~2.5 GB at Q4 | Fast (est., see model page) | MIT |
| Gemma 4 E2B | 2B effective | ~2 GB class | ~155 tok/s (est.) | Gemma |
The short version: LFM2.5-2.6B is the speed-and-footprint pick, Bonsai 27B is the quality ceiling, and the incumbents remain the predictable choices. The interesting story is how the ternary models got here at all.
The New On-Device Wave
LFM2.5-2.6B — the efficiency pick
Liquid AI's LFM2.5-2.6B arrived in August 2026 under the LFM Open license, and it is the closest thing the 8GB tier has to a default. It runs in under 2.5 GB of RAM, and Liquid vendor-reports 220 tok/s on an M5 Max. Even discounting vendor optimism, a 2.6B model with that profile leaves an 8GB Mac with more free memory than any other entry on this page — room for a real context window, a browser, and your editor at once.
Maple Preview 20B-A1B — natively ternary
DeepGrove's Maple Preview 20B-A1B (August 2026, MIT) is the more radical design: a natively-trained ternary MoE — weights constrained to three values during training, not quantized afterward — with roughly 1B active parameters. DeepGrove vendor-reports 281 tok/s on an M5 Pro and 200+ tok/s on a Mac mini M4. Those are drafter-free numbers on mainstream hardware, and if community benchmarks bear them out, Maple is the fastest genuinely capable model the entry tier has seen.
Bonsai 27B — a 27B on an 8GB Mac
Prism ML's Bonsai 27B (July 2026, Apache 2.0) is the headline act: a native 1-bit/ternary derivative of Qwen3.6-27B — the model that currently tops the LLMCheck Mac rankings — compressed to a 3.9-5.9 GB checkpoint. Prism ML vendor-reports roughly 90% quality retention against the full-precision original. If that holds up under community testing, it means 27B-class reasoning on a base MacBook Air, which would have sounded absurd a year ago.
Nanbeige4.2-3B — the vendor-benchmark outlier
Nanbeige4.2-3B (July 2026, Apache 2.0) uses a looped-transformer architecture that re-runs layers to trade compute for depth, and its vendor-reported 63.6 on SWE-bench Verified would be extraordinary for a 3B model — better than some 30B-class systems. That number is vendor-reported and not yet independently reproduced, so file it under "intriguing, verify before relying on it." As a lightweight coding assistant it is worth a try regardless.
According to the LLMCheck index, the common thread in this wave is provenance: nearly every headline figure is vendor-reported so far. Community-submitted Apple Silicon numbers are accumulating — check the benchmarks page for the current state of verification.
The Incumbents: Phi-4 Mini and Gemma 4 E2B/E4B
Phi-4 Mini is Microsoft's real small model — a 3.8B dense network under the MIT license, roughly 2.5 GB at Q4. It carries the phi family's signature strength: reasoning and structured tasks well above its weight class, at the cost of thinner long-tail world knowledge. It is the model the fabricated "Phi-5 Mini" entry was presumably an echo of, and it remains an excellent, boring, reliable choice.
Gemma 4 E2B and its larger sibling E4B are Google's efficiency pair. E2B is the fastest model in the LLMCheck index at roughly 155 tok/s (estimated) on high-end Apple Silicon, and both have the broadest ecosystem support in the tier — every runtime, every quant format, extensively community-tested. E4B is the pick when you want a step up in quality while staying comfortably inside 8GB.
The honest comparison with the new wave: the incumbents are slower on paper and smaller in headline capability, but their numbers are verified and their failure modes are known. The ternary models are more capable on paper and mostly vendor-benchmarked. Running one of each — an incumbent as your daily driver, a ternary as your experiment — is a perfectly sensible 8GB strategy right now.
Install Guide
The incumbents are one command away. Install Ollama, then:
ollama run phi4-mini
Gemma 4 E2B/E4B work the same way. Note that Ollama 0.32.x now opens an interactive coding agent when you type ollama with no arguments, and its MLX engine gained image input and M5 Neural Accelerator support — small models benefit disproportionately from both.
For the new wave: LFM2.5, Maple Preview, and Bonsai publish MLX checkpoints, and LM Studio (0.4.20+) lists all of them in its model browser — search the model name, download, and load. Nanbeige4.2-3B ships standard GGUF builds that run anywhere llama.cpp does. See our guides hub for the MLX setup walkthrough.
Limitations and Caveats
- Vendor numbers dominate this wave. The 220 tok/s (LFM2.5), 281 tok/s (Maple), 63.6 SWE-bench Verified (Nanbeige), and 90% quality-retention (Bonsai) figures are all vendor-reported. Treat them as claims until community data lands.
- Ternary trade-offs are real. Even at 90% retention, Bonsai loses something versus full-precision Qwen3.6-27B — typically on edge-case reasoning and rare knowledge. The floor is high, but it is not the original model.
- Small models hallucinate. Every entry here is strongest at reasoning, code, and structured tasks, and weakest as an encyclopedia. Verify factual claims.
- 8GB still constrains context. Weights fitting is not the whole story — the KV cache grows with context length, so long-document work remains the domain of 16GB+ machines.
The summary: the 8GB Mac has gone from "barely viable" to genuinely spoiled for choice in one summer. Start with LFM2.5-2.6B or Phi-4 Mini, keep Bonsai 27B loaded for the hard questions, and watch the benchmarks page as community verification fills in.