“Open-weight” means the trained model files are published and you can run them on your own hardware. It does not always mean open source: some licences add conditions, so we list each one. Sizes, licences and context lengths come from each model's Hugging Face page; benchmark scores are the ones each maker publishes, checked on 20 October 2026.
The short version
| Model | Maker | Sizes | Licence | Best for |
|---|---|---|---|---|
| Gemma 4 | Google DeepMind | E2B, E4B, 12B, 26B A4B (MoE), 31B | Apache 2.0 | Phones and laptops; images and audio; 140+ languages |
| Qwen3.5 | Alibaba (Qwen) | 0.8B to 397B A17B (MoE) | Apache 2.0 | The widest range of sizes; strong small models; Chinese |
| DeepSeek V4 | DeepSeek | Flash 284B (13B active), Pro 1.6T (49B active) | MIT | Top open reasoning and coding; 1M-token context |
| Kimi K3 | Moonshot AI | 2.8T (104B active) | Kimi K3 License | Long agentic coding sessions; text and images |
| GLM-5.3 | Z.ai | GLM-5.3, GLM-5.3-Flash | Custom (5.3); MIT (5.3-Flash) | English and Chinese; agents |
| Mistral Small 4 | Mistral AI | 119B; Devstral Small 2 24B for code | Apache 2.0 | European languages; a permissive licence |
| gpt-oss | OpenAI | 20B, 120B | Apache 2.0 | A 20B model that fits a 16 GB machine |
| Llama 4 | Meta | Scout 109B (17B active), Maverick | Llama 4 Community License, gated | Existing Llama setups; not updated since May 2025 |
| LFM2.5 | Liquid AI | 230M, 1.2B, 2.6B, 8B A1B | LFM Open License v1.0 | Fast, low-memory models for phones and edge devices |
MoE (mixture of experts) models list total and active parameters. All of them must sit in memory, but only the active ones work on each token, so they run faster than their total size suggests. “Gated” means you must accept the licence on Hugging Face before downloading.
Benchmarks, as each maker reports them
Two common tests: MMLU-Pro (broad knowledge) and GPQA Diamond (hard science questions). Makers run them with their own settings, so treat small gaps as noise.
| Model | MMLU-Pro (%) | GPQA Diamond (%) | Source |
|---|---|---|---|
| DeepSeek V4 Pro (max reasoning) | 87.5 | 90.1 | DeepSeek model card |
| Gemma 4 31B | 85.2 | 84.3 | Google model card |
| Gemma 4 26B A4B | 82.6 | 82.3 | Google model card |
| Qwen3.5 9B | 82.5 | 81.7 | Qwen model card |
| gpt-oss 120B | 80.8 | 80.1 | Qwen3.5-9B model card |
| Qwen3.5 4B | 79.1 | 76.2 | Qwen model card |
| gpt-oss 20B | 74.8 | 71.5 | Qwen3.5-9B model card |
| Gemma 4 E4B | 69.4 | 58.6 | Google model card |
| Gemma 4 E2B | 60.0 | 43.4 | Google model card |
What stands out: in 2026 a 9B open model (Qwen3.5) reports MMLU-Pro and GPQA scores above a 120B model from 2025 (gpt-oss), and the largest open models (DeepSeek V4 Pro) are close to closed frontier models on the maker's own comparison.
Model by model
Gemma 4 (Google DeepMind)
Five sizes: E2B (2.3B effective, 5.1B with embeddings), E4B, 12B, a 26B mixture-of-experts model with 3.8B active, and 31B. All read images; E2B, E4B and 12B also take audio. Context is 128K tokens on the small ones and 256K on the rest, and Google says they were pretrained on 140+ languages. Apache 2.0. The 26B A4B is the interesting one for laptops: near-31B scores, running close to 4B speed. Best pick if you want one family from phone to workstation.
Qwen3.5 (Alibaba)
The widest line-up: 0.8B, 2B, 4B, 9B, 27B, 35B A3B, 122B A10B and 397B A17B, all Apache 2.0 and all able to read images. 262K tokens of context natively (9B) and 201 languages and dialects according to Qwen. The 4B and 9B are the strongest small models on paper. In our own five-question test (below) the smallest ones were less reliable on plain facts.
DeepSeek V4 (DeepSeek)
V4-Flash (284B total, 13B active) and V4-Pro (1.6T total, 49B active), both with a one-million-token context and an MIT licence. DeepSeek reports the Pro version in maximum reasoning mode at 93.5 on LiveCodeBench and a Codeforces rating of 3206. A V4.1-Flash followed on 16 October 2026. These need data-centre hardware; most people use them through a hosted API. The older “R1 Distill” small models are not the same thing and did poorly in our test.
Kimi K3 (Moonshot AI)
2.8 trillion parameters with 104B active, text and images, one-million-token context, released under its own Kimi K3 License. Moonshot pitches it for long coding and knowledge-work sessions with little supervision. Data-centre only.
GLM-5.3 (Z.ai)
GLM-5.3 (custom licence) and GLM-5.3-Flash (MIT, reads images), both tagged for English and Chinese on Hugging Face, updated on 14 and 15 October 2026. Large models for servers and agent work.
Mistral Small 4 and Devstral (Mistral AI)
Mistral Small 4 is a 119B model under Apache 2.0 whose card lists 24 languages, including Russian, Chinese, Arabic and Hindi. Devstral Small 2 (24B, Apache 2.0) is Mistral's coding model and fits a well-equipped desktop in 4-bit.
gpt-oss (OpenAI)
Two models from August 2025, 20B and 120B, both Apache 2.0. Not the newest, but the 20B is still a common choice for a 16 GB machine, and both are widely supported by local tools.
Llama 4 (Meta)
Scout (109B total, 17B active) and Maverick, under the Llama 4 Community License, gated on Hugging Face. The weights there were last updated in May 2025. Smaller Llama 3.2 models (1B, 3B) are still common in phone apps.
LFM2.5 (Liquid AI)
Small models built for devices: 230M, 1.2B, 2.6B and an 8B with 1B active. The 1.2B runs in under 1 GB of memory according to Liquid AI, with a 32,768-token context and eight languages. Liquid itself recommends it for agents, data extraction and retrieval, and not for knowledge-heavy tasks or programming. Its licence is Liquid's own (LFM Open License v1.0), so read it before commercial use.
What hardware you need
A rough rule: a model quantised to 4 bits needs about 0.5 to 0.6 GB of memory per billion parameters, plus room for the conversation. For mixture-of-experts models, count the total parameters, not the active ones.
| Device | Realistic models (4-bit) |
|---|---|
| iPhone or Android phone | Up to about 4B: Gemma 4 E2B (about 3.6 GB in MLX 4-bit), Qwen3.5 2B or 4B, LFM2.5 1.2B (about 0.7 GB), Llama 3.2 1B or 3B |
| Laptop with 16 GB | Up to about 14B dense: Qwen3.5 9B, Gemma 4 12B, gpt-oss 20B (tight) |
| Mac or PC with 32 to 64 GB | Gemma 4 26B A4B or 31B, Qwen3.5 27B or 35B A3B, Devstral Small 2 24B |
| Workstation with 128 GB or more | Mistral Small 4 119B, gpt-oss 120B, Qwen3.5 122B A10B |
| Data centre | DeepSeek V4, Kimi K3, GLM-5.3, Qwen3.5 397B, Llama 4 Maverick |
Jan's documentation gives a similar guide for Macs: 8 GB for models up to about 3B, 16 GB up to about 7B, 32 GB up to about 13B. LM Studio recommends 16 GB or more on Apple silicon.
A small test: five questions on a Mac
Pinku's developers ran the small models they were considering through the same five factual questions (capital of France, days in a leap year, author of Don Quixote, chemical symbol for gold, largest planet), three in Spanish and two in English, with MLX 4-bit builds on a Mac. A model passed a question if it gave the right answer without rambling past the length limit. It is five questions, not a benchmark: it only shows whether a phone-sized model gets basic facts right.
| Model | Correct |
|---|---|
| Gemma 4 E2B | 5 of 5 |
| LFM2.5 1.2B | 5 of 5 |
| Qwen3.5 2B | 4 of 5 |
| Qwen3.5 4B | 4 of 5 |
| DeepSeek R1 Distill 7B | 4 of 5 |
| Qwen3.5 0.8B | 2 of 5 |
| DeepSeek R1 Distill 1.5B | 0 of 5 |
How to use them
On a computer
- Pinku for Mac: in the Mac App Store (version 1.0.7 on sale), Apple silicon with macOS 14 or later. Twelve curated models from 0.34 to 4.7 GB: Gemma 3 1B and Gemma 3n E2B, Llama 3.2 1B and 3B, Qwen3 0.6B, 1.7B, 4B and 8B, DeepSeek-R1 Distill Qwen 1.5B (4-bit and 8-bit), SmolLM3 3B and LFM2 1.2B. No terminal or setup: pick a model and chat. Free to download; larger models with Pinku Pro. It does not yet include Gemma 4 or Qwen3.5.
- Ollama: free and open source (MIT), macOS 14 or later, Windows and Linux. A command line plus a desktop chat app since July 2025. Its library has Gemma 4, Qwen3.5, gpt-oss, Llama 4 and LFM2.5 to run locally; the giant models (DeepSeek V4, Kimi K3, GLM-5.3) run on Ollama's cloud, paid with usage credits.
- LM Studio: free desktop app (not open source) with a model browser, running models with llama.cpp and Apple's MLX. On a Mac it needs Apple silicon and macOS 14; it recommends 16 GB of memory.
- llama.cpp: the open-source (MIT) engine behind many of these apps, for people who are happy in a terminal. It runs GGUF files, the format most models are shared in.
- Jan: open-source desktop chat app, version 0.8.6 (9 October 2026), macOS 13.6 or later; local models need Apple silicon on a Mac. It can also connect to cloud models.
- mlx-lm: Apple's MIT-licensed library for running models on Apple silicon, for developers.
On a phone
- Pinku (iPhone 12 or later with iOS 18.6, and Macs with Apple silicon): a curated list of small models (Gemma, Llama 3.2, Qwen3, SmolLM3, LFM and others) that it only shows if your device has the memory for them; no account; the default model is free and larger ones come with Pinku Pro. Chats never leave your device; Apple's privacy label lists only purchase history and an anonymous user ID, used for Pinku Pro purchases. Where Hugging Face is blocked it downloads from ModelScope.
- Locally AI: free, from the LM Studio team, iPhone and iPad (iOS 18.1) and Mac (macOS 26).
- PocketPal AI: free and open source, on iPhone and Android; loads any GGUF model from Hugging Face, so you pick the size yourself.
- Private LLM: $4.99 once, iPhone, iPad and Mac, with Siri and Shortcuts.
- Enclave: free, iOS 27, with optional cloud models in its Pro tier.
Compared in more detail elsewhere: the phone apps by the models they can load on Reviewland, and the desktop tools as Ollama alternatives on Reditor (both run by Saga Labs, in English).
More on phones: AI apps for iPhones without Apple Intelligence and what Apple Intelligence does offline.
Questions people ask
What is the best open-weight AI model in 2026?
It depends on the hardware. For data-centre use, DeepSeek V4 Pro reports the highest scores of the models here on its own card. For a laptop, Gemma 4 26B A4B or Qwen3.5 9B. For a phone, Gemma 4 E2B, Qwen3.5 2B to 4B or LFM2.5 1.2B.
What is the difference between open-weight and open source?
Open-weight means the trained model files are published so you can run them. The licence may still add conditions: Gemma 4, Qwen3.5, Mistral Small 4 and gpt-oss use Apache 2.0, DeepSeek V4 uses MIT, while Llama 4, Kimi K3, GLM-5.3 and LFM2.5 have their own licences.
How much memory do I need to run a local AI model?
Roughly 0.5 to 0.6 GB per billion parameters at 4-bit, plus room for the conversation. A 4B model fits a recent phone, about 14B fits a 16 GB laptop, and the trillion-parameter models need servers.
Which open model works best on an iPhone?
Models up to about 4B parameters. In Pinku's five-question test on a Mac, Gemma 4 E2B and LFM2.5 1.2B answered all five correctly, Qwen3.5 2B and 4B four, and DeepSeek R1 Distill 1.5B none.
How do I run an open model on my computer?
Install Ollama, LM Studio or Jan, choose a model that fits your memory and download it. llama.cpp and mlx-lm are the options for people comfortable with a terminal.