open models
Frontier-class work,
at home.
Heavy AI use gets expensive on subscriptions and metered APIs, and the price is always set by someone else. Open-weight models have changed what a Mac can do on its own. Depending on your chip and unified memory, a model you download once can get you near-frontier quality on many everyday tasks — on your desk, offline, for the cost of the electricity.
Photograph — Slava Auchynnikau via Unsplash
Where open models come from
Open-weight models are published by their makers on public model hubs. Orvlet's model library searches the two largest directly, shows each model's license before you download, and verifies every file by checksum:
- Hugging Face — the largest public model hub, including community MLX and GGUF conversions sized for Macs.
- ModelScope — a major hub where many model families publish their releases alongside Hugging Face.
Nothing is fetched until you search, and a download starts only when you press Download. Models you already have from LM Studio, Ollama or the Hugging Face cache are found on disk and reused — no second copy.
What “near-frontier” honestly means
The best open models of 2026 reason step by step, write and fix code, call tools, read images and hold very long documents in context. On many day-to-day tasks — drafting, summarizing, refactoring, answering questions about your own files — a well-chosen local model is close to the hosted frontier. On the hardest problems the largest hosted models can still be ahead, and the biggest open releases need far more memory than a laptop has.
Orvlet does not make you choose once. Keep a local model as your default, and send a single hard task to a cloud provider with your own key when it is worth it. Your projects stay on your Mac either way.
2026 in open models
What changed this year
Dated facts from the makers' own model cards and announcements. Benchmarks vary by task and setup, so we leave the scoreboards to the model cards.
Big models, small active part
Mixture-of-experts designs now dominate: a large total size with only a few billion parameters active per token. Qwen3.6-35B-A3B (April 2026) activates about 3B of 35B; Google's Gemma 4 26B A4B (April 2026) about 3.8B of 25B; Mistral Small 4 (March 2026) about 6B of 119B. Fewer active parameters per token means faster answers from the same memory.
Thinking you can dial
Nearly every 2026 release ships a reasoning mode you can turn on, off or up: Qwen3.5 and Qwen3.8 (thinking on by default; Qwen3.8 adds an effort setting), DeepSeek V4 (Non-think, Think High and Think Max), Mistral Small 4 and GLM-5.3-Flash (a reasoning-effort setting), IBM Granite 4.2 (think, non-think and low-effort modes).
Very long context
Qwen's 2026 models read 262,144 tokens natively and extend to about a million; Gemma 4's medium models hold 256K; DeepSeek V4, released April 24, 2026, offers a 1M context. Whole repositories and long reports fit in one request — memory permitting.
Low-bit weights from the source
Makers increasingly ship weights already quantized: OpenAI's gpt-oss models post-trained in MXFP4 (August 2025), Kimi K2.5 with native INT4 (January 2026), DeepSeek V4 with FP4 experts. Apple's MLX supports affine 1–8-bit, MXFP4, MXFP8 and NVFP4 quantization — the formats that make large models fit in unified memory.
Faster decoding built in
Multi-token prediction is now trained into many models — Qwen3-Next, Qwen3.5, Qwen3.6 and Qwen3.8 among them — and GLM-5 ships a configuration for speculative decoding with it. The Orvlet Engine uses prediction heads and draft models where a model provides them, and checks that the output matches plain decoding.
Licenses that permit real use
Qwen3.5, 3.6 and 3.8, Gemma 4, gpt-oss, Mistral Small 4 and Granite 4.2 are Apache 2.0; DeepSeek V4 and GLM-5 are MIT. Others use their makers' own licenses. Orvlet shows the license before every download and gates restrictive ones behind an explicit acceptance.
Macs built for it
Apple reported in November 2025 that the M5's GPU Neural Accelerators speed up prompt processing by up to 4x over M4. The Mac Studio announced August 25, 2026 offers M5 Max with up to 128 GB of unified memory and M5 Ultra with up to 512 GB at 1.2 TB/s, which Apple says lets users run massive models entirely on device.
Download it once. Run it offline. Nobody meters the next question.
Photograph — Donna Elliot via Unsplash
a practical guide
What can my Mac run?
Unified memory decides the size of model you can hold; the chip decides how fast it answers. Orvlet's library marks a model Fits when weights, KV cache for your chosen context and runtime overhead stay within 60% of memory, and Tight up to 80%. The rows below are rules of thumb for 4-bit weights (roughly 0.6 GB per billion parameters); the library computes the real figure from each model's files.
| Unified memory | Comfortable at 4-bit | Good for |
|---|---|---|
| 16 GB | Dense models up to about 8B; small MoE models with a few billion active parameters as a tight fit. | Chat, writing, summaries, quick code edits, embeddings and voice transcription. |
| 24 GB | Dense models in the 8B–14B range with room for a long context. | Everyday coding help, document Q&A over your knowledge collections, vision on screenshots. |
| 32 GB | The 27B–35B class: dense 27B models and 30B-class MoE models at 4-bit, usually as a tight fit. | Multi-step agent tasks, reasoning modes, longer refactors. |
| 64 GB | 27B–35B models comfortably or at 8-bit; 70B–80B-class models as a tight fit. | A strong daily driver for coding and research, with subagents on a second small model. |
| 96–128 GB | MoE models in the 100B–120B class at 4-bit, such as the gpt-oss-120b and Mistral Small 4 size class. | Near-frontier work on many tasks, long-context analysis, model comparisons side by side. |
| 256–512 GB | MoE models in the 200B–400B class at 4-bit. | The largest models a single Mac can hold. Bigger releases (trillion-parameter models) stay beyond one machine. |
Estimates only. Context length, a quantized KV cache (8- or 4-bit, a load option in the Engine), other running apps and the exact file sizes all move the line. Prompt processing is faster on newer chips; generation speed follows memory bandwidth.
Getting more from the memory you have
- Pick the right quant. The library lists each GGUF variant with its size, and a comparison matrix shows memory, speed and quality trade-offs between quantizations.
- Shrink the cache. An 8- or 4-bit KV cache roughly halves or quarters the memory a long context needs.
- Quantize at load. The Engine can quantize full-precision weights to 4- or 8-bit as it loads them, and convert GGUF files to MLX.
- Use two Macs. A GGUF model's layers can be split across Macs you pair on your own network, in proportion to each one's free memory.
Bring a runtime you already like
The Orvlet Engine serves MLX and GGUF models whose architecture it supports, and the library tells you before you load. Anything else can run through Ollama, LM Studio, llama.cpp, MLX LM or any OpenAI-compatible server, and Orvlet connects to it without a key. On macOS 27 with Apple Intelligence, Apple's on-device Foundation Models are one more option with no download at all.
Sources
Model facts are from the makers' published model cards and announcements: Qwen (qwenlm.github.io, huggingface.co/Qwen and the QwenLM release log), DeepSeek (api-docs.deepseek.com news, April 24, 2026), Google's Gemma 4 model card (ai.google.dev), Mistral AI (mistral.ai/news, March 16, 2026), Moonshot AI, Z.ai, OpenAI (github.com/openai/gpt-oss), IBM Granite, Apple Machine Learning Research (November 19, 2025), the Apple Newsroom (August 25, 2026) and the MLX documentation. Model names belong to their owners; Orvlet is not affiliated with any model maker.