How to read this list
Three numbers matter. Total parameters decide memory: at 4-bit quantization, roughly 0.5 to 0.6 GB per billion, plus context. Active parameters decide speed on local hardware, because each generated token only reads the experts it uses. And the licence decides whether legal will let you use it. I have not included benchmark scores: vendors run them with their own harnesses, and the only evaluation that matters is yours, on your tasks. The tier column assumes 4-bit weights and a moderate context; long contexts need more memory.
The list
| Model | Maker | Parameters | Context | Licence | Memory tier |
|---|---|---|---|---|---|
| gpt-oss-20b | OpenAI | 21B total / 3.6B active (MoE) | 128K | Apache 2.0 | 16–32 GB |
| Gemma 4 26B A4B | Google DeepMind | 25.2B total / 3.8B active (MoE) | 256K | Apache 2.0 | 16–32 GB |
| Gemma 4 31B / 12B / E4B | Google DeepMind | 30.7B / 12B / 4.5B effective (dense) | 256K (31B, 12B); 128K (E4B) | Apache 2.0 | 8–32 GB |
| Qwen3.8-27B | Alibaba Qwen | 27B (dense) | 262K native | Apache 2.0 | 32 GB |
| Muse Glimmer 30B | Meta | 29.6B incl. vision encoder (dense) | 131K+ | Apache 2.0 | 32 GB |
| gpt-oss-120b | OpenAI | 117B total / 5.1B active (MoE) | 128K | Apache 2.0 | 80–128 GB |
| Mistral Small 4 | Mistral AI | 119B total / 6.5B active (MoE) | 256K | Apache 2.0 | 96–128 GB |
| Qwen3.8-Flash-Next | Alibaba Qwen | 125B / 6B active, plus 51B n-gram embedding (MoE) | — | Qwen Community 1.0 | 128 GB+ |
| GLM-5.3-Flash | Z.ai | 320B total / 18B active (MoE) | — | MIT | 256 GB |
| MiniMax-M3 | MiniMax | ~428B total / ~23B active (MoE) | 1M | MiniMax Community | 256–512 GB |
| DeepSeek V4.1 Flash | DeepSeek | 552B backbone / 8–16B active (MoE) | 1M | MIT | 512 GB |
Parameters, context and licences are from each model's Hugging Face card, checked on 20 October 2026. A dash means I did not confirm the figure from the card. "Memory tier" is my estimate, not the vendor's.
Tier 1: 16 to 32 GB (laptops, entry desktops)
gpt-oss-20b is the default. OpenAI's model card says it runs within 16 GB of memory, it has only 3.6 billion active parameters so it is quick, and it is Apache 2.0.
Gemma 4 from Google DeepMind comes in five sizes, all Apache 2.0: E2B and E4B for phones and small laptops, 12B, a 26B A4B Mixture-of-Experts with 3.8 billion active parameters, and a dense 31B. The 26B A4B is the interesting one locally: near-30B capacity at the speed of a 4B model, with a 256K context. The family is multimodal, taking text and images, with audio on the smaller models.
Qwen3.8-27B is the dense option: 27 billion parameters, Apache 2.0, a native context of 262,144 tokens, and a model card aimed squarely at coding agents. Dense means slower per token than the MoE models at the same size, and often more consistent. Anyone searching for "Qwen 3" should know the family is now at 3.8.
Muse Glimmer 30B is Meta's August 2026 release, published by Meta's verified Hugging Face organisation under Apache 2.0, a change from the custom Llama licences. The card describes it as distilled from Muse Spark for agentic tasks on consumer hardware, with image input.
Tier 2: 64 to 128 GB (Mac Studio, DGX Spark, Strix Halo, one 96 GB GPU)
gpt-oss-120b is the standard choice: 117 billion parameters, 5.1 billion active, Apache 2.0, and per OpenAI it fits a single 80 GB GPU. On a 128 GB unified-memory machine it leaves room for context.
Mistral Small 4 is the main alternative: 119 billion parameters with 6.5 billion active, a 256K context and Apache 2.0, from a European vendor, which some procurement teams care about.
Qwen3.8-Flash-Next is newer and architecturally ambitious (125 billion parameters with 6 billion active, plus a 51-billion-parameter n-gram embedding table), but it ships under the Qwen Community licence rather than Apache 2.0. Read that licence before deploying it commercially. At 4-bit it is a tight fit for 128 GB.
Tier 3: 256 to 512 GB (M5 Ultra Mac Studio, multi-GPU workstations)
This is where open models start to look like the frontier. GLM-5.3-Flash from Z.ai has 320 billion parameters with 18 billion active, is MIT-licensed, and its card positions it as the first natively multimodal model in the GLM-5 series. MiniMax-M3 has about 428 billion parameters with about 23 billion active and a 1M-token context, under the MiniMax Community licence. DeepSeek V4.1 Flash has a 552-billion-parameter backbone with 8 to 16 billion activated, a 1M context and an MIT licence; at 4-bit it needs most of a 512 GB machine.
Speed at this tier is limited by bandwidth: 18 to 23 billion active parameters read per token is several times the work of the tier-2 models. A 1.2 TB/s M5 Ultra handles that far better than a 273 GB/s box would, which is why the hardware page puts the Mac Studio at the top for this tier.
Open, but not local
Some of the strongest open-weight models are simply too large for a desk. Kimi K3 is 2.8 trillion parameters under its own licence. DeepSeek V4 Pro is about 1.6 trillion (MIT). Qwen3.8-2.4T-A95B and the full GLM-5.3 (about 753 billion) are data-center models too. They matter because you can run them on your own rented GPUs or through hosts such as Ollama's cloud tier, keeping the open-weight benefits (no vendor lock-in, pinned versions) without owning the hardware.
How I would pick
- Shortlist two models from your memory tier, at least one Apache 2.0 or MIT.
- Run both on twenty real tasks from your team's work, with the runtime you will actually use. The options are compared in Ollama vs LM Studio vs llama.cpp vs vLLM.
- Measure quality, tokens per second and memory headroom with your real context lengths.
- Pin the version, and re-run the same tasks when a new release comes out. Something will within weeks.
Related
- Local AI for engineering teams — when running models yourself makes sense
- Best hardware for local LLMs — what each memory tier costs
- Ollama vs LM Studio vs llama.cpp vs vLLM — the runtime choice
- Neoclouds compared — renting GPUs for the models that do not fit
Frequently asked questions
What is the best local LLM in 2026?
There is no single best one; it depends on your memory. With 16 to 32 GB, gpt-oss-20b, Gemma 4 (26B A4B or 31B) and Qwen 3.8 27B are the strongest candidates. With 128 GB, gpt-oss-120b and Mistral Small 4 (119B) are the safest choices, both Apache 2.0. With 256 to 512 GB, GLM-5.3-Flash (320B, MIT) and DeepSeek V4.1 Flash (MIT) become possible.
What is gpt-oss?
gpt-oss is OpenAI's family of open-weight reasoning models, released in August 2025 under the Apache 2.0 licence. gpt-oss-120b has 117 billion parameters with 5.1 billion active per token and runs on a single 80 GB GPU; gpt-oss-20b has 21 billion parameters with 3.6 billion active and runs within 16 GB of memory, thanks to MXFP4 quantization of the expert weights.
What is the latest Qwen 3 model?
As of October 2026 the current Qwen release on Hugging Face is Qwen 3.8, published in August 2026. It includes Qwen3.8-27B, a dense 27-billion-parameter model under Apache 2.0 with a 262,144-token native context, and Qwen3.8-Flash-Next, a Mixture-of-Experts model with 125 billion parameters (6 billion active) plus n-gram embeddings, under the Qwen Community licence. The largest, Qwen3.8-2.4T-A95B, is far too big for local hardware.
Can I run DeepSeek locally?
Only the smaller variants, and only with a lot of memory. DeepSeek V4.1 Flash has a 552-billion-parameter backbone and is MIT-licensed; at 4-bit quantization it needs several hundred gigabytes, so a 512 GB Mac Studio or a multi-GPU server. DeepSeek V4 Pro is about 1.6 trillion parameters and is a data-center model.
Which local LLM is best for coding?
For a laptop, Qwen 3.8 27B and gpt-oss-20b are the common choices; Meta's Muse Glimmer 30B is built specifically for agentic tool use. With 128 GB, gpt-oss-120b and Mistral Small 4 work well with coding agents. Test on your own repository with your own agent harness before standardising; published coding benchmarks are run with specific harnesses that may not match yours.
Sources: official model cards on Hugging Face from OpenAI, Google, Alibaba Qwen, Meta, Mistral AI, Z.ai, MiniMax, DeepSeek and Moonshot AI (retrieved 20 October 2026). Licence names are as declared on the cards; read the full licence text before commercial use.