Local AI by Editorial Team

Best Local LLM for Everyday Tasks (Chat, Writing, Summarizing)

Which open-weight LLM should you actually run for everyday chat, writing, and summarizing? A tier-by-tier guide from 8 GB laptops to 32 GB desktops.

“Best local LLM” used to mean whichever 7B you could squeeze into 8 GB. The 2026 field is wider: a 1B model on a thin laptop, a 4B on an 8 GB card, a 24B on a 24 GB card, and a 70B on a workstation. Every figure below was pulled from the Ollama library page and the Hugging Face model card on 2026-09-27, so the sizes and licences are today’s, not last month’s.

TL;DR

  • 8 GB laptops and Steam Decks: llama3.2:3b (2.0 GB) for the lightest path, qwen3:4b (2.5 GB) if you can spare a little more RAM and want longer context.
  • 16 GB tier: qwen3:8b (5.2 GB, 40K context) or mistral-small:24b (14 GB, 32K context) if you have headroom. The 24B is the first tier where chat, writing, and summarizing feel coherent on long inputs.
  • 24 GB tier: mistral-small:24b is the practical sweet spot. qwen3:14b (9.3 GB) is the lower-cost alternative with 100+ language support.
  • 32 GB+ tier: qwen3:32b (20 GB) is the highest-quality fully open (Apache-2.0) model most people can run at home. llama3.3:70b (43 GB) is the frontier if you accept Meta’s licence terms.
  • License picks matter as much as size. Apache-2.0 (Qwen3, Mistral Small 24B, Phi-4) is the safe pick for commercial or redistributable work. Llama 3.2/3.3 and Gemma 3 have custom licences that require acceptance and have restrictions on use.

What “everyday tasks” actually covers

Everyday is the opposite of agentic coding and large-context research. It is a chat that does not lose the thread of the conversation, a paragraph rewrite that does not sound like a Wikipedia stub, and a long document or article that can be summarized without the model losing the first page by the time it reaches the last. Three properties drive almost every recommendation below:

  1. Instruction following - whether the model reads your prompt and does what you asked rather than guessing.
  2. Prose quality - whether the output reads naturally instead of stilted or repetitious.
  3. Long-context retention - whether the model can hold a 10-page article or a 30-message thread without losing the start.

Those three collapse into a simple hardware ladder, because every size jump buys more of all three at once.

The 2026 open-weight field for chat and writing

Five model families cover the realistic choices for an everyday local assistant. Three are Apache-2.0 (Qwen3, Mistral Small, Phi-4), one is Meta’s Llama 3.2/3.3 Community License, and one is Google’s Gemma license. The licence decision is made before the size decision: if your work is commercial or redistributable, narrow to Apache-2.0 and MIT first.

Qwen3 (Apache-2.0, Alibaba)

The Qwen3 family ships at 0.6B, 1.7B, 4B, 8B, 14B, 30B (MoE), 32B, and 235B (MoE), with 40K-token context for the dense sizes and 256K context for the 4B, 30B, and 235B variants. The 4B and 14B cards on Hugging Face list 32,768 tokens natively, extendable to 131,072 with YaRN. The Qwen3 README advertises 100+ languages and dialects, switching between a “thinking” mode for reasoning and a “non-thinking” mode for chat, and an Apache-2.0 licence across the family.

For everyday tasks the relevant Qwen3 sizes are qwen3:1.7b (1.4 GB, 2.03B parameters, Q4_K_M), qwen3:4b (2.5 GB, 38.1M downloads), qwen3:8b (5.2 GB), and qwen3:14b (9.3 GB). The 30B and 235B are MoE checkpoints; they need a 24 GB card minimum for the 30B and a multi-GPU workstation for the 235B, which is out of scope for everyday hardware.

Gemma 3 (Gemma Terms of Use, Google)

The Gemma 3 family ships at 270M, 1B, 4B, 12B, and 27B, with 128K context for the 4B, 12B, and 27B sizes and 32K for the 1B. The 270M and 1B are text-only; the rest are multimodal (text and image input). Gemma 3 is gated: the Ollama tag page points at the “Gemma Terms of Use” (last modified February 21, 2024), which requires users to agree to Google’s usage licence on Hugging Face before access.

The relevant sizes for everyday local use are gemma3:1b (815 MB, 32K context, text-only), gemma3:4b (3.3 GB, 128K context, multimodal), and gemma3:12b (8.1 GB, 128K context, multimodal). The 4B and 12B are the sizes that actually compete with Qwen3 at the same VRAM tier; the 27B (17 GB) is the 24 GB+ tier.

Mistral Small (Apache-2.0, Mistral AI)

mistral-small:24b is the headline tag: 14 GB, 23.6B parameters, Q4_K_M, 32K context, Apache-2.0. The Hugging Face Mistral-Small-24B-Instruct-2501 card confirms the Apache-2.0 licence and 32K context. A Tekken tokenizer with 131k vocabulary handles English, French, German, Spanish, Italian, Chinese, Japanese, Korean, Portuguese, Dutch, and Polish.

Mistral’s plain mistral:7b is still a viable 4.4 GB tag at 32K context for the 16 GB tier; the older Mistral-7B-Instruct-v0.3 card lists Apache-2.0. The 7B is a workhorse more than a frontier model: it is the one to pick when a 24B will not fit and you want a model that is not Qwen3 or Llama.

Llama 3.2 and 3.3 (Llama Community License, Meta)

Meta’s open-weight chat line is split across two licences: the Llama 3.2 Community License covers the 1B and 3B tags, and the Llama 3.3 Community License covers the 70B. Both are custom commercial licences that require acceptance; companies above 700 million MAUs must request a separate licence from Meta. Both 3.2 3B and 3.3 70B list 128K context.

The realistic Ollama tags for everyday local use are llama3.2:3b (2.0 GB, 128K context, 8 languages) for the low end and llama3.3:70b (43 GB) for the high end. The 70B is not the best fit for everyday tasks: it is a frontier model on a workstation, not a casual chat model on a 24 GB card.

Phi-4 (MIT, Microsoft Research)

microsoft/phi-4 is a 14B dense decoder-only Transformer, MIT-licensed, with 16K context, released December 12, 2024. The card lists strong reasoning numbers (MMLU 84.8, GPQA 56.1, MATH 80.4, HumanEval 82.6) but flags a relative weakness on factual recall (SimpleQA 3.0) and a primarily English training mix (~8% multilingual). For pure chat and writing in English, Phi-4 competes with qwen3:14b and mistral-small:24b at the same VRAM tier. For multilingual or summarization-heavy workloads, Qwen3 or Mistral Small is the safer pick.

Pick by hardware tier

Under 4 GB of usable VRAM (8 GB laptops, Steam Deck, integrated graphics)

The realistic options are llama3.2:1b (1.3 GB), qwen3:1.7b (1.4 GB, 2.03B params, Apache-2.0), gemma3:1b (815 MB, 32K context, text-only), and llama3.2:3b (2.0 GB). Pick by licence first:

  • Apache-2.0 only: qwen3:1.7b. It holds 40K context, supports 100+ languages, and is the strongest general-purpose 1-2B open-weight chat model on the Ollama library today.
  • Meta licence acceptable: llama3.2:3b (2.0 GB, 128K context, 8 languages) is the better chat quality at this tier.
  • Gemma licence acceptable: gemma3:1b (815 MB) is the fastest to load and responds at the highest tokens-per-second on this hardware, but the 32K context and 8-language support make it a smaller model than the 3B Llama.

8 GB usable VRAM (entry-level discrete GPUs, M1/M2 base)

qwen3:4b (2.5 GB, 38.1M downloads, Apache-2.0, 32K native context extendable to 131K with YaRN) is the sweet spot at this tier. It is the smallest Qwen3 with the full multilingual support and a usable context window for everyday summaries. gemma3:4b (3.3 GB, 128K context, multimodal) is the alternative if you want image input; the 3 GB difference is real on an 8 GB card.

16 GB usable VRAM (RTX 4060 Ti, M2 Pro, M3 Pro)

qwen3:8b (5.2 GB, 40K context, Apache-2.0) and mistral:7b (4.4 GB, 32K context, Apache-2.0) are the realistic picks. Both fit a 16 GB card with room for a long context. qwen3:8b is the stronger multilingual pick and the one with the lower licence friction. mistral:7b is the pick if you specifically want a non-Qwen3 non-Llama model and prefer Mistral’s tokenizer behaviour for European languages.

24 GB usable VRAM (RTX 4090, M3 Max, M4 Pro)

This is the tier where everyday chat stops feeling like a demo. Three real options:

  • mistral-small:24b (14 GB, 23.6B params, Q4_K_M, 32K context, Apache-2.0) - the recommended default. Strong instruction following, good multilingual (11 languages via Tekken), and 32K context for everyday documents and long threads.
  • qwen3:14b (9.3 GB, 40K context, Apache-2.0) - half the VRAM, similar chat quality, 100+ languages, but smaller context.
  • gemma3:12b (8.1 GB, 128K context, multimodal) - the Gemma option, gated licence, multimodal input.

mistral-small:24b is the most capable of the three on chat and writing tasks; qwen3:14b is the better pick if context length or multilingual coverage matters more than raw chat quality.

32 GB+ usable VRAM (workstations, dual GPUs)

qwen3:32b (20 GB, 40K context, Apache-2.0) is the highest-quality fully open-weight model most people can run at home. qwen3:30b-a3b (19 GB, 256K context) is the MoE alternative that opens up long-context work. llama3.3:70b (43 GB, 128K context, 8 languages) is the frontier chat model at this tier but requires Meta’s licence and most of a 48 GB card to run at full Q4.

For everyday tasks, qwen3:32b is the practical ceiling. Going higher buys benchmarks more than it buys a better chat experience.

What the licences actually let you do

The licence filter matters more than the size filter for anyone shipping product, redistributing weights, or using AI inside a company above Meta’s 700M-MAU threshold. The Apache-2.0 models (Qwen3, Mistral Small 24B, Mistral 7B) and the MIT model (Phi-4) impose no usage clause beyond attribution. Llama 3.2/3.3 require acceptance of Meta’s Llama Community License and carve out the 700M-MAU restriction. Gemma 3 requires acceptance of Google’s Gemma Terms of Use and is gated on Hugging Face.

For everyday personal use, any of the five works. For commercial or redistributable work, narrow to Apache-2.0 and MIT first; that is Qwen3, Mistral Small, Mistral 7B, and Phi-4. The open-model licences hub has the wider matrix.

A few things that are not on the cards

Tokens per second are not on the cards. This site does not publish tokens-per-second numbers for hardware it does not own. Two facts are documentable from the primary sources: the 1B-3B tier runs at tens of tokens per second even on integrated graphics, and the 24B tier is the first one where long-context summarization starts to feel slow on consumer GPUs. Run the model on the hardware in front of you and trust your own number.

The advertised context window is not what fits in VRAM. The context-window article covers the KV-cache math; the takeaway is that a 24 GB card with mistral-small:24b (32K context) holds the full advertised window, while a 16 GB card with qwen3:4b (advertised 32K native, extendable to 131K with YaRN) drops effective context well before 131K.

Thinking mode is not free. Several Qwen3 sizes can run in a “thinking” mode that improves reasoning but spends tokens. The thinking-mode cost article has the breakdown; for everyday chat the default is non-thinking.

The Bottom Line

For everyday chat, writing, and summarizing, the size ladder from 8 GB laptops to 32 GB workstations is short: qwen3:1.7b or llama3.2:3b at the bottom, qwen3:4b in the middle, qwen3:8b or mistral:7b at 16 GB, mistral-small:24b at 24 GB, and qwen3:32b at 32 GB. The two non-obvious calls are the licence filter (narrow to Apache-2.0 for commercial work) and the multimodal question (Gemma 3 is the only fully open-weight multimodal choice on this list). Pick the model whose licence you can live with, match the size to your hardware, and verify the context window on the hardware in front of you rather than trusting the advertised number.