Reasoning models spend tokens before they answer. They generate a chain of thought, weigh alternatives, and then produce the final reply, which is why a 7B reasoning model can out-think a 70B chat model for problems that benefit from working through steps. The cost is speed and VRAM: a reasoning pass for a hard math problem can run to thousands of tokens before the answer appears, and the working memory for those tokens has to live on your card.
The open-weight landscape for reasoning has settled into three usable families, plus a handful of specialist entrants. Every figure below was pulled on 2026-09-09 from the Hugging Face API or the live Ollama library tag page, not from a recap post.
Quick reference
| Job | Model | Ollama tag (size) | Licence |
|---|---|---|---|
| Smallest usable reasoning | DeepSeek-R1-Distill-Qwen-1.5B | deepseek-r1:1.5b (1.1 GB) | MIT |
| Phone-class reasoning | Phi-4-mini-reasoning | phi4-mini-reasoning:3.8b (3.2 GB) | MIT |
| 8 GB sweet spot | DeepSeek-R1-Distill-Qwen-7B | deepseek-r1:7b (4.7 GB) | MIT |
| 8 GB sweet spot, native thinking | Qwen3-8B | qwen3:8b (5.2 GB) | Apache 2.0 |
| 16 GB tier flagship | DeepSeek-R1-Distill-Qwen-14B | deepseek-r1:14b (9.0 GB) | MIT |
| 16 GB tier alternative | Qwen3-14B | qwen3:14b (9.3 GB) | Apache 2.0 |
| 16 GB tier, hybrid thinking | gpt-oss-20B | gpt-oss:20b (14 GB) | Apache 2.0 |
| 24 GB tier pick | DeepSeek-R1-Distill-Qwen-32B | deepseek-r1:32b (20 GB) | MIT |
| 24 GB tier alternative | QwQ-32B | qwq:32b (20 GB) | Apache 2.0 |
| 24 GB tier, NC licence | EXAONE-Deep-32B | exaone-deep:32b (19 GB) | Non-Commercial |
| 32 GB+ reasoning frontier | DeepSeek-R1-Distill-Llama-70B | deepseek-r1:70b (43 GB) | MIT |
| Top-tier MoE flagship | Qwen3-235B-A22B-Instruct-2507 | qwen3:235b (142 GB) | Apache 2.0 |
The two takeaways before the details: most picks are MIT or Apache 2.0, which removes the usual licence obstacle, and the sizes cluster on powers of two because most reasoning models are either distilled from a base Qwen or Llama, or are full MoEs.
Three families, plus a long tail
DeepSeek-R1 and its distillations. DeepSeek-R1 is the 671B MoE flagship (MIT, ungated, 128K context) that the January 2025 release put on the map. The May 2025 refresh, DeepSeek-R1-0528, sits at 685B parameters, also MIT and ungated. Neither fits a consumer card. The five distillations do: DeepSeek applied R1’s outputs as a fine-tuning signal onto Qwen 1.5B/7B/14B/32B and Llama 3.3 70B, each MIT and ungated. Sizes on the Ollama library are 1.1 GB, 4.7 GB, 9.0 GB, 20 GB and 43 GB respectively. A seventh tag, deepseek-r1:8b at 5.2 GB, is a Qwen3-8B distillation from the 0528 release, not the original R1 line; the Ollama page lists its architecture as qwen3.
Qwen3 with thinking mode, plus the standalone QwQ. Qwen3-32B (Apache 2.0, ungated, 32K native context extendable to 128K via YaRN) introduced hybrid thinking. The model emits a <think>...</think> block when enable_thinking=True (the default) and can be toggled per turn with /think and /no_think markers. Three sizes matter for consumer hardware: Qwen3-8B at 5.2 GB, Qwen3-14B at 9.3 GB, and Qwen3-32B at 20 GB on the Ollama library. Qwen3-235B-A22B-Instruct-2507 is the flagship MoE at 142 GB (qwen3:235b); the model card reports AIME 2025 at 70.3 and GPQA at 77.5 (the Qwen team’s own measurement), with a native context of 262,144 tokens extendable to 1,010,000 via Dual Chunk Attention. QwQ-32B is the earlier Qwen2.5-32B fine-tune for reasoning; Apache 2.0, ungated, 40K context, the same 20 GB Ollama tag. QwQ has not been superseded by Qwen3, which is its own hybrid model - the two coexist.
The long tail. Phi-4-reasoning is a 14B MIT Microsoft model that fits at phi4-reasoning:14b (11 GB) with a 32K context; Phi-4-mini-reasoning is its 3.8B sibling at 3.2 GB (phi4-mini-reasoning:3.8b) with a 128K context, the only reasoning model that runs at useful speed on a phone. Magistral-Small-2506 is Mistral’s 24B Apache 2.0 attempt, on Ollama as magistral:24b at 14 GB. EXAONE-Deep-32B ships at exaone-deep:32b (19 GB), but the LICENSE file is the EXAONE AI Model License Agreement 1.1 - NC: Section 3.1 prohibits commercial use of the model, derivatives or output. That is the one licence trap in the family. Finally, openai/gpt-oss-20b and openai/gpt-oss-120b are OpenAI’s August 2025 reasoning releases, both Apache 2.0 and ungated, exposing reasoning_effort (low/medium/high) and separate analysis and final output channels. On Ollama they are gpt-oss:20b (14 GB) and gpt-oss:120b (65 GB) with 128K context each.
Why thinking-mode hybrid is the actual change
A pure reasoning model is a one-way bet: every reply pays the chain-of-thought cost. Hybrid thinking, which Qwen3 and gpt-oss both ship, lets the same model answer simple chat without thinking and switch into a chain-of-thought block for hard problems. The practical effect is that you stop needing two models on the card, a small chat model for everyday messages and a reasoning model for the one math problem per week, because one Qwen3-32B can do both jobs at 20 GB.
A reasoning reply to a hard question can spend 2,000 to 8,000 tokens on the thinking block before the answer, and each token carries KV-cache weight on the card. A 20 GB model with a 32K context window loses headroom fast if the thinking block runs long; check the VRAM arithmetic before assuming a tier page applies.
Picks by tier
Under 4 GB. deepseek-r1:1.5b at 1.1 GB is the smallest reasoning model that is genuinely reasoning, not a chat model with a reasoning label. phi4-mini-reasoning:3.8b at 3.2 GB is the better pick if you have a few gigabytes spare, with a 128K context window the DeepSeek distill cannot match. This tier is where the no-GPU cluster entry point meets reasoning: these tags run acceptably on a modern CPU, slowly.
4 to 6 GB. deepseek-r1:7b at 4.7 GB is the standout at this budget, MIT and a real chain-of-thought model. qwen3:8b at 5.2 GB is the alternative if you want hybrid thinking for chat and reasoning from one download. exaone-deep:7.8b at 4.8 GB is here too, but the non-commercial licence disqualifies it for anything you charge money for. The full 8 GB tier guide lists the chat picks for this budget.
9 to 15 GB. deepseek-r1:14b at 9.0 GB and qwen3:14b at 9.3 GB are the two reasonable picks. gpt-oss:20b at 14 GB is the alternative if you specifically want OpenAI’s reasoning_effort dial, and magistral:24b at 14 GB is the Mistral option. See the 16 GB guide for the broader picks.
19 to 22 GB. This is the dense reasoning tier. deepseek-r1:32b, qwen3:32b and qwq:32b all sit at 19 to 20 GB and all run on a 24 GB card with headroom for context. EXAONE-Deep at 19 GB is here too, again non-commercial. The 24 GB tier guide covers the chat alternatives.
Above 24 GB. deepseek-r1:70b at 43 GB is the strongest MIT reasoning model that fits on a single high-end card. gpt-oss:120b at 65 GB is the Apache 2.0 MoE alternative with adjustable reasoning effort. qwen3:235b at 142 GB is the flagship MoE (235B total, 22B active); it competes with hosted front-ends but needs the kind of card the 32 GB tier guide covers plus offload. The 671B DeepSeek-R1 at 404 GB is data-centre territory only.
Licences, where the traps are
Most of the family is permissive. DeepSeek-R1, all five distillations, Phi-4-reasoning and Phi-4-mini-reasoning are MIT. Qwen3 (all sizes), QwQ, Magistral and gpt-oss are Apache 2.0. None gate access on Hugging Face. Read the broader patterns in the open-model licences guide, but for this category the single trap is EXAONE-Deep, whose LICENSE file is the LG AI Research non-commercial terms. If you want the same reasoning coverage commercially, deepseek-r1:32b or qwen3:32b cover the gap.
How to actually run one
A reasoning model wastes the most time if you do not cap the thinking budget. Ollama’s defaults are generous; set a tighter num_predict cap or, for Qwen3, prefix the prompt with /no_think for queries where you do not want the chain. The Ollama runners comparison lists which runners expose the relevant knobs. Quantization matters here: the GGUF, AWQ, GPTQ and MLX comparison covers the format trade-offs, and what quantization costs you covers the quality loss, which on a reasoning model is more visible than on a chat model because the thinking block amplifies any degradation step by step.
The bottom line
For most people with a 16 GB or 24 GB card, the answer is qwen3:14b or qwen3:32b if you want hybrid thinking, or deepseek-r1:14b or deepseek-r1:32b if you want a pure reasoning model under MIT. Below 8 GB, phi4-mini-reasoning:3.8b is the smallest reasoning model worth running. Above 24 GB, deepseek-r1:70b or gpt-oss:120b covers the single-card frontier. The only commercial-licence trap in the family is EXAONE-Deep; everything else in the table is MIT or Apache 2.0 and ungated, and the Ollama tag sizes quoted are the live library values pulled on 2026-09-09.