Best local LLM for reasoning · 2026
The best local LLM for reasoning and maths, ranked for your VRAM
The reasoning shortlist is ordered by what the vendors' own evaluations and the community report for maths, competition problems and step-by-step thinking — AIME, MathArena, Humanity's Last Exam, Codeforces — as of the catalogue date. For each memory class below we list every model on that list that actually fits at Q4_K_M with an 8K context, so the top row is the pick and the rest are what you give up by choosing it. Numbers are computed, not benchmarked; how the shortlist works.
Reasoning models think out loud before they answer, and the thinking is tokens: the trace for a hard problem routinely runs past 10K. Treat 8K as the floor, not the target — if the "max context" column reads under 32K, quantise the KV cache to q8_0 or step down a quantisation before you hand it anything difficult.
Best reasoning LLM for 6 GB VRAM
5.3 GB usable · speeds for a GeForce RTX 3060 Laptop everything that fits 6 GB →Qwen3.5 4B is the pick: Step-by-step reasoning in 8 GB. 1 other model on the reasoning shortlist also fits, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Qwen3.5 4B Alibaba · Apache 2.0 |
Step-by-step reasoning in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 66K | 78 | Runs great |
| 2 | Qwen3.5 2B Alibaba · Apache 2.0 |
Reasoning traces on 4 GB; expect long outputs. | 2.27B | Q4_K_M · 2.0 GB | 256K | 159 | Runs great |
Best reasoning LLM for 8 GB VRAM
7.0 GB usable · speeds for a GeForce RTX 4060 everything that fits 8 GB →Qwen3.5 9B is the pick: Thinking mode on a 12 GB card — ahead of the 2025 R1 distills. 2 other models on the reasoning shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Qwen3.5 9B Alibaba · Apache 2.0 |
Thinking mode on a 12 GB card — ahead of the 2025 R1 distills. | 9.65B | Q4_K_M · 6.3 GB | 31K | 30 | Tight fit |
| 2 | Qwen3.5 4B Alibaba · Apache 2.0 |
Step-by-step reasoning in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 120K | 63 | Runs great |
| 3 | Qwen3.5 2B Alibaba · Apache 2.0 |
Reasoning traces on 4 GB; expect long outputs. | 2.27B | Q4_K_M · 2.0 GB | 256K | 129 | Runs great |
Best reasoning LLM for 12 GB VRAM
10.6 GB usable · speeds for a GeForce RTX 3060 12 GB everything that fits 12 GB →Qwen3.5 9B is the pick: Thinking mode on a 12 GB card — ahead of the 2025 R1 distills. 2 other models on the reasoning shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Qwen3.5 9B Alibaba · Apache 2.0 |
Thinking mode on a 12 GB card — ahead of the 2025 R1 distills. | 9.65B | Q4_K_M · 6.3 GB | 146K | 40 | Runs great |
| 2 | Qwen3.5 4B Alibaba · Apache 2.0 |
Step-by-step reasoning in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 236K | 83 | Runs great |
| 3 | Qwen3.5 2B Alibaba · Apache 2.0 |
Reasoning traces on 4 GB; expect long outputs. | 2.27B | Q4_K_M · 2.0 GB | 256K | 171 | Runs great |
Best reasoning LLM for 16 GB VRAM
14.4 GB usable · speeds for a GeForce RTX 5080 everything that fits 16 GB →gpt-oss 20B is the pick: Low/medium/high reasoning effort in 13 GB. 3 other models on the reasoning shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | gpt-oss 20B OpenAI · Apache 2.0 |
Low/medium/high reasoning effort in 13 GB. | 20.9B MoE | MXFP4 · 11.6 GB | 126K | 120 | Runs great |
| 2 | Qwen3.5 9B Alibaba · Apache 2.0 |
Thinking mode on a 12 GB card — ahead of the 2025 R1 distills. | 9.65B | Q4_K_M · 6.3 GB | 256K | 107 | Runs great |
| 3 | Qwen3.5 4B Alibaba · Apache 2.0 |
Step-by-step reasoning in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 256K | 222 | Runs great |
| 4 | Qwen3.5 2B Alibaba · Apache 2.0 |
Reasoning traces on 4 GB; expect long outputs. | 2.27B | Q4_K_M · 2.0 GB | 256K | 455 | Runs great |
Best reasoning LLM for 24 GB VRAM
22.4 GB usable · speeds for a GeForce RTX 4090 everything that fits 24 GB →Gemma 4 31B is the pick: 89% AIME — the strongest maths of any model under 32B. 8 other models on the reasoning shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Gemma 4 31B Google · Apache 2.0 |
89% AIME — the strongest maths of any model under 32B. | 31.3B | Q4_K_M · 20.2 GB | 21K | 35 | Tight fit |
| 2 | Qwen3.8 27B Alibaba · Apache 2.0 |
Thinking mode plus the best agentic reasoning at 24 GB. | 27.8B | Q4_K_M · 16.7 GB | 98K | 39 | Runs great |
| 3 | Qwen3.6 27B Alibaba · Apache 2.0 |
Nearly Qwen3.8's reasoning, one release older. | 27.8B | Q4_K_M · 16.7 GB | 98K | 39 | Runs great |
| 4 | Qwen3.6 35B-A3B Alibaba · Apache 2.0 |
Reasoning traces at MoE speed. | 35.9B MoE | Q4_K_M · 20.9 GB | 82K | 126 | Tight fit |
| 5 | Ornith 1.5 35B-A3B Ornith AI · MIT |
Always thinks first; MIT-licensed. | 35.9B MoE | Q4_K_M · 20.9 GB | 82K | 126 | Tight fit |
| 6 | gpt-oss 20B OpenAI · Apache 2.0 |
Low/medium/high reasoning effort in 13 GB. | 20.9B MoE | MXFP4 · 11.6 GB | 128K | 126 | Runs great |
| 7 | Qwen3.5 9B Alibaba · Apache 2.0 |
Thinking mode on a 12 GB card — ahead of the 2025 R1 distills. | 9.65B | Q4_K_M · 6.3 GB | 256K | 112 | Runs great |
| 8 | Qwen3.5 4B Alibaba · Apache 2.0 |
Step-by-step reasoning in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 256K | 233 | Runs great |
| 9 | Qwen3.5 2B Alibaba · Apache 2.0 |
Reasoning traces on 4 GB; expect long outputs. | 2.27B | Q4_K_M · 2.0 GB | 256K | 478 | Runs great |
Best reasoning LLM for 32 GB VRAM
30.4 GB usable · speeds for a GeForce RTX 5090 everything that fits 32 GB →Gemma 4 31B is the pick: 89% AIME — the strongest maths of any model under 32B. 8 other models on the reasoning shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Gemma 4 31B Google · Apache 2.0 |
89% AIME — the strongest maths of any model under 32B. | 31.3B | Q4_K_M · 20.2 GB | 73K | 62 | Runs great |
| 2 | Qwen3.8 27B Alibaba · Apache 2.0 |
Thinking mode plus the best agentic reasoning at 24 GB. | 27.8B | Q4_K_M · 16.7 GB | 226K | 69 | Runs great |
| 3 | Qwen3.6 27B Alibaba · Apache 2.0 |
Nearly Qwen3.8's reasoning, one release older. | 27.8B | Q4_K_M · 16.7 GB | 226K | 69 | Runs great |
| 4 | Qwen3.6 35B-A3B Alibaba · Apache 2.0 |
Reasoning traces at MoE speed. | 35.9B MoE | Q4_K_M · 20.9 GB | 256K | 225 | Runs great |
| 5 | Ornith 1.5 35B-A3B Ornith AI · MIT |
Always thinks first; MIT-licensed. | 35.9B MoE | Q4_K_M · 20.9 GB | 256K | 225 | Runs great |
| 6 | gpt-oss 20B OpenAI · Apache 2.0 |
Low/medium/high reasoning effort in 13 GB. | 20.9B MoE | MXFP4 · 11.6 GB | 128K | 224 | Runs great |
| 7 | Qwen3.5 9B Alibaba · Apache 2.0 |
Thinking mode on a 12 GB card — ahead of the 2025 R1 distills. | 9.65B | Q4_K_M · 6.3 GB | 256K | 200 | Runs great |
| 8 | Qwen3.5 4B Alibaba · Apache 2.0 |
Step-by-step reasoning in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 256K | 414 | Runs great |
| 9 | Qwen3.5 2B Alibaba · Apache 2.0 |
Reasoning traces on 4 GB; expect long outputs. | 2.27B | Q4_K_M · 2.0 GB | 256K | 850 | Runs great |
Best reasoning LLM for 48 GB VRAM
46.4 GB usable · speeds for an RTX 6000 Ada everything that fits 48 GB →Gemma 4 31B is the pick: 89% AIME — the strongest maths of any model under 32B. 8 other models on the reasoning shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Gemma 4 31B Google · Apache 2.0 |
89% AIME — the strongest maths of any model under 32B. | 31.3B | Q4_K_M · 20.2 GB | 175K | 33 | Runs great |
| 2 | Qwen3.8 27B Alibaba · Apache 2.0 |
Thinking mode plus the best agentic reasoning at 24 GB. | 27.8B | Q4_K_M · 16.7 GB | 256K | 37 | Runs great |
| 3 | Qwen3.6 27B Alibaba · Apache 2.0 |
Nearly Qwen3.8's reasoning, one release older. | 27.8B | Q4_K_M · 16.7 GB | 256K | 37 | Runs great |
| 4 | Qwen3.6 35B-A3B Alibaba · Apache 2.0 |
Reasoning traces at MoE speed. | 35.9B MoE | Q4_K_M · 20.9 GB | 256K | 120 | Runs great |
| 5 | Ornith 1.5 35B-A3B Ornith AI · MIT |
Always thinks first; MIT-licensed. | 35.9B MoE | Q4_K_M · 20.9 GB | 256K | 120 | Runs great |
| 6 | gpt-oss 20B OpenAI · Apache 2.0 |
Low/medium/high reasoning effort in 13 GB. | 20.9B MoE | MXFP4 · 11.6 GB | 128K | 120 | Runs great |
| 7 | Qwen3.5 9B Alibaba · Apache 2.0 |
Thinking mode on a 12 GB card — ahead of the 2025 R1 distills. | 9.65B | Q4_K_M · 6.3 GB | 256K | 107 | Runs great |
| 8 | Qwen3.5 4B Alibaba · Apache 2.0 |
Step-by-step reasoning in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 256K | 222 | Runs great |
| 9 | Qwen3.5 2B Alibaba · Apache 2.0 |
Reasoning traces on 4 GB; expect long outputs. | 2.27B | Q4_K_M · 2.0 GB | 256K | 455 | Runs great |
Best reasoning LLM for 16 GB unified memory
10.7 GB usable · speeds for an M4 · 16 GB everything that fits 16 GB →Qwen3.5 9B is the pick: Thinking mode on a 12 GB card — ahead of the 2025 R1 distills. 2 other models on the reasoning shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Qwen3.5 9B Alibaba · Apache 2.0 |
Thinking mode on a 12 GB card — ahead of the 2025 R1 distills. | 9.65B | Q4_K_M · 6.3 GB | 149K | 12 | Runs great |
| 2 | Qwen3.5 4B Alibaba · Apache 2.0 |
Step-by-step reasoning in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 239K | 26 | Runs great |
| 3 | Qwen3.5 2B Alibaba · Apache 2.0 |
Reasoning traces on 4 GB; expect long outputs. | 2.27B | Q4_K_M · 2.0 GB | 256K | 53 | Runs great |
Best reasoning LLM for 64 GB unified memory
48.0 GB usable · speeds for an M4 Max · 64 GB everything that fits 64 GB →Gemma 4 31B is the pick: 89% AIME — the strongest maths of any model under 32B. 8 other models on the reasoning shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Gemma 4 31B Google · Apache 2.0 |
89% AIME — the strongest maths of any model under 32B. | 31.3B | Q4_K_M · 20.2 GB | 185K | 17 | Runs great |
| 2 | Qwen3.8 27B Alibaba · Apache 2.0 |
Thinking mode plus the best agentic reasoning at 24 GB. | 27.8B | Q4_K_M · 16.7 GB | 256K | 20 | Runs great |
| 3 | Qwen3.6 27B Alibaba · Apache 2.0 |
Nearly Qwen3.8's reasoning, one release older. | 27.8B | Q4_K_M · 16.7 GB | 256K | 20 | Runs great |
| 4 | Qwen3.6 35B-A3B Alibaba · Apache 2.0 |
Reasoning traces at MoE speed. | 35.9B MoE | Q4_K_M · 20.9 GB | 256K | 77 | Runs great |
| 5 | Ornith 1.5 35B-A3B Ornith AI · MIT |
Always thinks first; MIT-licensed. | 35.9B MoE | Q4_K_M · 20.9 GB | 256K | 77 | Runs great |
| 6 | gpt-oss 20B OpenAI · Apache 2.0 |
Low/medium/high reasoning effort in 13 GB. | 20.9B MoE | MXFP4 · 11.6 GB | 128K | 76 | Runs great |
| 7 | Qwen3.5 9B Alibaba · Apache 2.0 |
Thinking mode on a 12 GB card — ahead of the 2025 R1 distills. | 9.65B | Q4_K_M · 6.3 GB | 256K | 56 | Runs great |
| 8 | Qwen3.5 4B Alibaba · Apache 2.0 |
Step-by-step reasoning in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 256K | 116 | Runs great |
| 9 | Qwen3.5 2B Alibaba · Apache 2.0 |
Reasoning traces on 4 GB; expect long outputs. | 2.27B | Q4_K_M · 2.0 GB | 256K | 239 | Runs great |
Best reasoning LLM for 128 GB unified memory
96.0 GB usable · speeds for an M4 Max · 128 GB everything that fits 128 GB →Ling 3.0 Flash 124B-A5B is the pick: AIME 93 claimed at 5B active; hybrid attention keeps long reasoning traces cheap. 10 other models on the reasoning shortlist also fit, in order of capability:
| # | Model | Why | Params | Quant · total | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|
| 1 | Ling 3.0 Flash 124B-A5B inclusionAI · MIT |
AIME 93 claimed at 5B active; hybrid attention keeps long reasoning traces cheap. | 124B MoE | Q4_K_M · 70.4 GB | 256K | 50 | Runs great |
| 2 | Gemma 4 31B Google · Apache 2.0 |
89% AIME — the strongest maths of any model under 32B. | 31.3B | Q4_K_M · 20.2 GB | 256K | 17 | Runs great |
| 3 | Qwen3.8 27B Alibaba · Apache 2.0 |
Thinking mode plus the best agentic reasoning at 24 GB. | 27.8B | Q4_K_M · 16.7 GB | 256K | 20 | Runs great |
| 4 | Qwen3.6 27B Alibaba · Apache 2.0 |
Nearly Qwen3.8's reasoning, one release older. | 27.8B | Q4_K_M · 16.7 GB | 256K | 20 | Runs great |
| 5 | gpt-oss 120B OpenAI · Apache 2.0 |
Adjustable reasoning effort; designed for one 80 GB card. | 117B MoE | MXFP4 · 61.4 GB | 128K | 54 | Runs great |
| 6 | Qwen3.6 35B-A3B Alibaba · Apache 2.0 |
Reasoning traces at MoE speed. | 35.9B MoE | Q4_K_M · 20.9 GB | 256K | 77 | Runs great |
| 7 | Ornith 1.5 35B-A3B Ornith AI · MIT |
Always thinks first; MIT-licensed. | 35.9B MoE | Q4_K_M · 20.9 GB | 256K | 77 | Runs great |
| 8 | gpt-oss 20B OpenAI · Apache 2.0 |
Low/medium/high reasoning effort in 13 GB. | 20.9B MoE | MXFP4 · 11.6 GB | 128K | 76 | Runs great |
| 9 | Qwen3.5 9B Alibaba · Apache 2.0 |
Thinking mode on a 12 GB card — ahead of the 2025 R1 distills. | 9.65B | Q4_K_M · 6.3 GB | 256K | 56 | Runs great |
| 10 | Qwen3.5 4B Alibaba · Apache 2.0 |
Step-by-step reasoning in 8 GB. | 4.66B | Q4_K_M · 3.5 GB | 256K | 116 | Runs great |
| 11 | Qwen3.5 2B Alibaba · Apache 2.0 |
Reasoning traces on 4 GB; expect long outputs. | 2.27B | Q4_K_M · 2.0 GB | 256K | 239 | Runs great |
The whole reasoning shortlist, and what each one needs
Every model on the list, best first, with the smallest discrete card and the smallest Mac that run it with 15% headroom at the recommended quantisation and 8K context. Every class has its own page, Macs from 8 to 512 GB included — the full list.
| # | Model | Params | Weights | Needs (GPU) | Needs (Mac) | Ollama tag |
|---|---|---|---|---|---|---|
| 1 | DeepSeek V4 Pro 1.6T-A49B The 0813 refresh of the V4 flagship. Included as the honest ceiling; a terabyte of weights at Q4. |
1650B MoE | 927.8 GB @ Q4_K_M | more than one card | no single Mac | — |
| 2 | DeepSeek V4.1 Flash 552B-A16B New Twice V4 Flash's backbone (552B, 16B active per generated token) plus 196B of Engram lookup tables (189 GiB at FP8) that DwarfStar streams from SSD, so they are not counted here. A 512 GB Mac model, and as of October 2026 no mainline llama.cpp or Ollama build: DwarfStar on a Mac, vLLM across four GPUs. DeepSeek puts the cache at 890 bytes/token; modelled conservatively as 4 latent layers plus a 128-token window. |
552B MoE | 310.4 GB @ Q4_K_M | more than one card | 512 GB Mac | — |
| 3 | GLM-5.3 744B-A40B Same base as GLM-5.2, new post-training: Z.ai's most capable open-weights coder (Terminal-Bench 3.0 28.3, DeepSWE 66.9). Not MIT any more — its own GLM-5.3 licence. Same ceiling as 5.2: 512 GB of unified memory at Q4. |
753B MoE | 423.4 GB @ Q4_K_M | more than one card | no single Mac | — |
| 4 | GLM-5.3-Flash 320B-A18B The first natively multimodal GLM-5 and the first hybrid: 34 linear-attention blocks and 11 sparse-attention blocks with a 512-wide latent cache, so a 1M window stays affordable. Z.ai says it beats GLM-5.2 at 18B active; MIT. |
321B MoE | 180.5 GB @ Q4_K_M | more than one card | 512 GB Mac | — |
| 5 | DeepSeek V4 Flash 284B-A13B The V4 for 192–256 GB machines: Q3 squeezes into 192 GB, Q4 wants 256 GB, and 128 GB falls short even at Q2. Cache is modelled as a 576-wide latent; V4 compresses it further at long context, so this is conservative. |
284B MoE | 159.7 GB @ Q4_K_M | 192 GB card | 256 GB Mac | — |
| 6 | Qwen3.8-Flash-Next 180B-A6B The open preview of the Qwen4 architecture: a 125B-A6B hybrid (Gated DeltaNet + sparse attention, KV cache on 12 of 48 blocks) plus a 51B n-gram embedding and a 4B draft head — 180B on disk, 6B active. The hosted "Qwen3.8-Flash" is this model with a 1M window. |
180B MoE | 101.2 GB @ Q4_K_M | 141 GB card | 192 GB Mac | qwen3.8-flash-next:125b-a6b-nvfp4 |
| 7 | Ling 3.0 Flash 124B-A5B A 124B hybrid (5 linear-attention layers per MLA layer) with 5.1B active: SWE-bench Pro 56.6 and AIME 93 claimed. Built for 96–128 GB machines. |
124B MoE | 69.7 GB @ Q4_K_M | 141 GB card | 128 GB Mac | — |
| 8 | Ornith 1.5 397B-A17B The flagship Ornith on the Qwen3.5 397B-A17B architecture, MIT-licensed. A 256 GB Mac Studio at Q4, and it is in the Ollama library. |
397B MoE | 223.2 GB @ Q4_K_M | more than one card | 512 GB Mac | ornith-1.5:397b |
| 9 | Gemma 4 31B The dense flagship: strongest maths of the 24–32 GB class (89% AIME), clean prose, vision. Q4 is a tight 24 GB fit. |
31.3B | 17.6 GB @ Q4_K_M | 32 GB card | 32 GB Mac | gemma4:31b |
| 10 | Qwen3.8 27B The current default local Qwen: dense 27B, text + image + video, 262K context. Only 16 of its 64 blocks keep a KV cache, so long context is cheap. |
27.8B | 15.6 GB @ Q4_K_M | 24 GB card | 32 GB Mac | qwen3.8:27b |
| 11 | Qwen3.6 27B The 24 GB coding pick of spring 2026 (77.2 SWE-bench Verified). Same shape as 3.8, one generation behind. |
27.8B | 15.6 GB @ Q4_K_M | 24 GB card | 32 GB Mac | qwen3.6:27b |
| 12 | gpt-oss 120B Designed to land on one 80 GB card. Only ~5B parameters are active per token. |
117B MoE | 60.5 GB @ MXFP4 | 80 GB card | 128 GB Mac | gpt-oss:120b |
| 13 | Qwen3.6 35B-A3B Mixture of experts with ~3B active: the fastest serious model a 24 GB card runs, and the best MoE under 40B on agentic coding. |
35.9B MoE | 20.2 GB @ Q4_K_M | 32 GB card | 36 GB Mac | qwen3.6:35b-a3b |
| 14 | Ornith 1.5 35B-A3B A reasoning-first MIT build on the Qwen3.6 35B-A3B architecture (thinks before every answer). Same VRAM as its base. |
35.9B MoE | 20.2 GB @ Q4_K_M | 32 GB card | 36 GB Mac | ornith-1.5:35b |
| 15 | gpt-oss 20B Ships natively in MXFP4, so the 4-bit weights are the reference weights, not a lossy copy. Fits 16 GB. |
20.9B MoE | 10.8 GB @ MXFP4 | 16 GB card | 24 GB Mac | gpt-oss:20b |
| 16 | Qwen3.5 9B The default for 8–12 GB cards in 2026: beats every older 8B on every published benchmark, with vision. |
9.65B | 5.4 GB @ Q4_K_M | 10 GB card | 16 GB Mac | qwen3.5:9b |
| 17 | Qwen3.5 4B The 8 GB coding agent. Q4 lands near 3.4 GB, leaving room for a real context window. |
4.66B | 2.6 GB @ Q4_K_M | 6 GB card | 8 GB Mac | qwen3.5:4b |
| 18 | Qwen3.5 2B Phone-class, and multimodal. Replaces Llama 3.2 3B as the "it runs on anything" answer. |
2.27B | 1.3 GB @ Q4_K_M | 6 GB card | 8 GB Mac | qwen3.5:2b |
Thinking is a switch on most of these — Qwen's /think and /no_think, gpt-oss's low, medium and high effort. Leave it off for chat and turn it on for the problem: the speed per token does not change, but a reasoning answer is many more tokens, so the tokens-per-second column counts twice.
Other jobs