llmrun

Monthly14 min read

Best local LLMs, November 2026: picks for 8 to 48 GB

Three new open Score leaders, all 359 GB or more. For home use, Qwen3.8 27B remains the pick from 24 GB, and Qwen3.5 9B below.

The best open model available on 16 October by today's llmrun Score is Xiaomi's MiMo V2.6 Pro RL. Now a month later, it sits fourth behind three models released since 17 October, with Moonshot's Kimi K3.1 Pro on top at 139.8. The smallest of the three requires 359 GB of memory at Q4_K_M, so a desk with one graphics card is not enough. Our picks for the hardware most people own have remained the same as last month, with August's Qwen3.8 27B still the best on any GPU or Mac from 24 GB to 128 GB and Qwen3.5 9B still the best below 24 GB.

Key takeaways

  1. The month's new leaders require 359 to 1,712 GB of memory at Q4_K_M. The new 512 GB Mac Studio is the only Mac that holds any of them.
  2. From 24 GB to 128 GB, Qwen3.8 27B remains the best model. Any open model with a higher Score needs at least 183 GB, ten times the 17.9 GB for our pick.
  3. Qwen3.5 9B is still the best choice for cards below 24 GB. Bigger models can squeeze onto a 16 GB card, but only if you give up most of the context.
  4. The 27B is the best choice for people who have 24 GB and put answer quality first. A mixture-of-experts model writes about four times faster by our estimate, but scores about 15 points lower.
  5. The month's new 27B, Nous Research's Arete 27B, is a Qwen3.8 27B fine-tune that has not yet been scored.

Three new leaders, none of them for your desk

Moonshot announced Kimi K3.1 Pro on 23 October, with the weights arriving on Hugging Face in the first days of November. It is the same size as Kimi K3 at 2.8 trillion parameters, and has about 104 billion parameters active per token plus, per its model card, a 1-million-token context. Moonshot's own table gives it a result within a point of Claude Opus 5.5 on Terminal-Bench 4.0. Those figures are the vendor's, while ours come from public leaderboards and put Kimi K3.1 Pro about 5 points clear of every other open model.

Qualityllmrun Score leaderboard: Three of the five best open models arrived after mid-Octoberllmrun Score · higher is better · bar = 90% range
  1. Kimi K3.1 Pro#1 · range 136–143 · 1712 GB139.8
  2. DeepSeek V4.1 Flash 1017#2 · range 132–137 · 458 GB134.5
  3. Inkling 2 Small#3 · range 130–138 · 359 GB134.1
  4. MiMo V2.6 Pro RL#4 · range 131–137 · 614 GB133.5
  5. Kimi K3#5 · range 131–135 · 1670 GB132.8
Source: llmrun Score v3-irt-4, 46 benchmarks · Data as of Nov 21, 2026Full ranking: 108 open models →

Our rule makes second and third a tie, as each Score is inside the other's 90% range. Mira Murati's Thinking Machines Lab introduced Inkling 2 Small on 5 November as a hosted model only, and published the weights a week later, putting Inkling 2 Small under Apache 2.0 after outside safety testing. The other arrival, DeepSeek V4.1 Flash 1017, is a "minor update" to the September Flash according to DeepSeek's card, but enough of an update to reach the top three.

All three belong to the giant mixture-of-experts category. Kimi K3.1 Pro alone needs an estimated 1,712 GB at Q4_K_M, more memory than three of Apple's largest Mac Studios combined. For an 8-GPU server such as a DGX B300, our estimate is about 150 tok/s. This month's leaderboard is news to read about; for models to run at home, look at the picks below.

Below 183 GB, an August model still rules

In August, Qwen3.8 27B was released under Apache 2.0, and the Qwen team's card describes the 3.8 release as "the most capable generation in the Qwen open-model family to date". In our Score it ranks 15th among open models of any size, behind only mixture-of-experts models at least ten times its size.

121.9
Qwen3.8 27B, 17.9 GB at Q4_K_M
183GB
memory for the first higher Score (DeepSeek V4 Flash 0731, 127.6)
1,712GB
memory for the top open Score (Kimi K3.1 Pro, 139.8)

The reason for the 27B's lead is Qwen's rapid improvement at the same size. There have been three generations of the 27B this year, all with the same 17.9 GB file at Q4_K_M.

QualityBar chart: Same 17.9 GB, 18 more Score points in six monthsllmrun Score, Q4_K_M needs 17.9 GB for all three
  • Qwen3.5 27B (February)103.8
  • Qwen3.6 27B (April)111.2
  • Qwen3.8 27B (August)121.9
Source: llmrun.dev/benchmark/llmrun-score · Data as of Nov 14, 2026

The memory requirement has not changed at all. Anyone with a 24 GB card bought in February has had two free upgrades, each just a download. That is why a monthly list can stay unchanged for three months, with the next jump at this size coming with the next Qwen if the pattern holds.

Nous Research's new 27B this month is Arete 27B, a fine-tune of Qwen3.8 27B. Nous Research says in its model card that the supervised training was meant to "reinforce western values on a Chinese-trained model", with no RL and no changes to the architecture. The speed is about the same as its base's by our estimate, but without a llmrun Score it cannot be a pick yet. This month's smaller arrivals, such as Nanbeige4.3 33B and SenseNova Loop 9B Preview, also lack enough public benchmark results for a Score.

MemoryPickScoreNeeds at Q4_K_MEst. speedSince October
8 GBQwen3.5 9B99.46.6 GB45.8 tok/s, RTX 5060unchanged
12 GBQwen3.5 9B99.46.6 GB68.7 tok/s, RTX 5070unchanged
16 GBQwen3.5 9B99.46.6 GB98.1 tok/s, RTX 5080unchanged
24 GBQwen3.8 27B121.917.9 GB37.6 tok/s, RTX 4090unchanged
32 GBQwen3.8 27B121.917.9 GB66.9 tok/s, RTX 5090unchanged
48 GBQwen3.8 27B121.917.9 GB35.8 tok/s, RTX 6000 Adaunchanged

The speeds given are llmrun estimates for one reference card per size (how they're made), as of 2026-11-16. The 48 GB row is slower than the 32 GB row because the reference card is an RTX 6000 Ada with less memory bandwidth than an RTX 5090. For the full rankings, see 8 GB, 12 GB, 16 GB, 24 GB and 48 GB.

8, 12 and 16 GB: Qwen3.5 9B, still

For small cards, the pick has been Qwen3.5 9B since February. It supports images and text, with a 6.6 GB footprint at Q4_K_M and estimated text output of about 46 tokens a second on an 8 GB RTX 5060. By our estimate, even the cheapest card here writes several times faster than people read at about 238 words a minute, roughly four words a second.

A smaller version is worth considering. At only 3.4 GB, Qwen3.5 4B has a Score of 98.5 and is tied with the 9B by our rule. That leaves enough memory on an 8 GB card for a long conversation, or for a desktop using some of the GPU as well. The catalogue's 218 models that fit 8 GB include a lot of old ones, with Gemma 2 9B from 2024 about 31 points below the 9B. For this reason, our picks only go back twelve months.

At 12 GB the 9B still has no better, newer rival, and the extra memory can be used for context. Qwen3.8 27B can technically fit a 12 GB RTX 5070 at IQ3_XXS, but there is no memory left and the fit gets a C from us.

Things get more interesting, and tighter, at 16 GB. The 9B is still the safe choice, with an estimated speed of about 98 tok/s on an RTX 5080. At Q4_K_M, Qwen3.8 27B is 1.9 GB too large for the card, but it will load at Q3_K_L in 15.4 GB. The mixture-of-experts Gemma 4 26B A4B can be loaded at Q4_K_S. Both take up 97% of the card, leaving little room for context, and both get a B from us. Although it is from August 2025 and older than our cut-off, GPT OSS 20B has the best coding score that fits at Q4_K_M, with room to spare at 13.1 GB. For a comparison of all five 16 GB cards, see our 16 GB GPU review.

FitModel picks table: At 16 GB the 9B leads, with GPT OSS 20B the coding choiceRanked by llmrun Score · best quant that fits · speeds estimated on the RTX 5080
#ModelSpeed on RTX 5080llmrun ScoreQuant · VRAM
1
Q4_K_M · 6.6 GB · 98.1 t/s
98.1 tok/sQ4_K_M · 6.6 GB
2
Qwen3.5 4BAlibaba
Q4_K_M · 3.4 GB · 191 t/s
191 tok/sQ4_K_M · 3.4 GB
3
Q4_K_M · 13.1 GB · 183 t/s
183 tok/sQ4_K_M · 13.1 GB
Source: llmrun catalogue and llmrun Score · Data as of Nov 21, 2026All 310 models on 16 GB →

24 and 32 GB: Qwen3.8 27B, or a MoE for speed

The 27B needs 17.9 GB of an RTX 4090's memory, or three quarters of the card, and its estimated speed is about 38 tokens a second. That is faster than reading, though the wait is noticeable on long answers. Its own predecessor Qwen3.6 27B comes next among dense models, about 11 points behind. For every quant and context length, see How much VRAM does Qwen 3.8 27B need?

Speed is the trade here. Since a mixture-of-experts model reads only a few billion of its parameters per token, it writes much faster than a dense model with the same memory footprint.

SpeedBar chart: On an RTX 4090, MoE models run 4 to 5.3 times as fast as their dense counterpartstok/s, estimated, RTX 4090, Q4_K_M
  • Gemma 4 26B A4B (MoE)171.1
    Score 105.8
  • Qwen3.6 35B A3B (MoE)151
    Score 106.8
  • Qwen3.8 27B (dense)37.6
    Score 121.9
  • Gemma 4 31B (dense)32.1
    Score 106.3
Source: llmrun.dev/gpu/tier/24gb-vram · Data as of Nov 16, 2026

Look at the notes rather than the bars: both MoE models are about 15 points below the 27B in Score. Qwen3.6 35B A3B needs 22 GB on a 24 GB card, and this leaves little room for context. Gemma 4 26B A4B's memory requirement is 16.5 GB. We would prefer the 27B for answer quality and a MoE for less waiting on many short replies. We rate 244 of the 329 models that fit 24 GB as running great on an RTX 4090.

The same 27B takes up 56% of a 32 GB RTX 5090, with an estimated speed of about 67 tokens a second. The extra memory allows longer context and a second model, with estimated MoE speeds of over 200 tok/s. Prices and cost per token are covered in our RTX 5090 article.

FitModel picks table: The 27B also leads on a 32 GB RTX 5090Ranked by llmrun Score · best quant that fits · speeds estimated on the RTX 5090
#ModelSpeed on RTX 5090llmrun ScoreQuant · VRAM
1
Q4_K_M · 17.9 GB · 66.9 t/s
66.9 tok/sQ4_K_M · 17.9 GB
2
Q4_K_M · 17.9 GB · 66.9 t/s
66.9 tok/sQ4_K_M · 17.9 GB
3
Q4_K_M · 21.7 GB · 57.1 t/s
57.1 tok/sQ4_K_M · 21.7 GB
4
Q4_K_M · 16.5 GB · 250 t/s
250 tok/sQ4_K_M · 16.5 GB
Source: llmrun catalogue and llmrun Score · Data as of Nov 21, 2026All 349 models on the RTX 5090 →

48 GB: more room, the same answer

Of the 370 models that fit a 48 GB card such as the RTX 6000 Ada, Qwen3.8 27B is still the best by Score. The larger models now within reach fit only with 2-bit quants. At IQ2_M, Qwen3.5 122B A10B needs 42.7 GB and has a Score of 109.4, already below the 27B before the loss from 2-bit quantization. GPT OSS 120B is very fast, with an estimated speed of about 195 tok/s, and fits in the same way.

As at 24 GB, the extra memory is for longer context, higher quants and a second model alongside the first, with no better model by our Score until about 183 GB. For quality alone, do not choose a 48 GB card on this month's data.

Macs: only the 512 GB Studio changes the answer

With our 3.5 GB system allowance, a 32 GB Mac has only 28.5 GB available for a model to use. For Macs as for GPUs, the pick is the top Score at Q4_K_M and the letter is our fit grade.

MacUsable memoryTop pick at Q4_K_MEst. speed, grade
M5, 16 GB12.5 GBQwen3.5 9B16.9 tok/s, B
M5 Pro, 24 GB20.5 GBQwen3.8 27B12.3 tok/s, C
M5 Max, 36 GB32.5 GBQwen3.8 27B18.5 tok/s, B
M5 Max, 64 GB60.5 GBQwen3.8 27B24.7 tok/s, A
M5 Max, 128 GB124.5 GBQwen3.8 27B24.7 tok/s, A
M5 Ultra, 256 GB252.5 GBMiMo V2.6 Flash RL (Score 131.9)49.8 tok/s, A
M5 Ultra, 512 GB508.5 GBInkling 2 Small (Score 134.7)46.5 tok/s, A

The weak point is the 24 GB Mac mini M5 Pro, since the 27B uses 87% of its usable memory. On this slow prompt-reading chip, we estimate about 44 seconds before the first word of a reply to a 4,000-token prompt. Our prompt-reading estimates look about twice too cautious in light of community measurements, though even 20 seconds is a long wait. The 36 GB Mac Studio M5 Max gets a B, while the 27B has comfortable room from 64 GB. The chip choices are covered in Mac for local LLMs, late 2026.

FitModel picks table: A 64 GB Mac Studio M5 Max runs the 27B with room to spareRanked by llmrun Score · best quant that fits · speeds estimated on the Mac Studio M5 Max (64 GB)
#ModelSpeed on Mac Studio M5 Max (64 GB)llmrun ScoreQuant · VRAM
1
Q4_K_M · 17.9 GB · 24.7 t/s
24.7 tok/sQ4_K_M · 17.9 GB
2
Q4_K_M · 17.9 GB · 24.7 t/s
24.7 tok/sQ4_K_M · 17.9 GB
3
Q4_K_M · 21.7 GB · 21.1 t/s
21.1 tok/sQ4_K_M · 21.7 GB
4
Q4_K_M · 16.5 GB · 80.7 t/s
80.7 tok/sQ4_K_M · 16.5 GB
Source: llmrun catalogue and llmrun Score · Data as of Nov 21, 2026All 373 models on the Mac Studio M5 Max (64 GB) →

This month's Mac news is really about the Mac Studio M5 Ultra with 512 GB. Apple set the price on 28 October at $22,299, with the first deliveries expected in late November. Two 256 GB Studios cost $701 less, as Cult of Mac pointed out, but the machines do not share a memory pool. It is the only Mac with room for this month's new models, and our estimates put Inkling 2 Small at about 47 tok/s and DeepSeek V4.1 Flash 1017 at about 52. For the 256 GB model, the pick is MiMo V2.6 Flash RL at 186.5 GB. From 24 to 128 GB the Mac pick is still the 27B, with the same estimated 49 tok/s on an M5 Ultra with 96 or 512 GB, since the speed depends on memory bandwidth rather than capacity.

Best local LLM for coding, reasoning, vision and tool use

Here too, we use Q4_K_M on the reference card for each size. Coding and reasoning have separate domain scores. Vision and tool use have no domain scores of their own, and for those we use the overall Score of models with support for them; the speeds given are llmrun estimates, not measured.

Task24 GB, RTX 409016 GB, RTX 5080
CodingQwen3.8 27B, coding 118.4GPT OSS 20B, coding 101.9
ReasoningQwen3.8 27B, reasoning 124Qwen3.5 9B, reasoning 97
VisionQwen3.8 27B, Score 121.9Qwen3.5 9B, Score 99.4
Tool useQwen3.8 27B, Score 121.9Qwen3.5 9B, Score 99.4
Fast chatQwen3.5 4B, 201 tok/sQwen3.5 4B, 191.4 tok/s
FitModel picks table: For coding at 24 GB the 27B leads by about 12 pointsRanked by llmrun Score · best quant that fits · speeds estimated on the RTX 4090
#ModelSpeed on RTX 4090Coding ScoreQuant · VRAM
1
Q4_K_M · 17.9 GB · 37.6 t/s
37.6 tok/sQ4_K_M · 17.9 GB
2
Q4_K_M · 21.7 GB · 32.1 t/s
32.1 tok/sQ4_K_M · 21.7 GB
3
Q4_K_M · 22.0 GB · 151 t/s
151 tok/sQ4_K_M · 22.0 GB
Source: llmrun catalogue and llmrun Score · Data as of Nov 21, 2026All 330 models on 24 GB →

The 27B's agents score of 122.3 is the best that fits 24 GB, but with no other scored model in that domain at this size there is no ranking to give. On the site's task pages (coding, tool use, vision, reasoning), quants below Q4_K_M are allowed, which puts Qwen3.8 27B first at Q3_K_L for 16 GB. More about coding agents is in SWE-bench Verified for local models.

How we pick, and what to watch in December

RuleWhat it means
QualityThe llmrun Score combines 46 public benchmarks into one number with a 90% range (method). 107 open models have one today. We never pick a model without one.
FitMemory at Q4_K_M, with room for a 16K-token context. GPUs use the size's reference card, Macs the usable memory (RAM minus 3.5 GB).
RecencyOnly models released since 2025-11-15 can be a pick. Older ones appear as named alternatives with their date.
SpeedEstimated by llmrun's calibrated model from memory bandwidth and active parameters, not measured. Real runs vary with runtime, context and drivers.
TiesScores inside each other's 90% range are ties. Qwen3.5 9B (96 to 102.8) and 4B (93.8 to 103.1) are one; we prefer the larger model.

The December list could change for three reasons. Inkling 2 Small's weights have only been public since 12 November, with community quants and runtime support still being added. Arete 27B and the month's smaller models will have a Score once there are enough public benchmark results, and a fine-tune beating its base would mark the first change at 24 GB since August. The first 512 GB Mac Studios should reach buyers, and we will be able to check our Mac estimates against their reports. Every change against this list will be marked in next month's edition.

Questions

What is the best local LLM for 24 GB of VRAM in November 2026?
Qwen3.8 27B has been the leader at Q4_K_M since August. For speed, our estimate for Qwen3.6 35B A3B is about 151 tok/s, but its Score is 106.8.
Can I run Kimi K3.1 Pro locally?
A single GPU or Mac is not enough. It needs an 8-GPU server such as a DGX B300, for which we estimate about 150 tok/s.
Is 16 GB enough for coding and agents?
Yes, for a small model: GPT OSS 20B (coding 101.9) needs 13.1 GB and our speed estimate on an RTX 5080 is about 182.8 tok/s. Qwen3.8 27B has little room for context, since it only loads at Q3_K_L.
Will a 48 GB card give me a better model than a 24 GB one?
No, by Score the top pick is the same on both cards. More memory gives you room for higher quants, much longer context and several models at once.
Can I run Qwen3.8 27B on a Mac?
Yes, from 24 GB, after allowing about 3.5 GB of RAM for the system. We estimate an M5 Pro with 24 GB (grade C, tight) at 12.3 tok/s, an M5 Max with 36 GB at 18.5 and an M5 Ultra at 49.3.