Qwen3.8-27B, Muse Glimmer-30B, and K2-Horizon on 16, 24, and 32 GB

By LLM Hardware team · Methodology ·

Beam, Granite Switch, Dory, and DeepSeek-V4.1-Flash-1017 added 17 October 2026 · Qwen3.8-27B files still the 3 October check

AI and automation help produce this page; every figure is checked against the cited sources and the site formula before publishing.

Key facts

Qwen3.8-27B files
Unsloth UD-Q4_K_S 15.4 GB to UD-Q4_K_XL 17.6 GB. Formula 18.2 GB. Ollama qwen3.8:27b listed 18 GB.
K2-Horizon-7B
Q4 7.4 GB by the formula on the Hub's 9B count; fits 8 GB
K2-Horizon-MoVA-36B-A4B
36B total / 4B active, Q4 23.6 GB by the formula
Checked
2026-10-13

Checked 2026-10-13. Author: LLM Hardware team. Amazon prices are not stored or shown on this site.

Citeable facts

Key fact: Qwen3.8-Flash-Next UD-Q4_K_XL is 111.33 GB. UD-IQ4_XS is 93.68 GB. A 32 GB GPU fits neither file.

How we got it: Unsloth Qwen3.8-Flash-Next-GGUF shard sums on 9 October 2026. Formula on 180B (125B plus 51B n-gram embedding and 4B MTP) is 110 GB. The Qwen Community License is not Apache-2.0. Qwen Cloud Flash is the hosted API, not these weights.

Checked:

Key fact: Nemotron 3.5 Bolt at Q4 is 46.4 GB. Muse Spark Open at Q4 is 491.6 GB. Neither Q4 fits a 32 GB GPU.

How we got it: 74 × 0.5 × 1.2 + 2 and 816 × 0.5 × 1.2 + 2. NVIDIA and Hugging Face cards, 15 October 2026. Bolt’s shipped local file is 2-bit, aimed at 24 GB.

Checked:

Key fact: Granite Switch 4.1 8B at Q4 is 7.3 GB and fits 8 GB. The 30B at Q4 is 20.0 GB and fits 24 GB. Dory 1.0 4B at Q4 is 4.4 GB and fits 8 GB. Beam at Q4 is 302.8 GB and does not fit 128 GB.

How we got it: Safetensors totals on 17 October 2026: granite-switch-4.1-8b 8,776,400,896; granite-switch-4.1-30b 29,989,814,272; Dory BF16 3,975,229,440; Beam 501,267,612,352. Q4 is params_B × 0.6 + 2.

Checked:

Key fact: Qwen3.8-27B at Q4 with a 32K-token chat is about 20.3 GB counting its 16 attention layers, or 26.8 GB counting all 64 layers. The first fits a 24 GB card; the second needs about 32 GB.

How we got it: 27 x 0.5 x 1.2 + 2 = 18.2 GB. KV per token at f16 = 2 x layers x 4 KV heads x 256 head size x 2 bytes (Qwen3.8-27B card and config.json: 16 full_attention and 48 linear_attention layers). 16 layers: 65,536 bytes x 32,768 = 2.1 GB. 64 layers: 8.6 GB.

Checked:

Key fact: Muse Glimmer-30B at Q4 with a 32K-token chat is about 21.5 GB and fits a 24 GB card with 2.5 GB left.

How we got it: 29.6 x 0.5 x 1.2 + 2 = 19.8 GB. KV: 2 x 52 layers x 2 KV heads x 128 head size x 2 bytes = 53,248 bytes per token x 32,768 = 1.7 GB (Muse-Glimmer-30B model card).

Checked:

Key fact: An 8 GB RTX 4060 holds Dory 1.0 4B, Granite Switch 4.1 8B and K2-Horizon-7B at Q4; a 36 GB Mac Studio M5 Max (about 32 GB usable) holds every 24 GB-tier model here plus Nemotron 3.5 Bolt's 25GB 2-bit download.

How we got it: Site formula at 0.5 bytes per parameter: 3.98B 4.4 GB, 8.78B 7.3 GB, 9B (Hub count) 7.4 GB. Mac: 36 GB less the 4 GB macOS reserve used in data/hardware.json. Bolt file size from the Ollama library.

Checked:

Methodology explains the VRAM formula, the bytes per weight at each quantization, and the full source list.

The models worth a local install in that window are not the announcement posts. We checked the Hugging Face cards, then mapped them onto memory tiers. This guide quotes no tokens-per-second figures of its own. Where a vendor published a speed, the guide says whose number it is.

What fits

MemoryCatalog machineReasonable to loadLeave for a bigger box
8 GB RTX 4060 (8GB) K2-Horizon-7B Q4 (7.4 GB). Dory 1.0 4B Q4 (4.4 GB). Granite Switch 4.1 8B Q4 (7.3 GB). Qwen3.8-27B. Granite Switch 30B Q4 (20.0 GB). Muse Glimmer-30B.
16 GB None linked here 7B–14B Q4. Gemma 4 12B. Qwen3.8-27B Q4 is 18.2 GB and does not fit. Leave 27B-class models for 24 GB
24 GB RTX 4090 (24GB), no buy link Qwen3.8-27B Q4 (18.2 GB). Granite Switch 4.1 30B Q4 (20.0 GB). Muse Glimmer K-Quant-17GB. Bolt only as the 2-bit file. Bolt at Q4 (46.4 GB). Beam at Q4 (302.8 GB). Muse Spark Open.
32 GB RTX 5090 (32GB), no buy link; Mac Studio M5 Max (36GB) Muse Glimmer dynamic quant (their 32 GB target). Qwen3.8-27B with more KV cache. Bolt’s 2-bit file, with more context room than 24 GB. Bolt at Q4 (46.4 GB). Muse Spark Open. Qwen3.8-Flash-Next.
64–128 GB unified See the 64 GB to 128 GB guide Super VL formula Q4 is 74 GB (96 or 128 GB, not 64). Ollama q4_K_M is 90GB: 128 GB fits, 96 GB is the edge. Flash-Next UD-IQ4_XS is 93.68 GB on 128 GB. Beam Q4 (302.8 GB). DeepSeek-V4.1-Flash-1017 files (510.3 GB). Muse Spark Open Q4 (491.6 GB).

Run your own context length through the VRAM calculator. The figures above are short-context estimates (4K to 8K tokens), not a 32K chat; the KV cache worked examples below add that. The 15 October pair is written up on the Nemotron 3.5 Bolt page and the 64 GB to 128 GB guide.

Tier by tier: machines and models

Each tier names a machine from our hardware catalog and the headline models it holds by the site formula, parameters x bytes per parameter x 1.2 + 2 GB, at a short context. Macs lose about 4 GB to macOS, as in our catalog. This page does not show prices or stock; the listing does.

8 GB: NVIDIA GeForce RTX 4060 (8GB)

Holds the small models in this roundup: Dory 1.0 4B at Q4 (4.4 GB) or Q8 (6.8 GB), Granite Switch 4.1 8B at Q4 (7.3 GB, 0.7 GB left) and K2-Horizon-7B at Q4 (7.4 GB on the Hub's 9B count, 0.6 GB left). The last two leave little room for a long context. Qwen3.8-27B (18.2 GB) and Muse Glimmer-30B (19.8 GB) do not fit, and the Bolt E4B q4_K_M download (7.9GB) leaves no room.

MSI RTX 4060 Ventus 2X Black 8G OC on Amazon

16 GB: a step, not a target

16 GB covers 7B to 14B models at Q4, but every 27B to 30B headline model here misses it: Qwen3.8-27B at 18.2 GB, Muse Glimmer-30B at 19.8 GB and Granite Switch 4.1 30B at 20.0 GB. This guide links no 16 GB machine.

24 GB: NVIDIA GeForce RTX 4090 (24GB)

The discrete tier for this roundup. Qwen3.8-27B at Q4 is 18.2 GB (5.8 GB left), Muse Glimmer-30B 19.8 GB (4.2 GB left), Granite Switch 4.1 30B 20.0 GB (4.0 GB left) and K2-Horizon-MoVA-36B-A4B 23.6 GB on its 36B total (0.4 GB left). Our catalog listed this card's Amazon listing as unavailable on 17 October 2026, so it is named here without a link. The 5090 vs 4090 guide compares it with the 32 GB tier.

32 GB: NVIDIA GeForce RTX 5090 (32GB) or Apple Mac Studio (M5 Max, 36GB)

Both give about 32 GB to work with: the RTX 5090 has 32 GB of GDDR7, and the Mac Studio's 36 GB of unified memory leaves about 32 GB after macOS. Either holds everything in the 24 GB tier with room for a 32K chat (worked examples below), Nemotron 3.5 Bolt's 25GB 2-bit download with about 7 GB left, and Muse Glimmer's dynamic quant, which Meta's card aims at 32 GB. Neither holds Qwen3.8-27B at Q8 (34.4 GB) or Granite Switch 4.1 30B at Q8 (38.0 GB). The RTX 5090 is named without a link because our catalog listed its Amazon listing as unavailable on 17 October 2026. Apple lists 460 GB/s for this Mac Studio's M5 Max (18-core CPU, 32-core GPU).

Mac Studio M5 Max 36 GB on Amazon

64 GB to 128 GB

Full-precision runs and the large mixture-of-experts models in this roundup start here. The machines are in the 64 GB to 128 GB guide.

Qwen3.8-27B

Qwen/Qwen3.8-27B is a 27B dense model. The Hub license field is Apache-2.0. The card describes a vision-language model with a 262,144-token native context. On 3 October 2026 the short-context formula was 27 × 0.5 × 1.2 + 2 = 18.2 GB. Unsloth’s GGUF tree the same day showed UD-Q4_K_S at 15.4 GB, Q4_0 at 16.1 GB, UD-Q4_K_M at 16.5 GB, and UD-Q4_K_XL at 17.6 GB. Those files are the weights. The formula is larger because of the 1.2 factor and 2 GB. A 16 GB card does not have room for KV cache on top of a 17.6 GB file. 24 GB is the discrete tier. 32 GB is headroom for a longer chat, not a different model. Ollama’s library listed qwen3.8:27b at 18 GB.

The older Qwen3 hardware guide still covers the 8B/14B/32B generation. qwen3:32b is not this model. The tag is qwen3.8:27b. Gemma 4’s comparable local MoE is on the what-can-I-run guide: 17.1 GB at Q4 when you count all 25.2B parameters. Kolibri-1, also dated 3 October, is about 78 GB FP8 and is a datacenter serve, not a row in the table above.

Muse Glimmer-30B

meta-models/Muse-Glimmer-30B is a ~29.6B dense model with a perception encoder. The model card says Apache-2.0 and an August 2026 release. Their own fit table is the one to trust: a K-Quant around 17 GB aimed at 24 GB, a dynamic quant aimed at 32 GB, and full precision aimed at 64 GB. They also say the 4-bit language-model weights come in under 20 GB, with KV cache and the vision encoder on top. A 16 GB gaming card is below the target they published.

Meta's card reports generation speed for the 17 GB quant plus their DFlash drafter: 74.9 tok/s baseline and 233.4 tok/s with speculation on an RTX 5090, 23.7 / 37.8 on an M4 Max, measured by them with batch size 1 and greedy decoding. Those are Meta's numbers, not ours. Speculation changes the picture; do not compare that 233 figure to a plain Q4 llama.cpp run of a different model.

K2-Horizon

IFM/K2-Horizon-7B is the dense 7B-class model (the Hub sidebar lists about 9B parameters, which usually includes embeddings). The card calls the release fully open — weights, data, and code — and dates the artifact index 1 October 2026. Read the LICENSE file before you ship a product on it. On the Hub's 9B count, Q4 is 9 x 0.5 x 1.2 + 2 = 7.4 GB by the site formula, so an 8 GB GPU holds it with 0.6 GB left. It has a 512K context on the card; you will not get 512K on 8 GB, because the KV cache will not fit. K2-3.7B is the smaller sibling if 8 GB feels tight once you turn context up.

K2-Horizon-MoVA-36B-A4B is the 36B / ~4B-active MoE from the same family (3 September 2026 in the research pass, Apache-2.0 listed there — confirm the file). Q4 of the full expert set is 36 x 0.5 x 1.2 + 2 = 23.6 GB by the site formula. Speed can resemble a small model. Memory cannot. A 16 GB card does not hold it. 24 GB is the minimum discrete tier, and it will be tight with context.

K2-32B, if you grab that checkpoint, behaves like other dense ~32B models: Q4 around 20 GB of weights, so 24 GB again, 32 GB if you want Q5/Q6 or a long chat.

How to run the headline models

Every command below is copied from the Ollama library or from the model's Hugging Face card, read on 30 October 2026. Where a vendor gives no command for the 4-bit file that fits a consumer card, the table says so instead of guessing one. The vLLM commands on the cards load full-precision weights, so the memory column shows what that costs.

ModelCommand (source)What it loadsMemoryMachines on this page
Qwen3.8-27Bollama run qwen3.8:27b (Ollama library)Q4_K_M, 18GB download18.2 GB by the formulaRTX 4090, RTX 5090, Mac Studio M5 Max 36 GB
Qwen3.8-27Bpip install vllm, then vllm serve "Qwen/Qwen3.8-27B" (Qwen model card)Full BF16 weights66.8 GB by the formulaNone
Muse Glimmer-30Bpip install vllm, then vllm serve "meta-models/Muse-Glimmer-30B" (Meta model card)Full precision73.0 GB by the formula; the card's own table aims full precision at 64 GBNone. The card's 24 GB and 32 GB quants come with no run command; the card points to its quantizations list.
K2-Horizon-7Bpip install vllm, then vllm serve "IFM/K2-Horizon-7B" (IFM model card)16-bit weights, 9B on the Hub sidebar23.6 GB by the formulaRTX 5090. A 24 GB card leaves 0.4 GB. The Q4 that fits 8 GB (7.4 GB) has no command on the card.
Nemotron 3.5 Boltollama run nemotron-3.5-bolt (Ollama library)74b-a9b 2-bit build, 25GB download25GB file; the Q4 formula (46.4 GB) does not apply to this buildRTX 5090 or Mac Studio M5 Max 36 GB, about 7 GB left. Over a 24 GB card.

Granite Switch 4.1 has no Ollama tag, and Dory 1.0 needs NVIDIA's llama.cpp fork; both are in the 17 October section below. After loading, the PROCESSOR column of ollama ps should read 100% GPU, which Ollama's FAQ says means the model is entirely in GPU memory.

A 32K chat: KV cache worked examples

The tier figures stop at a short context, 4K to 8K tokens. A longer chat adds a key-value cache, which the methodology sizes per token as 2 x layers x KV heads x head size x bytes per element, with 2 bytes at the f16 default. Ollama starts with a 4,096-token context, per its FAQ, so a 32K chat has to be asked for. Two headline models at 32,768 tokens:

Muse Glimmer-30B

Meta's card lists 52 layers, 2 KV heads and a head size of 128. That is 2 x 52 x 2 x 128 x 2 = 53,248 bytes per token, or about 1.7 GB at 32K. Added to the Q4 estimate of 29.6 x 0.5 x 1.2 + 2 = 19.8 GB, the total is about 21.5 GB. The RTX 4090 keeps 2.5 GB, and the RTX 5090 and the Mac Studio M5 Max 36 GB keep about 10.5 GB. A 16 GB card does not hold it.

Qwen3.8-27B

Qwen's card gives a hybrid layout: 16 blocks of three Gated DeltaNet layers and one Gated Attention layer, with 4 KV heads, and the config sets a head size of 256. Its layer list marks the 16 Gated Attention layers as full_attention and the other 48 as linear_attention. Applied to the 16 attention layers, the KV formula gives 2 x 16 x 4 x 256 x 2 = 65,536 bytes per token, about 2.1 GB at 32K, and a total of about 20.3 GB with the 18.2 GB Q4 estimate. The methodology's KV formula does not cover the 48 linear-attention layers. Counting all 64 layers as attention layers gives 8.6 GB of cache and 26.8 GB in total, the conservative case.

To try it, start Ollama with OLLAMA_CONTEXT_LENGTH=32768 ollama serve, then run ollama run qwen3.8:27b. For a q8_0 cache, set OLLAMA_FLASH_ATTENTION=1 and OLLAMA_KV_CACHE_TYPE=q8_0 on the same serve command; Ollama's FAQ says cache quantization only applies with Flash Attention on. For other context lengths, use the VRAM calculator.

The 24 GB card

For this set, 24 GB is the useful discrete step: Qwen3.8-27B Q4, Muse's 17 GB quant, and MoVA Q4. A used RTX 3090 or a 4090 does that job. An RTX 5090 adds 32 GB and a lot of bandwidth. Extra VRAM does not make a 7B model faster once it already fits, and 32 GB still does not fit Qwen3.8-Flash-Next. The 5090 vs 4090 guide has the bandwidth comparison and the 70B options.

RTX 4090 (24GB): no buy link. Our catalog listed its Amazon listing as unavailable on 17 October 2026.

On a Mac, the same 24 to 36 GB band covers these models: Mac mini M5 Pro 24 GB and M4 Pro 24 GB for Qwen3.8-27B Q4 (tight) and K2-7B, and Mac Studio M5 Max 36 GB for Qwen3.8-27B plus K2-32B / 36B-A4B. DeepSeek-V4-Flash at an aggressive quant is the EVO-X2 (128 GB at ~256 GB/s) or an Apple Ultra, not those minis and not a 64 GB DDR5 box at ~90 GB/s. The Beelink GTR9 Pro matches the EVO-X2’s bandwidth and adds a NIC caveat; the ranking is in the Strix Halo guide. The 48 GB and 64 GB Macs are in the Apple Silicon guide.

Also on the list, with caveats

15 October 2026: Bolt and Muse Spark Open

Nemotron 3.5 Bolt is 74B total and 9B active. NVIDIA’s card says the shipped local checkpoint stores experts at about 2 bits so it fits a 24 GB budget at batch size 1 and 128K context. Our Q4 formula on the total is 46.4 GB, which does not fit 24 GB or 32 GB. On 17 October 2026 Ollama listed nemotron-3.5-bolt:latest at 25GB. That download is the edge of a 24 GB card and may spill. Do not load a Q4. The parent Super VL model is 120B total. The Q4 formula is 74 GB. Ollama’s 120b-a12b-q4_K_M tag was 90GB the same day: 128 GB fits, and 96 GB is the edge after macOS.

Muse Spark Open is about 816B total and about 42B active. It is not Muse Glimmer. Q4 is 491.6 GB. The Hugging Face llama.cpp post says BF16 is about 1.5 TB, and that a 2-bit quant used 304 GB and 18 tokens per second on a 512 GB M3 Ultra. That speed is theirs. Unsloth says start with dynamic 2-bit on 256 GB. A 32 GB GPU and a 128 GB mini PC do not run it. Bolt E4B now has Ollama tags (nemotron-3.5-bolt:e4b-q4_K_M was a 7.9GB download on 30 October 2026), but this page has no parameter count for it yet, so it stays off the table, as do weights that are still missing, such as Mistral Large 4.

17 October 2026: Beam, Granite Switch, and Dory

Reflection Beam is 501.3B total and 23.4B active, Apache-2.0, text, 1M context. Q4 is 302.8 GB, so 64 GB, 96 GB, and 128 GB all miss it. Unsloth UD-Q4_K_XL summed to 306.5 GB and misses 256 GB. UD-IQ1_S summed to 103.1 GB: that 1-bit file can sit on 128 GB and does not fit 96 GB. ollama run beam is beam:cloud, not a local weight. Size another quant in the VRAM calculator.

Granite Switch 4.1 8B is 8.78B. Q4 is 7.3 GB and fits 8 GB. Q8 is 12.5 GB and needs 16 GB. Granite Switch 4.1 30B is 29.99B. Q4 is 20.0 GB, so it misses 16 GB and fits 24 GB. Q8 is 38.0 GB and misses 32 GB. There is no ollama run granite-switch tag. ollama run granite4.1:8b is the older base, not this Switch bundle.

Dory 1.0 4B is 3.98B, OpenMDW 1.1, 256K context. Q4 is 4.4 GB and Q8 is 6.8 GB, so both fit 8 GB. The published Q5_K_M GGUF was 3,673,288,864 bytes. NVIDIA’s card says to use a llama.cpp fork, not a stock Ollama tag. The 24 GB side of Bolt is on the Bolt hardware page, and the 64–128 GB miss for Beam is on the 64 GB to 128 GB guide.

Announced, not a recommendation

Nex-N2.5-Pro, DeepSeek-V4.1-Pro, and MiMo-V2.6 showed up as announcements or incomplete drops. This guide does not say which GPU "runs" them. When a repo has files, a license, and a size, it can go in the calculator. Until then it is a press post.

Questions

Does a 16 GB GPU run Qwen3.8-27B?

No, not as a comfortable fit. The Q4 formula is 27 × 0.5 × 1.2 + 2 = 18.2 GB. On 3 October 2026 Unsloth’s UD-Q4 files were 15.4 GB, 16.5 GB, and 17.6 GB, and Ollama listed qwen3.8:27b at 18 GB. A 16 GB card has no room left for KV cache. Use 24 GB. Muse Glimmer’s own card targets a 17 GB quant at 24 GB, not 16 GB.

Why is K2-Horizon-MoVA slower to load than it is to run?

It is a 36B-total Mixture-of-Experts with about 4B active parameters. Active parameters set tokens per second. Total parameters set VRAM, because every expert has to be resident. Q4 of the full model is 23.6 GB by the site formula (36 x 0.5 x 1.2 + 2), so plan a 24 GB card, where 0.4 GB is left, or a 32 GB machine for anything beyond a short chat, even though each token only touches a 4B slice.

Can I run DeepSeek-V4 on one gaming GPU?

No. DeepSeek-V4.1-Flash-1017, checked 17 October 2026, is a 763.2B safetensors total whose files sum to 510.3 GB. Q4 on that total is 459.9 GB. Neither figure fits 128 GB, 256 GB, or a 512 GB machine after a macOS reserve. DeepSeek-V4.1-Pro still has no verified local weight drop. Weights: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017

Can I run Muse Spark Open or Nemotron 3.5 Bolt on a gaming GPU?

Bolt’s published 2-bit checkpoint is sized for 24 GB. A Q4 of its 74B total is 46.4 GB and does not fit 24 GB or 32 GB. Muse Spark Open is 816B total. Q4 is 491.6 GB, so no gaming GPU and no 128 GB unified machine runs it. A 2-bit Muse file is a 256 GB-class download.

Which new models should I ignore until weights exist?

Nex-N2.5-Pro, DeepSeek-V4.1-Pro, and MiMo-V2.6 were announcements or incomplete releases in the 29 May–1 October 2026 window. Do not buy hardware for them until a Hugging Face repo you can download is up, with a license and a real file size.

Which machine in this guide holds Qwen3.8-27B with a 32K chat?

By the site formula, Qwen3.8-27B at Q4 is 18.2 GB. Its 16 attention layers add about 2.1 GB of f16 cache at 32K tokens, for about 20.3 GB, which fits a 24 GB card such as the RTX 4090 with 3.7 GB left. Counting all 64 layers as attention layers gives 26.8 GB, which needs about 32 GB: the RTX 5090, or the Mac Studio M5 Max with 36 GB, which keeps about 32 GB after macOS.

Is an 8 GB RTX 4060 enough for anything in this roundup?

Yes, for the small models: Dory 1.0 4B at Q4 (4.4 GB) or Q8 (6.8 GB), Granite Switch 4.1 8B at Q4 (7.3 GB) and K2-Horizon-7B at Q4 (7.4 GB on the Hub's 9B count). The last two leave under 1 GB for context. None of the 27B to 30B headline models fit.

How do I run these models locally?

Qwen3.8-27B at Q4: ollama run qwen3.8:27b, an 18GB download. Nemotron 3.5 Bolt's 2-bit build: ollama run nemotron-3.5-bolt, 25GB. Muse Glimmer-30B and K2-Horizon-7B: their model cards give vllm serve commands that load full-precision weights, about 73.0 GB and 23.6 GB by the site formula, and neither card gives a command for its 4-bit files.