Open-weight LLMs in late 2026, sorted by the GPU memory they need

The open-weight LLM I would start with in November 2026 is Qwen3.8-27B: it is Apache-2.0, its FP8 build fits one 48 GB GPU with room for a long context, and only one other single-GPU model matches its score on Artificial Analysis's Intelligence Index. I sort this list by GPU memory first, because memory decides what you can run at all, and by licence second, because several of these models are open-weight without being open source. Mistral's Devstral 2 and Medium 3.5 give no rights to a company with more than $20 million in monthly revenue, MiniMax M3 asks commercial users for a notice and a credit, and Qwen3.8-Flash-Next, Kimi K3 and Meta's Muse Spark Open put conditions on selling inference: Qwen's at any revenue, the other two above $20 million a year.

How the tiers are sized#

The rule I follow is that the weights have to fit inside 92% of the memory the GPU reports, with room left for the KV cache, because 0.92 is the share vLLM claims by default. Every size below is the maker's own checkpoint, read from the Hugging Face API on 2026-11-12, and every fit is our calculation, not a test. Mixture-of-experts models are sized on total parameters: only a few experts run per token, but all of them have to sit in memory.

Here is the arithmetic once. Qwen3.8-27B's FP8 checkpoint is 30,866,866,928 bytes, which is 30.87 GB or 28.75 GiB. A 48 GB card with ECC on reports 46,068 MiB, and 92% of that is 41.39 GiB, so 12.64 GiB is left. Qwen3.8-27B keeps 64 KiB of cache per token (2 x 16 attention layers x 4 KV heads x 256 x 2 bytes), so that is room for about 207,000 tokens of context before activations. The KV cache guide explains the formula, and the LLM VRAM calculator runs it for your own context and number of users.

Up to 48 GB: one RTX A6000-class card#

Eight current models fit one 48 GB card with their maker's own checkpoint, and all eight are Apache-2.0.

All eight serve with vLLM, and Ollama's library has a tag for seven of them: gpt-oss:20b, gemma4:12b, gemma4:31b, devstral-small-2:24b, granite4.2:30b, qwen3.8:27b and qwen3.6:35b. I found no Ollama tag for Nanbeige4.3-33B on 2026-11-12. The 48 GB cards are Ampere or Ada generation. FP8 math needs compute capability 8.9 or newer, so on the Ampere RTX A6000 vLLM runs an FP8 checkpoint weight-only: you keep the memory saving and lose the FP8 speed. The two smallest Gemma 4 sizes, E2B and E4B, are built for laptops and phones, and OpenAI says gpt-oss-20b runs within 16 GB, so none of those needs a rented GPU for one user.

What I would start with: Qwen3.8-27B in FP8, because it leaves room for about 207,000 tokens of context and scores 20 on the Intelligence Index, against 4 for Gemma 4 31B. vLLM's recipe for it lists the FP8 build for Hopper and Blackwell GPUs only, so on these 48 GB cards treat that combination as untested. Nanbeige4.3-33B, whose repository is under three weeks old, matches that score, but its cache costs 208 KiB per token, so the 8.3 GiB it leaves holds about 42,000 tokens.

80 to 96 GB: one A100, H100 or RTX PRO 6000#

The 80 and 96 GB cards hold the 27B to 35B models at full BF16 precision, plus three larger models that ship in 4-bit formats.

Under about 8 GiB of headroom is a fit on paper only, so I treat the last three rows as 96 GB models. NVFP4 is the 4-bit format that Blackwell GPUs compute natively, and the Blackwell card in the on-demand catalog is the RTX PRO 6000. NVIDIA lists one RTX PRO 6000 as a deployment target for Nemotron 3.5 Super VL. The NVFP4 guide covers the format.

What I would start with: Qwen3.8-27B in BF16 on the RTX PRO 6000, because no quantization is involved and 36.2 GiB holds two full 262,144-token contexts.

141 to 192 GB: one H200 NVL or two RTX PRO 6000s#

The models in this tier are 119B to 284B parameters, and only one of the four is plain open source at a precision that fits a single card.

Mistral's repositories hold two copies of the weights, and these figures are for one. DeepSeek V4 Flash holds on two RTX PRO 6000s only on paper: vLLM's recipe sizes it at 200 GB and verifies it on eight of them. DeepSeek's older R1 and V3 models and the R1 distills are sized in DeepSeek GPU requirements.

What I would start with: still Qwen3.8-27B, with the extra memory spent on context and concurrent users, because Mistral Small 4 and Medium 3.5 score 4 and 5 on the same index. DeepSeek V4 Flash is the exception when you need its 1M-token context under an MIT licence.

Beyond 192 GB: a multi-GPU server, or more than one#

The models people search for most this autumn start at 173 GiB and end at 2.3 TiB, so they need two to eight GPUs in one server, or two servers. The last column is the smallest shape whose combined budget holds the weights alone, counting two, four or eight H200 NVLs or RTX PRO 6000s and eight-GPU B200 or B300 servers.

The shapes are a floor, and they are arithmetic: they leave little for the KV cache, and they do not say that an engine supports that checkpoint's format on that GPU. The makers' own serving examples use more, and DeepSeek's card serves V4 Flash on a four-GPU GB300 node. Qwen3.8-Flash-Next would squeeze onto two RTX PRO 6000s with 3.1 GiB to spare, which is not enough to serve from, so I list it in this tier. Kimi K3.1 Pro is the newest entry, with a repository created on 2026-11-02, and it is within half a gigabyte of K3's size. On QuantaCloud, two H200 NVLs or four RTX PRO 6000s are on-demand configurations when the catalog lists them. Four or eight H200s, eight RTX PRO 6000s and anything on B200 or B300 are reserved capacity that we order and build to your spec: send a capacity brief with the model and the context length you need.

What I would start with: Qwen3.8-Flash-Next in FP8 for the most capability per GiB, unless you sell inference or a coding or office assistant, in which case GLM-5.3-Flash is MIT and scores within a point of it.

What each licence allows#

The licence column decides more than the benchmark does, because the same weights can be free for an internal tool and off-limits for a product. Read the licence file in the repository before you build on a model, and treat this table as a summary, not legal advice.

Only the first two rows are open-source licences in the usual sense. The rest are why I say open-weight.

What the public leaderboards say#

The honest answer on quality is that no single ranking holds for every workload, so I point to leaderboards that publish their method and I date them. Artificial Analysis runs its Intelligence Index v5.0.1 on open and closed models alike: nine evaluations across knowledge work, coding and science. These are its figures on 2026-11-12, rounded to whole points, where the top closed model scores 65.

Two things stand out. Size does not buy rank: two models of about 30B parameters that fit one card outscore every 119B to 128B model it scores in the tier above them. And the gap to the largest models is real, because the top of the table needs 1.42 TiB of GPU memory.

Two other leaderboards are worth checking before you commit. The official Terminal-Bench 4.1 board, last updated on 2026-11-11, lists two open-weight models: GLM-5.3 at 41.2% and GLM-5.3-Flash at 35.5%, against 65.8% for the top entry. SWE-rebench scores models on problems collected from GitHub repositories in a dated window, and in its window of 2026-05-15 to 2026-07-01 it has GLM-5.2 at 62.9%, MiniMax M3 at 47.2%, DeepSeek V4 Pro at 40.2%, Qwen3.6-27B at 31.2% and Qwen3.6-35B-A3B at 24.7%. For code specifically, open-weight coding models goes through the benchmarks one model at a time.

Which QuantaCloud GPU holds which tier#

Every on-demand GPU has at least 48 GB, so the first tier runs on anything in the catalog.

Prices checked 14 Nov 2026, 09:11 UTC

Whatever you pick, a QuantaCloud instance starts with an empty disk, and stopping it deletes the disk with the weights on it. A 30 GB checkpoint comes back in minutes and a 500 GB one does not, which is one more reason the largest models belong on reserved hardware. Self-hosting an LLM covers the rest of the stack, and vLLM with Docker has the commands for a first server.

Questions about open-weight LLMs#

What is the best open-source LLM right now?

By the one index that scores most of them the same way, it is Kimi K3.1 Pro, at 46 on Artificial Analysis's Intelligence Index on 2026-11-12. It needs about 1.42 TiB of GPU memory. The best that fit a single GPU are Qwen3.8-27B and Nanbeige4.3-33B, both at 20 and both Apache-2.0.

What is the best local LLM for one GPU?

On a 48 GB card I would run Qwen3.8-27B in FP8. On a 16 to 24 GB card or a laptop, the makers' own small builds are the place to look: OpenAI says gpt-oss-20b runs within 16 GB, Meta's 4-bit Muse Glimmer build targets 24 GB, and Gemma 4 E4B is made for laptops.

Are open-weight models open source?

Some are. Apache-2.0 and MIT models can be used, changed and sold without asking anyone. The others publish their weights under a licence with conditions: a revenue ceiling, a credit line, or a separate agreement before you sell access to the model.

Can I run Kimi K3 or DeepSeek V4 Pro on one server?

On paper, yes, on the right server. Kimi K3 is 1,453.74 GiB on disk, which the budget of eight B300s covers and eight H200s does not. DeepSeek V4 Pro is 831.43 GiB, which fits eight H200 NVLs with about 200 GiB left for the cache. Both are built-to-order configurations.

Which of these models accept images?

Most of the 2026 releases do: Gemma 4, Qwen3.8-27B, Qwen3.8-Flash-Next, Qwen3.6, Muse Glimmer, Muse Spark Open, Devstral Small 2, Mistral Small 4, Medium 3.5 and Large 4, Nemotron 3.5 Super VL, GLM-5.3-Flash, DeepSeek V4.1 Flash, MiniMax M3 and M3.1 Flash, and Kimi K3 and K3.1 Pro. gpt-oss, Granite 4.2, Nanbeige4.3-33B, GLM-4.7-Flash, Devstral 2, DeepSeek V4, GLM-5.3 and Qwen3.8-2.4T-A95B are text only.

My decision rule: pick the licence you can live with, then the largest model that leaves real room for context on the GPU you are willing to pay for, and check it against a dated leaderboard before you trust a model card. For most people that is Qwen3.8-27B on a 48 GB card, moving to the RTX PRO 6000 for full precision and long context. LLM inference on QuantaCloud lists what the instances come with.