Best open-weight model in November 2026: three leaderboards and the memory each top model needs

On 7 November 2026 the best open-weight model is Kimi K3.1 Pro: it scores 46 on the Artificial Analysis Intelligence Index v5.0.1, where the best closed model, Google's Gemini 4 Carbon (High), scores 60. On LMArena's text leaderboard (dated 6 November), the top open model is still its predecessor Kimi K3 (Max) at overall rank 19, 38 points behind first place. Kimi K3.1 Pro needs 1.56 TB for its weights, though; the best open models that fit one 128 GB DGX Spark at 4 bits score about 20 on the same index.

AA scores are shown rounded on the site; the decimals come from the data embedded in the leaderboard page. Gaps are my subtraction. The LMArena #1 is flagged preliminary on 4,483 votes; against the non-preliminary #2, Gemini 4 Argon (High) at 1523, the gap is 35 points.

Which open models lead the Artificial Analysis index?

Kimi K3.1 Pro leads the next open model by 13.9 points, about the same distance (13.6) that separates it from the closed leader. The index is in version 5.0.1, introduced in November 2026 per the methodology page, and weights four categories at 25% each: knowledge work (AA-Briefcase v1.1, AutomationBench-AA, GDP.pdf), coding (Terminal-Bench 4.0, FrontierCode 2.0), science (CritPt v2 Hard Subset, Terminal-Bench-Science 0.1) and general (AA-Omniscience v1.1, HLE-Diamond). Artificial Analysis runs every evaluation itself, so these are not vendor numbers. Its open-weights page counts any model whose weights can be downloaded, whatever the license; the leaderboard says 40 of its 99 ranked models qualify.

Read on artificialanalysis.ai and the per-model pages (for example Kimi K3.1 Pro) on 7 November 2026. The leaderboard shows no update date. Cost per task uses each model's cheapest listed API price, so it says little about self-hosting cost. Further down: Nemotron 3.5 Super VL 13.6, Inkling Small 12.9, Inkling (Xhigh) 12.5. gpt-oss-120b and Qwen3.8-Flash-Next have model pages but no v5 score yet.

The token column matters if you run these models yourself. Ling 3.1 Flash spent 240 million output tokens on the index and Step 5 Preview 129 million for almost the same score (29.6 against 28.5), so at equal tokens per second Ling takes nearly twice as long to finish the same work. Reasoning budgets get their own planned article on token budgets for reasoning models.

Where do open models rank on LMArena?

The best open model on LMArena's text board sits at overall rank 19, and the top 18 are all proprietary. The board moved to arena.ai; its version dated 6 November 2026 counts 9,140,146 votes across 419 models, 228 of them under open licenses. LMArena scores are pairwise human preferences, so they reward answer style as well as correctness.

Read on arena.ai with the open-source license filter and on the overall board, both showing "Nov 6, 2026", on 7 November 2026. The page is rendered in the browser, so I read it through a text extractor rather than raw data; two separate reads gave the same rows. The closed top five are Gemini 4 Carbon (High) 1526, Gemini 4 Argon (High) 1523, Muse Wonder (Max) 1513, Claude Opus 5.5 (High) 1505 and Claude Fable 5 (High) 1504.

Several new models have no row yet: Kimi K3.1 Pro, Muse Spark Open, MiniMax-M3.1-Flash, Ling 3.1 and the October DeepSeek update. Step 5 Preview appears at rank 83 with 1448 points but is labelled proprietary there, while Artificial Analysis counts it as open weights (StepFun publishes the weights under its own license).

What does Epoch's capabilities index add?

Epoch AI's ECI puts the best open model at rank 19, but it does not include Kimi K3.1 Pro yet, so it shows a narrower gap than Artificial Analysis. The ECI page loads its table in the browser; I sorted the linked CSV, labelled "Updated Nov. 7, 2026", myself.

• Closed top five: GPT-6 Astra 169.07, GPT-6.1 Sol 168.94, Claude Opus 5.5 168.36, GPT-6.1 Terra 167.82, Claude Fable 5.1 166.90.
• Open: DeepSeek V4.1 Flash 1017 159.31 (#19), Kimi K3 158.51 (#21), GLM-5.3 156.91 (#29), DeepSeek V4.1 Flash 156.49 (#33), DeepSeek V4 Pro 0813 155.87 (#37), Muse Spark Open 155.42 (#39), Mistral Large 4 154.66 (#45), GLM-5.3-Flash 153.64 (#49), Qwen 3.8 27B 152.73 (#54).

Epoch labels access too, and its labels are stricter than Artificial Analysis's: it marks Kimi K3 as "non-commercial", and GLM-5.3, Muse Spark Open and Mistral Large 4 as "restricted use". The license terms behind those labels are in the planned article on open-weight model licenses.

How much memory does each top open model need?

None of the top eight open models on the Artificial Analysis index fits one 128 GB Spark at 4 bits, and the top one needs about 1.56 TB. I compared the smallest builds on Hugging Face (file totals, checked 7 November 2026) with about 120 GB usable per Spark (121.6 GiB addressable, minus room for the KV cache) and 28 GB on a 32 GB GPU. The models guide explains the 121.6 GiB figure.

Sizes in GB (10⁹ bytes) from each repo's file listing on Hugging Face, 7 November 2026. "Spark" means one 128 GB DGX Spark; two Sparks pool about 240 GB across a network link, not as one memory space. "About 4 Sparks' memory" is a capacity comparison only, not a tested configuration.

The quality cost of the sub-2-bit builds is real. The site's GLM-5.3-Flash page covers what a 1- to 2-bit Unsloth Dynamic quant gives up, and the two-node cluster guide explains why a second Spark adds memory but not single-user speed.

Why do the three leaderboards disagree?

They measure different things, and they update at different speeds. Artificial Analysis runs agentic task suites on its own schedule and already scores Kimi K3.1 Pro, announced on 23 October. LMArena needs thousands of human votes per model, which takes weeks; MiMo-V2.6-Pro's ±9 interval on 4,189 votes shows how wide the error is early on. Epoch's ECI is a statistical blend of many benchmarks, some of them older.

Effort settings also change the ranking more than most readers expect. Artificial Analysis lists each reasoning effort as a separate row: Qwen3.8 27B scores 20.1 at Xhigh, 17.5 at Medium and 10.0 with reasoning off. If you serve a model with thinking turned down to save time, the number that applies to you is the lower one.

The open label is not consistent either. Step 5 Preview is "proprietary" on LMArena and open on Artificial Analysis. Kimi K3 is "non-commercial" on Epoch, while its license allows commercial use with conditions. Read the license, not the badge.

Which open model should you pick?

Pick by the hardware you can actually run, then by license, and only then by rank.

• GPU server or API: Kimi K3.1 Pro if its MaaS and display terms work for you, DeepSeek V4.1 Flash 1017 if you need MIT. DeepSeek costs $0.60 per index task through APIs against $3.68 for Kimi.
• Two Sparks: GLM-5.3-Flash in NVIDIA's NVFP4 build (204.4 GB, MIT, 27.8), the highest-scoring model that fits two Sparks at 4 bits. MiMo-V2.6-Flash at MXFP4 (167.4 GB, MIT, 26.7) leaves more room for context.
• One Spark: GLM-5.3-Flash at 2 bits if you can accept the quantization loss and want the score (27.8 in its native precision), Qwen3.8-27B or Nanbeige4.3-33B in FP8 or BF16 if you want full precision.
• 32 GB GPU: Qwen3.8-27B or Nanbeige4.3-33B at 4 bits.

My verdict: for one 128 GB box, Qwen3.8-27B in FP8 is the safest choice in November 2026. It is Apache 2.0, its 30.9 GB of weights leave about 93 GiB for context, and it gives up 26 index points to Kimi K3.1 Pro, whose weights are 50 times larger. Leaderboards will not tell you how either handles your documents, so score candidates on your own prompts first. GPUwerk rents a Spark at $0.79/hour for exactly that kind of test (pricing). For coding-specific rankings, see the companion article on the best open-weight coding models.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• LLM Leaderboard: Comparison of AI models, Artificial Analysis
• Open source (open weights) models, Artificial Analysis
• Intelligence benchmarking methodology, Artificial Analysis
• Kimi K3.1 Pro: intelligence, performance and price analysis, Artificial Analysis
• Text Arena: LLM Rankings, LMArena (arena.ai), 2026-11-06
• Text Arena, open-source license filter, LMArena (arena.ai), 2026-11-06
• Epoch Capabilities Index, Epoch AI, 2026-11-07
• ECI model scores CSV, Epoch AI, 2026-11-07
• moonshotai/Kimi-K3.1-Pro, Hugging Face, 2026-11-05
• unsloth/GLM-5.3-GGUF, Hugging Face
• unsloth/DeepSeek-V4.1-Flash-1017-GGUF, Hugging Face
• nvidia/GLM-5.3-Flash-NVFP4, Hugging Face