Best open-weight coding models in November 2026: independent scores, vendor claims, memory and speed

The best open-weight coding model in November 2026 is Kimi K3.1 Pro: Artificial Analysis measured it at 67.2% on Terminal-Bench 4.0, above every closed model it lists, but it needs 1,561 GB of weights. Among models you can run at home or on one DGX Spark, the strongest independent scores belong to GLM-5.3-Flash (32.8%, fits one 128 GB Spark only at 2 bits) and Nanbeige4.3-33B (29.8%, a 20.8 GB 4-bit file that fits a 32 GB GPU). SWE-bench Verified, LiveCodeBench and Aider's leaderboard have no entries newer than February 2026, so I ranked on Terminal-Bench 4.0 and Artificial Analysis's coding runs, read on 7 November 2026.

Scores in percent. Terminal-Bench 4.0 and FrontierCode 2.0 are Artificial Analysis's own runs, read from the per-evaluation data on its leaderboard page; the Coding Agent Index v2.0 (FrontierCode v1.1, Terminal-Bench 4.0, SWE-Atlas-QnA, equal weights) is from its coding agents page, where Claude Opus 5.5 in Claude Code tops the board at 61.7. All read 7 November 2026; neither page shows an update date. Build sizes are Hugging Face file totals; the Qwen3.8-Flash-Next figure is the community NVFP4 recipe on our Qwen3.8-Flash-Next page. "2 Sparks" means two 128 GB DGX Sparks over a network link.

What do the independent leaderboards say?

Three independent sources have current open-weight coding results. They cover different sets of models, but wherever they overlap, GLM-5.3 is the strongest open model after Kimi K3.1 Pro.

• Artificial Analysis runs Terminal-Bench 4.0 and FrontierCode 2.0 itself for every model on its index (table above), and separately runs whole coding agents for its Coding Agent Index. GLM-5.3 in Opencode leads the open models there at 48.9; Kimi K3.1 Pro has no agent run yet.
• The official Terminal-Bench 4.0 board (record updated 23 October 2026, 330 trials per row) lists two open models: GLM-5.3 at 41.82% in Claude Code at max effort, dated 14 August, and GLM-5.3-Flash at 35.76% with reasoning off, dated 26 August. The top row is Claude Opus 5.5 in Claude Code at 64.85%. No Kimi, DeepSeek or MiniMax row exists yet.
• SWE-rebench (Nebius) reruns fresh GitHub tasks with its own agent. Its newest window, 15 May to 1 July 2026, predates most autumn releases but covers the mid-size models: GLM-5.2 resolves 62.9%, MiniMax M3 47.2%, MiMo V2.5 Pro 46.5%, DeepSeek-V4 Pro 40.2%, Qwen3.6-27B 31.2% and Qwen3.6-35B-A3B 24.7%.

The gap at the top is smaller than in general rankings. On the overall open-weight leaderboard, Kimi K3.1 Pro trails the best closed model by 14 index points; on Terminal-Bench 4.0 it leads it by 2.

What happened to SWE-bench Verified, LiveCodeBench and Aider?

They stopped updating before this year's models arrived, so they cannot rank anything released since spring 2026.

• SWE-bench: the newest leaderboard entry is dated 26 February 2026. In the maintainers' own mini-SWE-agent 2.0.0 runs of 17 February, the best open models were MiniMax M2.5 (high) at 75.8%, GLM 5 (high) at 72.8% and Kimi K2.5 (high) at 70.8%, against 76.8% for Claude 4.5 Opus (high).
• LiveCodeBench: the newest problem on the public board is from April 2025.
• Aider Polyglot: "last updated November 20, 2025"; the best open entry is DeepSeek-V3.2-Exp (Reasoner) at 74.2% against 88.0% for gpt-5 (high).

Vendors still quote SWE-bench Verified on some cards, but they run it themselves, which brings us to the self-reported numbers.

How do vendor claims compare with independent runs?

On Terminal-Bench 4.0, vendor numbers land anywhere from 12 points above to 6 points below what Artificial Analysis measured, mostly because each vendor picks its own agent and settings.

Self-reported figures from each model card, read 7 November 2026; differences are my subtraction. DeepSWE v1.1 has no independent leaderboard I could find, so that column is vendor-only.

Two newer cards report only older or vendor-run tests. Ripple-65B-Preview (dense 65B, Apache 2.0, 4 November) claims 76.5% on SWE-bench Verified and 80.5% on Terminal-Bench 2.1, both self-run. JetBrains' Mellum2.1 (12B, 2.5B active) claims 47.0% on SWE-bench Verified and 82.0 on LiveCodeBench v6. Neither has an independent result yet. Treat any score in this paragraph as a claim, not a measurement.

How fast do these models run on a DGX Spark and an RTX 5090?

On one Spark, the fast coding models are the small-active MoEs; the strongest model that fits runs at about 24 tok/s on code. Reported single-stream decode speeds; only the first row is a GPUwerk measurement:

One user, one request, short prompts unless a context is given. Speculative decoding (MTP) gains depend on how predictable the text is.

I found no published RTX 5090 or Spark measurements for Nanbeige4.3-33B, so here is the bandwidth arithmetic instead. A dense model reads all its weights for each token, so the ceiling is bandwidth divided by file size. For the 20.8 GB Q4_K_M file: 1,792 GB/s ÷ 20.8 GB ≈ 86 tok/s on an RTX 5090, and 273 GB/s ÷ 20.8 GB ≈ 13 tok/s on a Spark. Real decode lands below the ceiling. That is the trade on this hardware: by this arithmetic a 32 GB card runs mid-size dense coders about six times faster than a Spark, and the Spark holds the 100 GB-plus models a card cannot load at all. Measured RTX 5090 figures for other models are collected in the planned RTX 5090 LLM benchmarks article; GPUwerk has published no RTX 5090 measurements.

Which agent tools work with these models?

Most of them run in the same agents you would use with a closed model, because vendors now benchmark in those agents. From the cards and Artificial Analysis's runs:

• Claude Code: Z.ai's GLM-5.3 results use Claude Code 2.1.207; DeepSeek used it for Terminal-Bench 4.0; Artificial Analysis ran Step 5 Preview in it.
• Codex: MiniMax ran Terminal-Bench 4.0 in Codex 0.149.1; Artificial Analysis ran DeepSeek V4.1 Flash in Codex (47.3 on its index).
• Opencode: Artificial Analysis's best open-model agent run, GLM-5.3 at 48.9.
• Pi and mini-SWE-agent: JetBrains evaluates Mellum2.1 in Pi v0.73.1, and most DeepSWE numbers above come from mini-SWE-agent.
• Vendor CLIs: Kimi Code for Kimi, DeepSeek's in-house agent for DeepSeek, Mistral Vibe for Mistral Large 4.

Two integration details trip people up. DeepSeek-V4.1-Flash-1017 ships without a Jinja chat template; the card points to a Python reference encoder and the deepseek-recipe library, so an engine that relies on the template needs its own DeepSeek V4.1 support. Kimi K3.1 Pro requires each assistant message to be sent back with its reasoning_content and tool_calls intact, which some agent frameworks strip. Parser flags for each family are in the companion article on tool calling with open models. For connecting Continue, Cline or aider to your own endpoint, follow the coding agents guide; Cursor and Continue specifics are in the Cursor and Continue page.

Which coding model should you run?

Match the model to the memory you have; at every size there is now a reasonable answer.

• 32 GB GPU: Nanbeige4.3-33B at Q4_K_M. It has the best independent Terminal-Bench 4.0 score of anything that fits (29.8%), a 512K context and an Apache 2.0 license. Qwen3.6-27B is the fallback with a fresh-task SWE-rebench result (31.2%).
• One 128 GB Spark, quality first: GLM-5.3-Flash at UD-Q2_K_XL, MIT licensed, about 24 tok/s on code. The 2-bit build gives up some of the 32.8% measured at native precision; our GLM-5.3-Flash page covers that trade.
• One Spark, speed first: Qwen3.6-35B-A3B in NVFP4 at 86 tok/s, or the measured Qwen3-Coder-30B-A3B at 80.9. Both are weaker on hard agentic tasks (24.7% on SWE-rebench for Qwen3.6-35B-A3B), but fast enough for tight edit loops.
• Two Sparks: GLM-5.3-Flash in NVIDIA's 204.4 GB NVFP4 build, the full 4-bit model.
• A GPU server: Kimi K3.1 Pro. If its license terms do not fit, GLM-5.3 or DeepSeek-V4.1-Flash-1017 (MIT).

My pick for one developer with one Spark is GLM-5.3-Flash at 2 bits: of the models that fit one Spark, it is the only one whose full-precision version scores above 30% on Terminal-Bench 4.0. If latency matters more than hard multi-step tasks, run Qwen3.6-35B-A3B instead. Either way, test on your own repository. A rented Spark costs $0.79/hour, billed per minute (pricing), and the self-hosted assistant vs Copilot comparison covers what you give up against a hosted tool.

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 (per-evaluation scores), Artificial Analysis
• AI Coding Agent Benchmarks and Leaderboard (Coding Agent Index v2.0), Artificial Analysis
• Terminal-Bench 4.0 leaderboard, Terminal-Bench (Laude Institute), 2026-10-23
• SWE-rebench Leaderboard, Nebius
• SWE-bench Leaderboards, SWE-bench
• Aider LLM Leaderboards, aider, 2025-11-20
• LiveCodeBench Leaderboard, LiveCodeBench
• moonshotai/Kimi-K3.1-Pro model card, Hugging Face, 2026-11-05
• deepseek-ai/DeepSeek-V4.1-Flash-1017 model card, Hugging Face, 2026-10-17
• MiniMaxAI/MiniMax-M3.1-Flash model card, Hugging Face, 2026-11-05
• XiaomiMiMo/MiMo-V2.6-Flash-RL model card, Hugging Face, 2026-10-28
• Nanbeige/Nanbeige4.3-33B model card, Hugging Face, 2026-10-28
• ricursiveAI/Ripple-65B-Preview model card, Hugging Face, 2026-11-06
• JetBrains/Mellum2.1-12B-A2.5B-Thinking model card, Hugging Face, 2026-10-07
• Qwen3.8-27B-NVFP4 on a single DGX Spark: vLLM+MTP measurements, NVIDIA Developer Forums, 2026-08-15
• Qwen3.6-35B-A3B on DGX Spark, SparkBench, 2026-08-04
• GLM-5.3-Flash on one DGX Spark, ai-muninn, 2026-08-30