Interlinked.

Best local LLM for coding
in 2026.

Every open-weight model ranked by SWE-bench, with thinking mode status, VRAM requirements, and hardware cost. Nobody else shows you which scores had thinking on or off.

VRAMSWE-BENCHMiniMax M2.5GLM-5Kimi K2.5*Qwen 3.5 397BCoder-NextDeepSeek V3.2

TL;DR: Key takeaways

01

Thinking mode breaks every leaderboard. Published SWE-bench scores mix models that always reason, models benchmarked with thinking off, and one that codes worse with it on, with no disclosure. Kimi K2.5's 76.8% was measured with thinking OFF, so its true ceiling is unknown.

02

MiniMax M2.5 leads the open-weight field at 80.2% SWE-bench, just 0.6 points behind Claude Opus 4.6 (80.8%), but needs ~457 GB of VRAM to run.

03

Gemma 4 26B MoE is the consumer king: 77.1% LiveCodeBench on a single ~$2K RTX 5090 (32 GB). That $2K tier captures ~95% of a $200K rig's practical coding value.

04

Scaffolding beats the model. SWE-bench Pro shows a 22+ point swing from the agent harness alone, roughly 3x the ~13 point spread across every model here.

05

Retrieval is the lever, and Interlinked Files is the scaffolding: machine-wide filename search in 85 µs and content search in ~7 to 9 ms, where ripgrep takes 93.8 s on the same 4.47M-file machine. Free forever.

Models ranked

9

every open-weight model with coding benchmarks

Thinking-handicapped

3

scores measured with thinking off or hurting

The problem

Thinking mode breaks every leaderboard.

When MiniMax M2.5 reports 80.2% on SWE-bench, that score includes interleaved thinking, a multi-step reasoning pass that runs on every single token. There is no way to turn it off.

When Kimi K2.5 reports 76.8% on SWE-bench, that score was measured in instant mode, thinking explicitly turned off. The 3.4-point gap might not exist.

And then there's DeepSeek V3.2, which is actually worse at coding with thinking enabled: 50.7% vs 56.1% on LiveCodeBench. The published SWE-bench score uses non-thinking mode.

Every existing leaderboard compares these numbers side by side without telling you any of this.

Thinking mode status

What each model was actually running.

ModelThinking modeStatus in benchmarksImpact
Claude Opus 4.6Adaptive (always on)Always on, model decides how much to reasonBaked in, no off switch
MiniMax M2.5Always on (no off switch)Always on, interleaved thinking is architecturalUnfair advantage in comparisons
Kimi K2.5Toggle (instant vs thinking)INSTANT MODE (thinking OFF) for SWE-bench, AIME, GPQAHandicapped, real score likely higher
GLM-5Always onAlways on, reasoning integrated into architectureBaked in
DeepSeek V3.2ToggleThinking HURTS coding (50.7 vs 56.1 non-thinking)Published coding scores use non-thinking
Gemma 4 26B / 31BToggleThinking ON for all published benchmarksFair
Qwen 3.5ToggleThinking ON for published benchmarksFair

Highlighted rows are the ones where thinking mode changes the interpretation of the score. Kimi's published SWE-bench score is with thinking OFF: its real capability is unknown.

Full coding leaderboard

Open-weight models ranked by SWE-bench.

Claude Opus 4.6 included as an unranked API reference. Highlighted rows indicate thinking mode status changes the interpretation.

RankModelReleasedThink?SWE-benchLiveCodeBenchVRAMNotes
unrankedClaude Opus 4.6Mar 2026Adaptive80.8%~80%API only#1 overall. Not runnable offline
1MiniMax M2.5Feb 2026Always on80.2%not published~457 GBOnly 0.6% behind Opus. GPQA only 47%
2GLM-5Feb 2026Always on77.8%52.0%*~860 GB*LCB score suspicious
3Kimi K2.5Jan 2026OFF (instant)76.8%85.0%~600 GBThinking was OFF. Real score likely higher
4Qwen 3.5 397BFeb 2026On76.4%83.6%~200 GBStrong all-rounder
5Qwen3-Coder-Next2026On70.6%not published~45 GB80B MoE, 3B active. Best efficiency ratio
6DeepSeek V3.2Dec 2025OFF (for code)67.8%74.1%~400 GBThinking hurts its coding
unrankedGemma 4 31B DenseApr 2026Onnot published80.0%31 GBNo SWE-bench published
unrankedGemma 4 26B MoEApr 2026Onnot published77.1%26 GBNo SWE-bench published. Consumer king

Kimi K2.5 holds the LiveCodeBench crown at 85.0% despite being benchmarked with thinking off. MiniMax M2.5 leads SWE-bench but has no LiveCodeBench score published. Google didn't publish SWE-bench for Gemma 4.

Consumer hardware

What runs on your GPU.

32 GB of VRAM. That's what a single RTX 5090 gives you. Q8 quantization or better: aggressive quants destroy quality.

ModelReleasedArchitectureVRAM (Q8)CodingELOCost
Gemma 4 26B MoEApr 2026MoE (3.8B active)26 GB77.1% LCB1441~$2K (1x RTX 5090)
Gemma 4 31B DenseApr 2026Dense (31B active)31 GB80.0% LCB1452Won't fit at Q8: no KV cache room
Qwen 3.5 27BFeb 2026Dense~27 GBnot published~1430No coding benchmarks published
Mistral Small 3.1 24BMar 2025Dense~24 GBnot published~1380Over a year old, -61 ELO

Gemma 4 26B MoE is the undisputed consumer king. 128 experts, 2 active, a 1.6% activation ratio nobody else has achieved at this scale. Fits at Q8 with 6 GB of headroom for KV cache and CUDA context.

Diminishing returns

Hardware cost vs coding quality.

HardwareVRAMCostBest modelCoding
1x RTX 509032 GB~$2KGemma 4 26B MoELCB 77.1%, ELO 1441
2x RTX 509064 GB~$4KQwen 3.5 72BELO ~1460, no coding bench
Workstation200 GB$15-30KQwen 3.5 397B MoESWE 76.4%, LCB 83.6%
4x H100 rack~460 GB$100-120KMiniMax M2.5SWE 80.2% (coding only)
8x H100 rack~600 GB$180-240KKimi K2.5SWE 76.8%+, LCB 85.0%

$2K tier

95%

of the $200K tier's practical value

$2K → $200K

+8 pts

100x the cost for 8% improvement on SWE-bench

First real step up

$15K

Qwen 3.5 397B, SWE-bench 76.4%

Scaffolding swings 22+ points. Model choice swings ~13.

What wraps the model matters 3x more than which model.

The real insight

Scaffolding matters more than the model.

SWE-bench Pro shows a 22+ point swing between basic and optimized agent scaffolds using the exact same model. That's 3x the gap between the best and worst models on this entire leaderboard.

The ceiling isn't the model. The ceiling is what's wrapped around the model: how the agent navigates repositories, retrieves context, breaks down tasks, and feeds relevant code back into the prompt. A mediocre model with great retrieval beats a frontier model with no context.

That's what Interlinked Files provides. Machine-wide file search in microseconds. Content search across every code file on the disk. The agent scaffolding that turns any model (whether it's Gemma 4 running locally or Claude Opus 4.6 through an API) into a better coding assistant.

Model gap

~13 pts

Best vs worst on SWE-bench

Scaffold gap

22+ pts

Same model, different agent

Implication

3x

Scaffolding impact vs model choice

Common questions

Local LLM coding, answered.

What is the best local LLM for coding in 2026?

For a single-GPU rig, Gemma 4 26B MoE: 77.1% on LiveCodeBench, running at Q8 on one ~$2K RTX 5090 (32 GB). With workstation VRAM, Qwen 3.5 397B (76.4% SWE-bench) and MiniMax M2.5 (80.2%) rank higher, but the $2K tier already captures ~95% of a $200K rig's practical value.

Can a local model match Claude Opus 4.6?

On raw benchmarks, nearly: MiniMax M2.5 tops the open-weight field at 80.2% SWE-bench, only 0.6 points behind Opus's 80.8%, but it needs ~457 GB of VRAM. On consumer hardware you trade a few points for running fully offline, with no API keys and no per-token cost.

Why do local LLM coding leaderboards disagree?

Thinking mode. MiniMax and GLM-5 always reason, Kimi K2.5 was benchmarked with thinking OFF, and DeepSeek V3.2 actually codes worse with thinking on (50.7% vs 56.1%). Most leaderboards line these numbers up without saying which mode produced them.

Does the model or the agent scaffolding matter more?

The scaffolding. SWE-bench Pro shows a 22+ point swing from the agent harness alone, about 3x the ~13 point gap across the entire model field. How the agent retrieves context and navigates the repo moves the score more than swapping models.

How does Interlinked Files make a local model code better?

It hands the agent instant, machine-wide search instead of blind grep crawls: filename lookups in 85 µs and content search in ~7 to 9 ms, where ripgrep takes 93.8 s on the same 4.47M-file machine. A single 16 ms indexed query can replace a 6m57s, 71-call search, cutting retrieval from ~58% of the context budget toward ~0%. Everything local is free, forever: the app, whole-machine search, MCP. Hosting starts at $5.99/mo.

Benchmark scores from published papers, official model cards, and Arena Leaderboard as of April 2026. Thinking mode status verified from official documentation. Hardware costs approximate retail/cloud.