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.
TL;DR: Key takeaways
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.
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.
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.
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.
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
every open-weight model with coding benchmarks
Thinking-handicapped
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.
| Model | Thinking mode | Status in benchmarks | Impact |
|---|---|---|---|
| Claude Opus 4.6 | Adaptive (always on) | Always on, model decides how much to reason | Baked in, no off switch |
| MiniMax M2.5 | Always on (no off switch) | Always on, interleaved thinking is architectural | Unfair advantage in comparisons |
| Kimi K2.5 | Toggle (instant vs thinking) | INSTANT MODE (thinking OFF) for SWE-bench, AIME, GPQA | Handicapped, real score likely higher |
| GLM-5 | Always on | Always on, reasoning integrated into architecture | Baked in |
| DeepSeek V3.2 | Toggle | Thinking HURTS coding (50.7 vs 56.1 non-thinking) | Published coding scores use non-thinking |
| Gemma 4 26B / 31B | Toggle | Thinking ON for all published benchmarks | Fair |
| Qwen 3.5 | Toggle | Thinking ON for published benchmarks | Fair |
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.
| Rank | Model | Released | Think? | SWE-bench | LiveCodeBench | VRAM | Notes |
|---|---|---|---|---|---|---|---|
| unranked | Claude Opus 4.6 | Mar 2026 | Adaptive | 80.8% | ~80% | API only | #1 overall. Not runnable offline |
| 1 | MiniMax M2.5 | Feb 2026 | Always on | 80.2% | not published | ~457 GB | Only 0.6% behind Opus. GPQA only 47% |
| 2 | GLM-5 | Feb 2026 | Always on | 77.8% | 52.0%* | ~860 GB | *LCB score suspicious |
| 3 | Kimi K2.5 | Jan 2026 | OFF (instant) | 76.8% | 85.0% | ~600 GB | Thinking was OFF. Real score likely higher |
| 4 | Qwen 3.5 397B | Feb 2026 | On | 76.4% | 83.6% | ~200 GB | Strong all-rounder |
| 5 | Qwen3-Coder-Next | 2026 | On | 70.6% | not published | ~45 GB | 80B MoE, 3B active. Best efficiency ratio |
| 6 | DeepSeek V3.2 | Dec 2025 | OFF (for code) | 67.8% | 74.1% | ~400 GB | Thinking hurts its coding |
| unranked | Gemma 4 31B Dense | Apr 2026 | On | not published | 80.0% | 31 GB | No SWE-bench published |
| unranked | Gemma 4 26B MoE | Apr 2026 | On | not published | 77.1% | 26 GB | No 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.
| Model | Released | Architecture | VRAM (Q8) | Coding | ELO | Cost |
|---|---|---|---|---|---|---|
| Gemma 4 26B MoE | Apr 2026 | MoE (3.8B active) | 26 GB | 77.1% LCB | 1441 | ~$2K (1x RTX 5090) |
| Gemma 4 31B Dense | Apr 2026 | Dense (31B active) | 31 GB | 80.0% LCB | 1452 | Won't fit at Q8: no KV cache room |
| Qwen 3.5 27B | Feb 2026 | Dense | ~27 GB | not published | ~1430 | No coding benchmarks published |
| Mistral Small 3.1 24B | Mar 2025 | Dense | ~24 GB | not published | ~1380 | Over 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.
| Hardware | VRAM | Cost | Best model | Coding |
|---|---|---|---|---|
| 1x RTX 5090 | 32 GB | ~$2K | Gemma 4 26B MoE | LCB 77.1%, ELO 1441 |
| 2x RTX 5090 | 64 GB | ~$4K | Qwen 3.5 72B | ELO ~1460, no coding bench |
| Workstation | 200 GB | $15-30K | Qwen 3.5 397B MoE | SWE 76.4%, LCB 83.6% |
| 4x H100 rack | ~460 GB | $100-120K | MiniMax M2.5 | SWE 80.2% (coding only) |
| 8x H100 rack | ~600 GB | $180-240K | Kimi K2.5 | SWE 76.8%+, LCB 85.0% |
$2K tier
of the $200K tier's practical value
$2K → $200K
100x the cost for 8% improvement on SWE-bench
First real step up
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
Best vs worst on SWE-bench
Scaffold gap
Same model, different agent
Implication
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.