Use case · Coding
Best local LLM for coding, and the hardware to run it
For local coding, Qwen3.6 35B-A3B is the fast pick: ~24 GB at 4-bit, so it runs on a GMKtec EVO-X2 64GB ($2,200) at ~72 tok/s by our estimate. For bigger models such as Qwen3.5 122B-A10B (~79 GB), the cheapest tracked rig is the GMKtec EVO-X2 128GB at $3,500.
Coding models and the cheapest hardware for each
| Coding model | 4-bit memory | Cheapest rig that fits | Cheapest complete computer | Fastest tracked rig |
|---|---|---|---|---|
| Qwen3.8 27Bcoding, vision, general · 27B | ~19 GB | GMKtec EVO-X2 64GB$2,200 · ~12 tok/s | same | NVIDIA GeForce RTX 5090 Founders Edition$8,500 · ~83 tok/s |
| GLM-4.7-Flash 30B-A3Bcoding agents · 30B (3B active) | ~21 GB | GMKtec EVO-X2 64GB$2,200 · ~72 tok/s | same | NVIDIA GeForce RTX 5090 Founders Edition$8,500 · ~269 tok/s |
| Qwen3 30B-A3Bfast chat & coding · 30.5B (3.3B active) | ~21 GB | GMKtec EVO-X2 64GB$2,200 · ~67 tok/s | same | NVIDIA GeForce RTX 5090 Founders Edition$8,500 · ~260 tok/s |
| Qwen3 32Bgeneral, coding, reasoning · 32.8B | ~22 GB | GMKtec EVO-X2 64GB$2,200 · ~10 tok/s | same | NVIDIA GeForce RTX 5090 Founders Edition$8,500 · ~70 tok/s |
| Qwen3.6 35B-A3Bfast coding & chat · 35B (3B active) | ~24 GB | GMKtec EVO-X2 64GB$2,200 · ~72 tok/s | same | NVIDIA GeForce RTX 5090 Founders Edition$8,500 · ~269 tok/s |
| Mistral Small 4 119B-A6Bchat, reasoning, coding, vision · 119.4B (6.5B active) | ~78 GB | GMKtec EVO-X2 128GB$3,500 · ~41 tok/s | same | PNY NVIDIA RTX PRO 6000 Blackwell$17,899 · ~194 tok/s |
| Qwen3.5 122B-A10Bstrong general / coding · 122B (10B active) | ~79 GB | GMKtec EVO-X2 128GB$3,500 · ~29 tok/s | same | PNY NVIDIA RTX PRO 6000 Blackwell$17,899 · ~152 tok/s |
Estimated single-user decode at ~8K context; GPU-card figures assume the model is fully in VRAM. Models tagged for coding in our catalog; we don’t rank code quality, so check current coding benchmarks for that. How we estimate
Too big for any single tracked rig at 4-bit: Muse Spark Open 816B-A42B (~521 GB)
Dense or MoE for coding?
Mixture-of-experts models read only their active weights per token, so they generate much faster on the same hardware. On the GMKtec EVO-X2 64GB ($2,200), we estimate Qwen3.8 27B (dense) at ~12 tok/s and Qwen3.6 35B-A3B (3B active) at ~72 tok/s. Agentic coding tools generate many tokens per task, so that gap matters.
Coding agents also fill long contexts with files and diffs. Our fit assumes ~8K tokens; plan extra memory headroom for 32K–128K sessions (see the VRAM calculator).
Measured speeds on coding models
Public single-user decode results we calibrate against. Runs differ in build and settings, so compare like with like.
| Hardware | Model / run | Measured tok/s | Our formula | Source |
|---|---|---|---|---|
| DGX Spark273 GB/s | Qwen2.5-Coder 7B Q8_0llama.cpp tg32 | 29.4 | 26.0 | source |
| DGX Spark273 GB/s | Qwen3-Coder 30B-A3B Q8_0llama.cpp tg32 | 61.1 | 49.8 | source |
| DGX Spark273 GB/s | GLM-4.7-Flash Q8_0llama.cpp tg32 | 48.7 | 53.9 | source |
| DGX Spark273 GB/s | Qwen3.5-122B-A10B Q5_K_Sllama.cpp tg128 | 23.0 | 27.7 | source |
| Strix Halo256 GB/s | Qwen3 30B-A3B Q4_K_Mllama.cpp Vulkan tg128 | 85.5 | 67.3 | source |
| Strix Halo256 GB/s | Qwen3 30B-A3B UD-Q4_K_XLllama.cpp Vulkan tg128 | 75.3 | 67.3 | source |
| Strix Halo256 GB/s | Qwen3-Coder-Next 80B-A3B Q4_K_Mllama.cpp Vulkan tg128 | 57.5 | 71.5 | source |
| M5 Max 40c614 GB/s | Qwen3.6-27B Q8_0llama.cpp decode@128 | 18.0 | 18.9 | source |
| M5 Max 40c614 GB/s | Qwen3.6-35B-A3B Q8_0llama.cpp decode@128 | 93.0 | 90.3 | source |
| RTX 50901,792 GB/s | Qwen3.6-27B Q4_K_Mllama.cpp tg128 | 77.1 | 81.3 | source |
| RTX 50901,792 GB/s | Qwen3.6-35B-A3B UD-Q4_K_Mllama.cpp tg128 | 271.0 | 268.6 | source |
| RTX 40901,008 GB/s | Qwen3 32B Q4_K_Mllama.cpp tg128 | 42.0 | 40.4 | source |
| RTX 40901,008 GB/s | Qwen3 30B-A3B Q4_K_Mllama.cpp tg128 (fusion off) | 198.4 | 204.4 | source |
| Strix Halo256 GB/s | Qwen3.6 35B-A3B Q4_K_MOllama/llama.cpp | 63.0 | 71.5 | source |
| RTX 50901,792 GB/s | Qwen3 30B-A3B Q4_K_Mllama.cpp + GGML_CUDA_GRAPH_OPT=1 | 352.1 | 260.1 | source |
Frequently asked questions
What is the best local LLM for coding?
It depends on your memory. Qwen3.6 35B-A3B is a fast MoE coding model that needs ~24 GB at 4-bit; Qwen3.8 27B is a dense alternative at ~19 GB; Qwen3.5 122B-A10B needs ~79 GB. We track hardware and speed, not code quality, so compare current coding benchmarks too.
What hardware do I need to run a coding model locally?
For Qwen3.6 35B-A3B, ~24 GB of VRAM or unified memory: the cheapest tracked option is the GMKtec EVO-X2 64GB at $2,200.
Can I run a coding model on a 24GB GPU?
Yes for Qwen3.8 27B, GLM-4.7-Flash 30B-A3B, Qwen3 30B-A3B at 4-bit with ~8K context. Longer contexts need more headroom.
Why are MoE models faster for coding?
Decode speed depends on how many bytes are read per token. A MoE model reads only its active experts, so a 35B model with 3B active generates far faster than a dense 27B model on the same hardware.
Can I run Muse Spark Open 816B-A42B locally?
Not on any single machine we track: it needs ~521 GB at 4-bit, more than the largest tracked rig's usable memory.