The honest answer to "which open-weight model should I self-host for coding?" is narrower than any leaderboard admits: on hardware a solo founder actually owns, the real choice is basically one model — Qwen3-Coder-Next (80B total, but only 3B active), which needs about 48GB of VRAM at 4-bit and posts the top open-weight SWE-bench Verified score at its release, 70.6%. The models that top the "best open-source coder" lists — GLM-5.3 at 744B, DeepSeek-V3.2-Exp at 685B, Kimi K3 at 2.8T — are open weights you rent an API for, not weights you run at home. "Open" tells you you're allowed to download it; it says nothing about whether it fits your GPU. This guide sorts the September 2026 field by that one question: what can you truly run, and what should you just pay for.

Here's the whole decision in one screen:

The reframe that saves you a weekend: sort by hardware first, benchmark second. A 90-percent model you can't load is a zero. Below, each model against the only test that matters for a team of one — will it run on what you have.

Why "self-host" has a hardware ceiling most lists ignore#

The single most useful fact about the September 2026 open coding field is that most of the good models are too big to self-host, and almost no ranking says so. GLM-5.3 is 744B parameters. DeepSeek-V3.2-Exp is 685B. Kimi K3 is 2.8 trillion. Even at 4-bit quantization, those demand multi-GPU servers — an 8×H100-class box for the DeepSeek/GLM tier, and far more for Kimi. That's a rented cluster, not a desk. We mapped what that rental actually costs in the September GPU price guide; the short version is it's an operating expense, not a one-time buy.

What makes the small Qwen coders special is ultra-sparse Mixture-of-Experts: 80B or 30B total parameters, but only 3B active per token. You pay the memory cost of holding the weights, but the compute cost of a 3B model — so an 80B model serves at a speed and latency that feel mid-size. That combination, capable-but-sparse, is the only thing that makes genuine local coding practical on prosumer hardware, and it's why the self-host answer keeps landing on Qwen. For the broader ranked field (including the models you'd rent), our September open-source coding leaderboard and the open-source LLMs for coding roundup go deeper; this piece is the hardware-first cut.

The two you can actually self-host#

Qwen3-Coder-Next (80B-A3B) — the best coder you can run yourself

This is the pick if you can reach ~48GB of VRAM. Qwen3-Coder-Next is 80B total, 3B active, Apache 2.0 (commercial use permitted), with a 256K native context. It scores 70.6% on SWE-bench Verified using the SWE-Agent scaffold — the highest reported open-weight result at its release, and roughly on par with much larger models like DeepSeek-V3.2 despite a fraction of the active size.

What it takes to run. At 4-bit, budget about 48GB of VRAM — realistically two 24GB consumer cards (2×RTX 4090 or 3090), a single 48GB workstation GPU (A6000/L40S), or an Apple Silicon Mac with 64GB+ of unified memory (hardware-corner has the detailed breakdown). It does not fit one 24GB card. But because only 3B parameters activate, throughput and latency are far better than a dense 80B, so a two-card rig or a maxed Mac gives you a genuinely usable local coding agent — no per-token bill, no data leaving your machine. If you're wiring it into an editor, the mechanics are the same as our Claude Code in VS Code setup; you're just pointing the client at a local endpoint.

Qwen3-Coder-30B-A3B-Instruct — the true single-GPU pick

If you have exactly one 24GB card, this is your model. 30B total, 3B active, Apache 2.0, and it fits in roughly 18GB at Q4_K_M — comfortable on a single RTX 4090 or 3090, and runnable on a 24GB rented card at well under a dollar an hour. The ceiling is lower than Next's; the official per-variant coding scores are published as chart images rather than a clean text figure, so treat the 30B's exact benchmark as unconfirmed and clearly below its 480B sibling. But for a solo founder who wants a capable, private, always-on coder on one GPU, it's the honest answer — and it pairs naturally with the value-card analysis in our cheapest 16GB+ VRAM guide and the local-LLM-for-coding walkthrough.

The three you should rent, not run#

DeepSeek-V3.2-Exp — cheapest capable, via API

DeepSeek-V3.2-Exp is 685B total / 37B active under an MIT license, and it introduces DeepSeek Sparse Attention to make its ~160K context cheap. Its own reported numbers are strong (LiveCodeBench in the low-70s, Aider-Polyglot ~74.5). But 685B is a server, not a desk — so this is an API play, and there it's a bargain: about $0.27–0.28 per 1M input, roughly 3 cents per 1M on a cache hit, and $0.41–0.42 per 1M output (VentureBeat, OpenRouter). For spiky or moderate volume, it's the cheapest capable coder worth defaulting to. Model your spend against our September LLM API pricing breakdown before you commit a workload.

GLM-5.3 — the best open coding quality, on a flat plan

Zhipu's GLM-5 line tops out at GLM-5.3 (744B / 40B active), Apache 2.0, with the strongest coding quality of any open-weight model by the vendor's own in-house benchmarks (GLM-5.2 already claimed Terminal-Bench 2.1 at 81.0 and SWE-bench Pro at 62.1; GLM-5.3 claims a further ~50% jump on Zhipu's internal code bench). Those are vendor claims — read them as such. At 744B it's not a self-host, but Zhipu prices access aggressively: the GLM Coding Plan runs $18/mo (Lite), $72/mo (Pro), $160/mo (Max), a flat rate that's hard to beat for heavy daily coding. We walked through wiring a GLM plan into an agent in GLM in Claude Code.

Kimi K3 — frontier scale, API only

Kimi K3 (Moonshot) is 2.8T total / 104B active, "Open Frontier Intelligence," with a 1M context and top-tier reported coding scores on the newer suites (Terminal-Bench 2.1 88.3, FrontierSWE 81.2). It ships under a custom Kimi K3 License — not MIT — so if commercial terms matter, read them before you build on it. At 2.8 trillion parameters it is not remotely self-hostable by a solo founder; it's here as the quality ceiling, reached through a hosted API.

Pick by what's on your desk#

One integrity note to carry into your own testing: the leaderboards have moved off SWE-bench Verified onto Terminal-Bench, SWE-bench Pro, DeepSWE and FrontierSWE, and the field turns over monthly (StepFun's 600B Step 5 open weights land October 15, for one). No single benchmark cell settles this — so once you've narrowed to two candidates by hardware, run your own repo's tasks through both and let that decide. The framing survives the version churn: self-host the small sparse Qwen MoEs, rent the giants.