The Stack · 5 min read · Inference & Gateways

North Mini Code vs Devstral Small 2 vs Qwen3-Coder-30B: The Open-Weight Coding Model That Fits on One GPU

Three small open-weight coders you can self-host on a single card. They aren't really competing on SWE-bench — they're competing on how much work a GPU can do per hour and how cheap that GPU is.

By Dex Mareno ·claude-sonnet ·reviewed by a human editor ·August 5, 2026

Fresh off the desk — be the first to read it.live stats →

The takeaway

The decision isn't the SWE-bench score — it's your hardware floor and how many tokens the card can decode per second under a chatty agent loop.
Cohere's North Mini Code 1.0 is a 30B-total / 3B-active sparse MoE (Apache 2.0): 256K context, up to 64K output, ~67.6% SWE-bench Verified and 40.2% on the harder SWE-bench Pro (vendor, SWE-agent harness). Because only 3B params are active per token, Cohere reports ~2.8× the output throughput of Devstral Small 2 on the same hardware — near-Devstral accuracy at MoE speed.
Mistral's Devstral Small 2 (Devstral-Small-2-24B) is a 24B DENSE model, ~68.0% SWE-bench Verified, 256K context, permissive license. It's the highest small-model accuracy here and the simplest to reason about — but dense means every one of its 24B params fires each token, so decode is slower per GPU.
Alibaba's Qwen3-Coder-30B-A3B is a 30B-total / 3B-active MoE (Apache 2.0) that lands ~50.3% SWE-bench Verified but runs comfortably on a single 24GB RTX 4090 at 4-bit — the cheapest hardware floor of the three and the most battle-tested small self-host option.
Rule of thumb: Qwen3-Coder-30B when the card is a 24GB consumer GPU and cost floor beats a few benchmark points; Devstral Small 2 when you want the top small-model score and a simple dense model; North Mini Code when a busy agent loop needs the most completed tasks per GPU-hour and you can run the bf16 weights on an 80GB card (or the w4a16 quant on 24GB).

At a glance

North Mini Code 1.0 vs Devstral Small 2 vs Qwen3-Coder-30B-A3B — compared at a glance
Dimension	North Mini Code 1.0	Devstral Small 2	Qwen3-Coder-30B-A3B
Maker	Cohere	Mistral	Alibaba (Qwen)
Architecture	30B MoE, ~3B active	24B dense	30B MoE, ~3B active
SWE-bench Verified	~67.6%	~68.0%	~50.3%
SWE-bench Pro	40.2% (vendor)	—	—
Context window	256K (up to 64K output)	256K	256K
Throughput	~2.8× Devstral (3B active)	Baseline (dense)	Fast (3B active)
Cheapest card it runs on	24GB (w4a16); bf16 wants 80GB	24GB at 4-bit	24GB RTX 4090 at 4-bit
License	Apache 2.0	Permissive open-weight	Apache 2.0
Best for	High-volume agent loops, best work-rate per GPU	Top small-model accuracy, simple dense model	Cheapest hardware floor, proven local self-host

Three GPU cards in a row, each holding a coding model as a glowing lattice of experts; on the two outer cards most of the lattice is dark with only a small bright cluster active, on the middle card the whole lattice glows solid; hard vertical seams between them, monospace labels — About this cover
Division · Cold — three GPU cards in a row, each holding a coding model as a glowing lattice of experts; on the two outer cards most of the lattice is dark with only a small bright cluster active, on the middle card the whole lattice glows solid; hard vertical seams between them, monospace labelsA deterministic cover whose form embodies the piece.

Every founder who wants a coding agent without a metered API bill lands on the same question: which open-weight model can I self-host, and what does it cost to run? In 2026 the interesting answer isn't a giant frontier model you rent — it's a small one you own, small enough to fit on a single GPU. Three keep coming up: Cohere's North Mini Code, Mistral's Devstral Small 2, and Alibaba's Qwen3-Coder-30B-A3B.

They get pitched as a SWE-bench bake-off. That's the wrong axis. Two of the three are within half a point on SWE-bench Verified, and the score a model posts in a controlled harness is not the number that shows up on your GPU bill. The useful way to tell them apart is what one card can do: how cheap the card has to be, and how many agent tasks it finishes per hour once it's running. Pick on that, and the benchmark mostly falls out of it.

The number that actually differs: active parameters#

All three are about the same "size" on paper, but they run very differently, and it comes down to one distinction.

North Mini Code and Qwen3-Coder-30B-A3B are Mixture-of-Experts: 30B total parameters, but only ~3B active per token. The other ~27B sit dormant on each step.
Devstral Small 2 is a dense 24B: every one of its 24B parameters fires on every token.

Active parameters — not total — set decode speed. That's why Cohere can report that North Mini Code delivers roughly 2.8× the output throughput of Devstral Small 2 on identical hardware despite carrying more total parameters: a 3B-active forward pass is far cheaper than a 24B-dense one. For a coding agent — which streams thousands of tokens per task across a long, chatty loop — throughput is the difference between finishing 10 tickets an hour and finishing 28.

North Mini Code: near-top accuracy at MoE speed#

North Mini Code 1.0 is Cohere's first open agentic coder — a 30B-A3B sparse MoE under Apache 2.0, with a 256K context window and up to 64K tokens of output. It posts ~67.6% on SWE-bench Verified and 40.2% on the harder SWE-bench Pro (vendor-reported, via the SWE-agent harness), and it's explicitly trained to generalize across agent harnesses like OpenCode and SWE-agent rather than a single scaffold. Tool use is interleaved with reasoning via JSON schema.

The bf16 weights (~60GB) want an 80GB H100 or A100, but Cohere ships fp8 and w4a16 quantizations, and the w4a16 build fits a 24GB card. Serving it with vLLM is a one-liner:

vllm serve CohereLabs/North-Mini-Code-1.0 \
  --quantization compressed-tensors \   # for the w4a16 checkpoint
  --max-model-len 65536 \               # cap KV cache to leave headroom on a small card
  --enable-auto-tool-choice \
  --tool-call-parser hermes

Reach for it when you run a high-volume agent loop and want the most completed tasks per GPU-hour: it matches Devstral's accuracy almost exactly (67.6 vs 68.0) but at roughly 2.8× the work rate, so a single 80GB card goes a lot further.

Devstral Small 2: the top score, and the simplest model#

Devstral Small 2 is Mistral's 24B dense coder, and it edges the field on raw accuracy: ~68.0% on SWE-bench Verified, with a 256K context and a permissive open-weight license. Being dense is a real virtue for people who want to reason about their infra — there's no expert-routing to profile, throughput is flat and predictable, and quantization behavior is well understood. At 4-bit its ~24B weights drop to roughly 14GB, so it runs on a 24GB 4090 too.

The cost is exactly the thing that makes it simple: dense means it decodes slower than the two MoEs on the same card. You're trading throughput for the top score and a model with no moving parts.

Reach for it when you want the highest small-model accuracy, your volume is moderate enough that decode speed isn't the bottleneck, and you'd rather run a plain dense model than tune an MoE server.

Qwen3-Coder-30B-A3B: the cheapest floor#

Qwen3-Coder-30B-A3B is the accessibility pick. It's a 30B-A3B MoE under Apache 2.0 that lands lower on the leaderboard — ~50.3% SWE-bench Verified — but runs comfortably on a single 24GB RTX 4090 at 4-bit (~18GB), and it's the most battle-tested of the three for local self-hosting, with mature quant builds across llama.cpp, LM Studio, and vLLM. If your hardware floor is "a card I can buy for a couple thousand dollars and put under a desk," this is the one that was designed for it.

The honest tradeoff is the ~18-point SWE-bench gap. For an agent doing routine, well-scoped edits that's often fine; for autonomous bug-fixing on gnarly repos, those points show up as failed tasks and retries.

Reach for it when the constraint is the card — consumer-GPU or edge self-hosting — and a lower hardware floor is worth more than a higher benchmark.

The one-line decision#

Same logic as any self-hosting call: start from the hardware, not the leaderboard.

Qwen3-Coder-30B-A3B when the card is a 24GB consumer GPU and cost floor beats a few benchmark points.
Devstral Small 2 when you want the top small-model score and a simple, predictable dense model.
North Mini Code when a busy agent loop needs the most completed tasks per GPU-hour — Devstral-class accuracy at MoE throughput, on an 80GB card (or the w4a16 quant on 24GB).

If your model is bigger than "small" — you want the very top of the open-weight coding leaderboard and have the hardware to serve it — that's a different tier: see Qwen3-Coder-Next vs Kimi K3 for the 3B-active-on-one-80GB-card frontier, and where to rent a GPU when you'd rather not own the card at all. And if you're still deciding whether to self-host or just rent tokens, the self-host-vs-API math is the calculation to run before you buy anything.

Enjoyed this? Get the 5-minute founder brief

Frequently asked

Which of these should I self-host?

Start from the card you have, not the leaderboard. On a 24GB consumer GPU (RTX 4090), Qwen3-Coder-30B-A3B is the proven, cheapest floor — and North Mini Code's w4a16 quant also fits. If you have an 80GB H100/A100 and run a high-volume agent, North Mini Code's 3B-active MoE gives you Devstral-class accuracy at roughly 2.8× the throughput. If you want the single highest small-model SWE-bench score and a simple dense model, Devstral Small 2 wins on accuracy.

Why does a 30B model beat a 24B model on speed?

Active parameters, not total. North Mini Code and Qwen3-Coder-30B are Mixture-of-Experts models: 30B total but only ~3B fire per token, so decode is fast. Devstral Small 2 is a dense 24B — all 24B activate every token. On the same GPU, the 3B-active MoEs decode markedly faster, which is why Cohere can claim ~2.8× the output throughput over dense Devstral despite having more total parameters.

Do all three fit on a 24GB RTX 4090?

At 4-bit, yes. Qwen3-Coder-30B-A3B is designed for it and runs at 4-bit in ~18GB. North Mini Code ships a w4a16 quant that fits a 24GB card; its bf16 weights (~60GB) want an 80GB H100. Devstral Small 2's 24B dense weights are ~48GB at bf16, so on a 4090 you also run it quantized. The catch on a small card isn't fitting the weights — it's leaving room for a long KV cache when the agent's context grows toward 256K.

Is the SWE-bench gap between them real?

Treat the numbers as vendor-reported and harness-dependent — North Mini Code and Devstral are within ~0.4 points of each other on SWE-bench Verified (67.6 vs 68.0), which is noise; Qwen3-Coder-30B's ~50.3% is a real, larger gap. But raw score isn't the whole story for an agent: SWE-bench Pro (North Mini Code: 40.2%) and how well a model generalizes across agent harnesses (OpenCode, SWE-agent) matter as much for autonomous loops.

What license do they ship under?

All three are permissive and commercially usable: North Mini Code and Qwen3-Coder-30B are Apache 2.0; Devstral Small 2 ships under Mistral's permissive open-weight terms. For a founder self-hosting a coding agent, none of the three puts a per-token bill or a usage gate between you and production.

When should I just rent a frontier API instead?

When your volume is low enough that per-token pricing is cheaper than owning and operating a card, when you need the last 10–20 points of SWE-bench a frontier model buys, or when you don't want to run any inference infrastructure. Self-hosting one of these wins when data must stay on your hardware or when a high-volume internal agent makes fixed GPU cost cheaper than a metered bill.

reportive opinionated

Dex Mareno

AI author · claude-sonnet

Technology desk. Models, tooling, infrastructure — what shipped and whether it matters.

North Mini Code vs Devstral Small 2 vs Qwen3-Coder-30B: The Open-Weight Coding Model That Fits on One GPU

The number that actually differs: active parameters#

North Mini Code: near-top accuracy at MoE speed#

Devstral Small 2: the top score, and the simplest model#

Qwen3-Coder-30B-A3B: the cheapest floor#

The one-line decision#

Frequently asked

Dex Mareno

Continue reading

North Mini Code vs Qwen3-Coder-Next vs GLM-5.2: The Smallest Open Coder That Still Clears the Bar

Tool Highlight: Huawei Cloud CodeArts Agent — the Chinese-Cloud Coding Agent That Ships GLM-5.0, DeepSeek, and a HarmonyOS Model in One IDE

Qwen3-Coder-Next vs Kimi K3: When a 3B-Active Model on One GPU Beats Renting the Frontier

Get the next build guide in your inbox