There is a single question that decides which open-weight coding model you actually run, and it is almost never "which one scores highest." It's this: what is the smallest, cheapest box that runs this thing without turning deployment into a science project? For most of 2026 the honest answer for anything near frontier quality was a quantization project on an 80GB card, or a multi-GPU node, or a hosted API. Cohere's North Mini Code is the model that makes the answer "one H100, FP8, ship it."
North Mini Code 1.0 landed on June 9 under Apache 2.0 — Cohere's first developer model and the debut of its North family. It is a 30-billion-parameter sparse Mixture-of-Experts that activates just 3B parameters per token (128 experts, 8 fire per token, with sliding-window and global attention interleaved 3:1). On SWE-bench Verified it posts 67.6%, and it clocks 33.4 on the Artificial Analysis Coding Index. Those are frontier-adjacent numbers from a model a third the total size of Qwen3-Coder-Next and a twenty-fifth the size of GLM-5.2.
The benchmark spread is noise. The deployment spread is the story.#
Line the three credible single-GPU-or-cheaper open coders up by score and they're all in the same class:
- GLM-5.2 — 74.4% on FrontierSWE (within a point of Claude Opus 4.8)
- Qwen3-Coder-Next — ~70.6% on SWE-bench Verified
- North Mini Code — 67.6% on SWE-bench Verified
A seven-point spread across different benchmark harnesses, on tasks with real run-to-run variance, is not what decides your architecture. If you swapped these three under a production agent for a week, the thing you'd notice is not the SWE-bench delta. It's the bill and the box.
All three are frontier-adjacent coders. Only one of them fits a single H100 at native precision with no quantization step. That's the number that changes what you ship.
Here's the box, which is where North separates:
| Model | Total / active | Fits where | Quant to fit? |
|---|---|---|---|
| North Mini Code | 30B / 3B | 1× H100, FP8 | No — FP8 ships native |
| Qwen3-Coder-Next | 80B / 3B | 1× 80GB card | Yes — ~48.8GB at Q4_K_M |
| GLM-5.2 | 753B / 40B | ~8-GPU H200 node | Rent the hosted API |
Qwen3-Coder-Next also fits one card, and we've made that exact case — but you get there by running a Q4_K_M GGUF at ~48.8GB, which means picking a quant, validating it didn't cost you accuracy, and living with the quirks of a quantized MoE. GLM-5.2 is a different weight class entirely; full precision is roughly an 8-GPU H200 node, so for most teams the real deployment is the hosted API at $1.40/M input. North's 30B total is small enough that Cohere just ships an FP8 checkpoint that runs on one H100, no quant tuning, and — the detail that matters for an agent — vLLM-native tool-call and reasoning parsing built in.
Running it: one command, one card#
The FP8 build is designed for vLLM and is not loadable through plain transformers (the quantization format is vLLM-specific). The serve command Cohere ships is a single line:
vllm serve CohereLabs/North-Mini-Code-1.0-fp8 \
-tp 1 \
--max-model-len 320000 \
--tool-call-parser cohere_command4 \
--reasoning-parser cohere_command4 \
--enable-auto-tool-choice \
--moe-backend triton
Note -tp 1: tensor parallelism of one — a single card. The --tool-call-parser and --reasoning-parser cohere_command4 flags are what make it drop straight into an agent loop; you get structured tool calls and interleaved reasoning without writing your own parser. If you're on a smaller card or the edge, Cohere also publishes a North-Mini-Code-1.0-w4a16 build that trims memory further with, per Cohere, negligible quality loss over BF16.
That's the whole setup. No node, no NVLink topology to reason about, no quant to bless.
Where North loses — and it does lose#
This is a decision piece, not a coronation. North Mini Code gives up real things:
- Raw score. 67.6% is the lowest of the three. If your workload is genuinely at the edge of what these models can do — gnarly multi-file refactors, long-horizon debugging where every point compounds — Qwen's extra ~3 points and GLM's extra ~7 are worth wanting, and you should run your own eval on your actual repo before you trust any of these numbers.
- Context reach. 256K is plenty for most repos, but GLM-5.2's 1M and Qwen's YaRN-extended ~1M matter if your agent routinely swallows a monorepo in one shot.
- Ecosystem maturity. GLM-5.2 and the Qwen coders have months of community quants, harness integrations, and routing recipes. North is newer; you're earlier on that curve.
The decision, in one line each#
- North Mini Code — you have (or can rent) one H100, you want the coding agent self-hosted with the least possible serving drama, and you'll trade a few benchmark points for FP8-native, no-quant, tool-calling-included deployment. This is the low-ops single-GPU default now.
- Qwen3-Coder-Next — you want a couple more points, you're comfortable running a GGUF/quant, and your card is anywhere from a 24GB 4090 (with offload) to an 80GB H100.
- GLM-5.2 — you want the top open-weight score, your volume is low enough that per-token pricing beats owning hardware, or you'd simply rather not operate an inference box at all.
All three are Apache/MIT open weights, so all three end the per-token bill the moment you self-host. The only real question left is how much machine that costs you — and North Mini Code just set the floor at one card.



