Software teams

Coding agents, code review and internal tools, often with a rule that source code stays in-house.

A lens over 4 jobs. Draft copy, under review; the models and numbers on each job page come from today's data.

The jobs it bundles

Each links to its job page (or the nearest scoreboard view while the page doesn't exist yet).

Write, fix and review code with an agent in the loop.

42 models fit

Drive CI, tickets and internal APIs.

284 models fit

Read a whole repository at once.

238 models fit

Keep the code in-house on one GPU.

313 models fit

Each constraint opens a job page or the scoreboard with the filter already in the URL.

Code stays in-house

If source code can't go to a hosted API, the model has to fit the GPU you own.

Coding models at 24 GB → · At 64 GB (Mac) →
License gates

Shipping a model inside a product is simpler under a permissive license.

Permissive coding models →
Audit trail

Prefer models whose behaviour we checked on a live endpoint, with the probe, the reply and the date on file.

Models with first-look receipts →

How people wire a model into this job. Notes are written by hand from the cited cards and docs; the model links go to the facts.

Model: moonshotai/Kimi-K3
Harness: Any OpenAI-compatible agent (OpenCode, Aider, Claude Code via proxy)
Weights: Provider-hosted (native MXFP4 weights)

Point the agent at the provider's /v1 base URL and the provider's model ref from the where-to-run table.
Set max context per the provider row, not the config value; providers cap below the model maximum.
Measure cost per completed task on your own repo before committing; the model's own eval files are the lab's numbers, not ours.

Model: Qwen/Qwen3-Coder-30B-A3B-Instruct
Harness: vLLM or SGLang serving + Aider/OpenCode
Weights: FP8 (fits a single 80GB GPU with headroom; see hardware-fit estimate)

Architecture is registered in both vLLM and SGLang (see servability chips).
Enable tool calling in the server flags; check the model card's chat template notes.