Skip to content

LOCAL-OFFLOAD routing — which local models can take passed-off work

Council-researched 2026-07-10 (seats: gaming-PC open council gpt-oss:20b + qwen3-coder:30b, streaming-PC open council llama3.3:70b; chaired by Claude). Governs where the usage governor routes work when Claude tiers are YELLOW+ and/or cloud CLIs (Codex/Gemini/Grok) are limited.

Offload ladder (top = try first)

  1. Codex (codex exec) — full lifting: implementation, review, tests.
  2. Gemini (agy -p) — full lifting.
  3. Grok (grok -p) — full lifting. Caveat: -p arg dies above ~32KB — pipe or file-feed larger inputs.
  4. Local-model tier (free, unlimited) — route by table below.
  5. Queue non-critical work until a window resets.

Gaming PC (DESKTOP-D249KD4 — RTX 4090 24GB + 48GB RAM, Ollama serialized one-GPU, seconds-fast)

Task classes: a=bulk mechanical edits, b=feature impl from clear spec, c=diff review, d=test writing, e=summarize/compress, f=doc writing, g=search/triage/labeling.

Model Reliable classes Input cap Latency Never give it
qwen3-coder:30b a, b, c, d — primary local implementer ~60KB diff / 256K ctx fast (~seconds) ambiguous specs, cross-cutting refactors
qwen3-next:80b b, c, e, g — strongest local reasoning on this box ~32k tokens medium (hybrid VRAM+RAM) bulk repetitive edits (token cost), huge diffs
devstral:24b a, b (small agentic edits), small c ~32KB fast deep tests, large search loads
gpt-oss:20b e, g, small c — fastest triage/summarize seat ~32KB fastest doc writing, complex tests
gemma4:26b e, f, light c — docs/summaries voice ~48KB fast deep code review, test generation

Streaming PC (DESKTOP-SCDT9HI — 94GB RAM, CPU-only) — batch/overnight queue ONLY

Model Reliable classes Input cap Latency Never give it
gpt-oss:120b c (deep review), b (clear spec), e, f — strongest local model in fleet ~20KB diff / 8k ctx per call 10–30 min/response anything interactive-path, bulk edits
llama3.3:70b c (small diffs), e, g ~10KB diff / 8k ctx 1–5 min bulk edits, tests, long docs
all-minilm:l6-v2 embeddings only any generation task

Hard rules (chaired, binding)

  • Never local: ceiling skip-list work — security-sensitive, architecture, migrations, concurrency, ambiguous specs.
  • Streaming PC is never on the interactive path; overnight/detached queue with polling only.
  • All local output is reviewed by Claude (or a cloud CLI) before merge; local review findings are advisory, never a gate by themselves — small local models are known-wrong on big diffs.
  • Cap local diffs per table; above cap → cloud CLI or queue.
  • Council seat self-reports were chaired: qwen3-coder:30b's "80b handles everything" rejected as overconfident; llama3.3's 2048-token self-cap raised to 8k (conservative CPU cap, model supports more).