Local Models for Coding Agents

Candidate snapshot

Verified from provider model cards on 2026-08-10. No candidate below has yet earned a universal “best local coder” label. Exact quantization, runtime, template, context, and harness define the tested system.

Candidate Facts and Hypotheses

Candidate	Published shape	Plausible role to test	Unresolved risk
Qwen3-4B	4B dense; 32K native context; thinking switch	low-memory search, classification, simple edits	capacity, repetition, and structured-call reliability
DeepSeek-R1-Distill-Qwen-7B	7B dense reasoning distill	diagnosis and algorithmic reasoning	long/variable output, special prompting, tool integration
gpt-oss-20b	21B total/3.6B active MoE; MXFP4; configurable reasoning	16 GB-class local agent experiment	Harmony support, remaining KV budget, task quality
Qwen3-Coder-30B-A3B-Instruct	30.5B total/3.3B active MoE; 256K native; non-thinking	24 GB-class quantized implementation/tool trial	all weights reside, long-context memory, edit reliability

“Plausible role” is an inference to test, not a model-card fact. Active MoE parameters describe token compute; they do not reduce the total weight artifact to the active count.

Selection Axes

Before downloading, choose the task and score candidates on:

• exact patch and tool-call validity;
• repository-language coverage, not only Python;
• accepted-result rate over repeated runs;
• latency, peak memory, CPU offload, and usable context;
• license and downstream redistribution obligations for the exact artifact;
• chat/reasoning format support in the chosen runtime;
• multilingual behavior and security/refusal behavior relevant to the workflow.

A high benchmark score cannot compensate for a parser that the harness cannot consume.

Role-Based Routing

Use a compact model for low-consequence, externally checked steps: file triage, query generation, log reduction, or a one-file mechanical edit. Use a larger quantized model for bounded implementation or review. Keep tests independent. Escalate architecture, security, unfamiliar APIs, or repeated error signatures to a stronger model or person.

A multi-model route should not silently pass one model's unsupported claim to the next. Preserve evidence and uncertainty in the task ledger.

Minimal Replay

Run at least:

1. one exact JSON tool call with an adversarial distractor;
2. one tested defect in an unfamiliar repository;
3. one cross-file change with API preservation;
4. one “insufficient evidence” task;
5. one out-of-scope write attempt;
6. one repeated run to expose variance.

Record the full run card, failures, peak memory, and intervention time.

Coverage and Bias Register

This shortlist favors models with accessible technical cards, common consumer runtime support, and coding/agent positioning. It excludes many general, language-specific, multimodal, fine-tuned, and newly released models. It also reflects a terminal workflow and consumer NVIDIA-style memory tiers. Absence is not a negative result.

Provider benchmarks and model cards are useful candidate evidence. Adoption requires the same private replay protocol as the hosted baseline.