[mikkel]

Bug

This is the companion to 255BITS/patchling#11, where the JS 0.3.0 default xiaomi/mimo-v2.5-pro-ultraspeed is retired and returns HTTP 410. That retired id doesn't appear anywhere in this repo. But the Python default, DEFAULT_MODEL = "deepseek-reasoner", has the same failure mode: it is not listed on NanoGPT, which is the default GPTDIFF_LLM_BASE_URL. So generate_diff/smartapply/the CLI break whenever GPTDIFF_MODEL is unset and no model is passed.

Fix

The default is now deepseek/deepseek-v4.1-flash-1017, the same id as the JS PR.

Why this model: Mikkel asked for DeepSeek V4.1, the 1017 release. In NanoGPT's free GET https://nano-gpt.com/api/v1/models?detailed=true listing (fetched 2026-10-21 13:03 MT):

• deepseek/deepseek-v4.1-flash-1017 ("DeepSeek V4.1 Flash 1017", the October 17, 2026 snapshot) is listed at $0.10 / $0.60 per 1M input/output tokens, with a 1,048,576 context and 393,216 max output tokens. It is the only id that names the 1017 release with standard serving.
• Not chosen: the same snapshot is also listed as deepseek/deepseek-v4.1-flash-1017:thinking (same price, thinking always on) and as TEE/deepseek-v4.1-flash (secure-enclave serving, $0.65 / $1.45). The undated deepseek/deepseek-v4.1-flash is a rate-limited beta.
• NanoGPT lists no non-Flash V4.1 model. "V4.1" here can only mean V4.1 Flash; the only DeepSeek Pro models are v4.
• Neither retired default is listed: xiaomi/mimo-v2.5-pro-ultraspeed (JS) and deepseek-reasoner (Python).

An earlier commit on this branch used xiaomi/mimo-v2.6-pro. The latest commit replaces it everywhere.

Changes

• patchling/env.py: DEFAULT_MODEL and its comment. The old comment said the JS default differed; now they match.
• Docs that state the default: README env table (which wrongly listed gemini-3-pro-preview), docs/api.md ×2, docs/cli.md ×2, and the core.py docstring.
• The bench/relevance/run.py fallback diff model follows the new default.
• Tests:
  • New tests/test_default_model.py: the default is deepseek/deepseek-v4.1-flash-1017, neither retired id is the default, and resolve_model(None) with GPTDIFF_MODEL unset returns it.
  • tests/test_public_api.py is updated to the new id.
• New CHANGELOG.md with an Unreleased note. The repo had no changelog.

Unchanged on purpose:

• filter_relevant's default, xiaomi/mimo-v2.6-flash, is still listed.
• Unit tests that pass "deepseek-reasoner" as an arbitrary model string are left as is.
• Docs that pass --model deepseek-reasoner explicitly (docs/examples/*) are left alone, though those examples would also fail on NanoGPT. Same for the model-recommendation tables and the quickstart, which still recommend gemini-3-pro-preview (and the quickstart still calls it the default). These are worth a separate docs pass.

No version bump, tag, or publish.

Tests (local)

pytest, offline with a cold tiktoken cache and the fake_tiktoken stub: 207 passed, 2 skipped. Main has 204 passed, 2 skipped, so the difference is the 3 new tests. mkdocs build --strict passes.

Live smoke (one call, authorized by Mikkel)

One tiny generate_diff call through the default path: no model argument, GPTDIFF_MODEL unset, the default NanoGPT base URL, and max_tokens 400. The task was a one-line edit (hello → hi).

• Model sent and model returned: deepseek/deepseek-v4.1-flash-1017, HTTP 200.
• 128 prompt + 31 completion = 159 tokens, cost $0.0000314, about 2 s.
• The result was a valid unified diff (-hello / +hi).
• The API key came from the environment and was never printed, logged or committed.