Skip to content

FlyCoder 0.3: three Ollama models, automatic publishing from GitHub - #10

Merged
Tromset merged 2 commits into
mainfrom
claude-code/peaceful-hamilton-89ge2k
Oct 8, 2026
Merged

Tromset merged 2 commits into
mainfrom
claude-code/peaceful-hamilton-89ge2k

Conversation

@Tromset

@Tromset Tromset commented Oct 8, 2026

Copy link
Copy Markdown
Owner

Why

  • ollama run Tromset/flycoder failed: the registry has no latest or fast tag, only *-gguf tags.
  • Names disagreed: the page said 0.3, the tags 0.2-beta, and the model introduced itself as "FlyCoder 0.2 beta".
  • Gemma 4 12B, the base of the main variant, does not run on the owner's 16 GB Mac.
  • Publishing was manual, and the page text was pasted by hand.

What changes

Model Base (GGUF) Weights Mac Context Min Ollama
Tromset/flycoder0.3 qwen3.5:9b 6.6 GB 16 GB 64K 0.30.0
Tromset/flycoder0.3fast qwen3.5:4b 3.3 GB 8 GB 32K 0.30.0
Tromset/flycoder0.3pro qwen3.8:27b 18 GB 32 GB+ 64K 0.32.12
Tromset/flycoder0.3:lite / :router qwen3.5:2b / 0.8b FlyBrain experts
  • Portable builds: all tags are GGUF, so they run on any Mac, Linux or Windows. MLX is opt-in with install.sh --mlx.
  • Modelfiles: Qwen's sampling for coding and 64K of context. The system prompt now asks the model for a final self-check and a single final version.
  • Automatic publishing: the new .github/workflows/publish-ollama.yml runs on every push to main that touches a Modelfile, docs/ollama/ or the publishing scripts. It rebuilds and pushes the changed models with scripts/publish.sh, then updates the ollama.com pages with scripts/update-ollama-pages.sh. It can also be started by hand.
  • Short pages in English: docs/ollama/<model>.md, about 20 lines each. npm test checks that each page shows the right name, size, memory and context. The old Tromset/flycoder page gets a "moved" notice.
  • Rest of the repository: FlyBrain, install.sh, the bench, the README and Documentation.md use the new names.

Setup needed (one time, by the owner)

Add these repository secrets:

  • OLLAMA_KEY: the private ~/.ollama/id_ed25519 of a machine signed in to the Tromset account.
  • OLLAMA_API_KEY: an API key from ollama.com/settings/keys.

The page update reuses the request behind the page's Edit button, which Ollama does not document. If ollama.com refuses it, the workflow shows a warning and the model pushes still succeed.

Verification

  • npm test: 65 tests pass.
  • actionlint and shellcheck pass.
  • With Ollama 0.40.1 on Linux, flycoder0.3, flycoder0.3fast, :lite and :router were built from these Modelfiles.
    • flycoder0.3 introduces itself as "FlyCoder 0.3, … built on Qwen3.5 9B" and writes correct code.
    • It uses 8.8 GB at 64K context (ollama ps).
  • scripts/publish.sh was run end to end with push and rm faked.
  • The 20-problem bench run on CPU is in progress; its results will be added to this PR.
  • Not tested here: the pro model (its 18 GB base does not fit this machine) and real pushes to ollama.com (no account key in this environment).

🤖 Generated with Claude Code

https://claude.ai/code/session_013QgiejZTjqQAUv4N4EFskK


Generated by Claude Code

claude added 2 commits October 8, 2026 18:56
- Tromset/flycoder0.3 (Qwen3.5 9B, 16 GB Macs), flycoder0.3fast (Qwen3.5 4B)
  and flycoder0.3pro (Qwen3.8 27B), all GGUF so they run on any Ollama >= 0.30
  (0.32.12 for pro); Gemma 4 dropped, it did not run on every 16 GB Mac
- 64K context by default (8.8 GB measured for flycoder0.3), Qwen coding sampling,
  system prompt asks for a final self-check
- publish-ollama workflow: every change to main rebuilds the changed models,
  pushes them and updates the ollama.com pages from docs/ollama/
- short English ollama.com pages, one per model, checked by npm test
- FlyBrain, installer, bench and docs moved to the new names

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QgiejZTjqQAUv4N4EFskK
- six-step agent workflow in the system prompt: contract, plan, code,
  tests, check, report; plus rules for coding agents
- a worked example (MESSAGE) showing the expected answer shape; npm test
  runs the example code against its own test cases
- multi-token prediction made explicit (draft_num_predict 4); context
  memory computed from the GGUF header (about 4.4 GB at 64K)
- every Modelfile declares its minimum Ollama version (REQUIRES)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013QgiejZTjqQAUv4N4EFskK
@Tromset
Tromset marked this pull request as ready for review October 8, 2026 19:04
@Tromset
Tromset merged commit 9fbabe4 into main Oct 8, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants