Skip to content

feat: add ollama/qwen3.5-9b-q4_k_m model params - #37

Open
zunami wants to merge 2 commits into
mnfst:mainfrom
zunami:main
Open

feat: add ollama/qwen3.5-9b-q4_k_m model params#37
zunami wants to merge 2 commits into
mnfst:mainfrom
zunami:main

Conversation

@zunami

@zunami zunami commented May 27, 2026

Copy link
Copy Markdown

Adds parameter definitions for qwen3.5:9b-q4_K_M running via Ollama.

Intentionally omits thinking.type — Ollama reports capabilities: ["thinking"]
via /api/show, but does not accept the Anthropic-style thinking API parameter.
Sending it causes "Failed to reach upstream provider" errors.

Ollama handles thinking internally via prompt prefix (/no_think), not via
API parameters.

@vercel

vercel Bot commented May 27, 2026

Copy link
Copy Markdown

@zunami is attempting to deploy a commit to the Manifest Team on Vercel.

A member of the Team first needs to authorize it.

@guillaumegay13

Copy link
Copy Markdown
Collaborator

Thanks for the contribution, @zunami — and sorry for the slow look! 🙏 A few things to sort before this can go in:

1. Provider should be the maker, not the runtime. The catalog is keyed by who makes the model, not how it's served — so Ollama-served models live under their maker. Qwen is Alibaba, and we already track it at models/alibaba/ (qwen3.5.yaml, qwen3.5-flash.yaml, …). A new top-level ollama provider would be the first gateway/runtime in the tree, which we've been avoiding (same reason OpenRouter/OpenCode-Go aren't providers here).

2. Param surface. The file mixes OpenAI-compatible params (max_tokens, top_p) with Ollama-native ones (num_ctx, top_k). Worth picking one surface — and note Ollama's own max-output option is num_predict, not max_tokens. Also missing the schema header the other files use:

# yaml-language-server: $schema=https://modelparams.dev/api/v1/schema.json

3. CI is red + branch is stale. It needs a rebase on main, and the generated package files have to be regenerated and committed (npm run codegen --workspace=modelparams) or the "generated files in sync" build fails.

One open question worth your thoughts: since alibaba/qwen3.5 already exists, is a separate entry for the local quantized :9b-q4_K_M tag adding configurable params that the existing entry doesn't cover? If the knobs are the same, it might be redundant; if the quant exposes genuinely different Ollama options, that's the case to make. Happy to help reshape it whichever way you'd like to take it. 🙂

brunobuddy added a commit that referenced this pull request Aug 3, 2026
* fix: compare the param guard against the merge base

The guard diffed the checked-out tree against the tip of the base branch.
On a pull request that tree is GitHub's refs/pull/N/merge commit, which is
only recomputed when the PR or its base is pushed to — so a re-run, or a
`labeled` event, compares a merge built on an older base against a base
tip that has moved on. Every parameter the base gained in between reads as
a removal.

That is how run 26753498917 blocked PR #37: re-run four days after it was
queued, it reported mistral/magistral-* losing `prompt_mode`, a file the PR
never touched. Diffing against the merge base puts both sides of the
comparison on the same snapshot. Removals the PR really does make are still
caught: PR #132 still reports dropping `reasoning_effort` from z-ai/glm-5.2.

* test: stop assuming HEAD has a parent

CI clones shallow, so `git rev-parse HEAD^` fails there. The remaining
cases cover the contract that matters — a resolvable ref narrows to a
shared commit, an unresolvable one yields null and the guard falls back.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

model Add a model that's missing

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants