Free Zen models with reasoning=true show zero thinking variants (reasoning_options=[] blocks fallback)

Summary

Several free OpenCode Zen models advertise reasoning = true in models.dev but expose zero thinking variants in OpenCode (/models picker + variant_cycle). Users cannot select low / medium / high thinking effort for these models.

Affected models (provider = opencode)

Verified against https://models.dev/api.json (opencode provider) and Zen docs (2026-09-13):

Model ID	reasoning	reasoning_options	Variants shown
big-pickle	true	[]	none (bug)
mimo-v2.5-free	true	[]	none (bug)
nemotron-3-ultra-free	true	[]	none (bug)
nemotron-3.5-lightning-free	true	[]	none (bug)
minimax-m2.5-free	true	[]	none (bug)
ling-3.0-flash-fin-free	true	[{type="toggle"}]	low/medium/high via fallback (OK)
muse-spark-1.3-contributor-free	true	[{type="effort", values=[minimal,low,medium,high,xhigh]}]	5 variants (OK)
deepseek-v4-flash-free	true	[{type="effort", values=[low,high,max]}]	OK
kimi-k2.5-free / glm-5-free / qwen3.6-plus-free	true	[{type="toggle"}] (+budget for qwen)	none — fallback denylisted (see below)

Zen endpoint table serves 5/6 advertised free models via https://opencode.ai/zen/v1/chat/completions (@ai-sdk/openai-compatible); only muse-spark-1.3-contributor-free uses .../v1/responses (@ai-sdk/openai).

Expected vs actual

• Expected: models with reasoning = true served over OpenAI-compatible API expose generic low / medium / high variants (same as ling-3.0-flash-fin-free does today).
• Actual: models with reasoning_options = [] show no variants at all.

Root cause (two layers)

1. reasoningVariants() returns {} instead of undefined for []

packages/opencode/src/provider/transform.ts:1704:

export function reasoningVariants(model: ModelsDev.Model, target: Provider.Model) {
  const options = model.reasoning_options
  if (options === undefined) return
  if (options.length === 0) return {}
  // ...
}

2. ?? fallback never runs on {}

packages/opencode/src/provider/provider.ts:1314:

const variants = ProviderTransform.reasoningVariants(model, base) ?? ProviderTransform.variants(base)

{} is not nullish, so generic variants(base) never runs for the [] models. Result: zero variants despite reasoning=true.

3. Hardcoded denylist in variants() also returns {}

packages/opencode/src/provider/transform.ts:777:

if (
  id.includes("deepseek-chat") ||
  id.includes("deepseek-reasoner") ||
  id.includes("deepseek-r1") ||
  id.includes("deepseek-v3") ||
  id.includes("minimax") ||
  (id.includes("glm") && !glm52) ||
  id.includes("kimi") ||
  id.includes("k2p") ||
  id.includes("qwen") ||
  id.includes("big-pickle")
)
  return {}

This blocks fallback for big-pickle, kimi-*, qwen-*, glm-* (non-5.2), minimax-* even when reasoning=true. For toggle-only models on @ai-sdk/openai-compatible, reasoningToggle() returns {} -> nonEmptyVariants({}) -> undefined -> fallback -> denylist {}. Net: no variants.

Models.dev evidence

Raw TOMLs (anomalyco/models.dev@dev:providers/opencode/models/):

• mimo-v2.5-free.toml: reasoning = true, reasoning_options = []
• nemotron-3-ultra-free.toml: reasoning = true, reasoning_options = []
• nemotron-3.5-lightning-free.toml: reasoning_options = []
• big-pickle.toml: reasoning = true, reasoning_options = []
• minimax-m2.5-free.toml: reasoning = true, reasoning_options = []
• ling-3.0-flash-fin-free.toml: reasoning_options = [{ type = "toggle" }] (works via fallback)
• muse-spark-1.3-contributor-free.toml: reasoning_options = [{ type = "effort", values = [...] }] + [provider] npm = "@ai-sdk/openai" (works)

Suggested fix (either side, ideally both)

• Option A (code, anomalyco/opencode): treat empty variants as missing. E.g. reasoningVariants() returns undefined for [], or change call site to fall back when Object.keys(variants).length === 0.
• Option B (data, anomalyco/models.dev): give free OpenAI-compatible reasoning models real effort options, e.g. reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }], or document that [] means fixed always-on reasoning with no knob.
• Re-evaluate denylist entry for big-pickle if reasoningEffort passthrough works on Zen chat-completions endpoint.

Workaround

Custom variants in opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "opencode": {
      "models": {
        "mimo-v2.5-free": {
          "variants": {
            "low": { "reasoningEffort": "low" },
            "medium": { "reasoningEffort": "medium" },
            "high": { "reasoningEffort": "high" }
          }
        }
      }
    }
  }
}

Environment

• opencode v2.0.3 (Windows, C:\Users\baban\.bun\bin\opencode.exe)
• Zen docs last updated 2026-09-13, /zen/v1/models checked 2026-09-14
• models.dev opencode provider, 102 models