Skip to content

refactor(model): remove hand-maintained model injection now covered by models.dev #171

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
cnjack merged 1 commit into from
Jul 24, 2026
+2 −99
Merged

refactor(model): remove hand-maintained model injection now covered by models.dev #171

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
101 changes: 2 additions & 99 deletions internal/model/registry.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -427,8 +427,8 @@ func sortModels(models []*RegistryModel) {
// internal-doc/model-research.md for the per-provider context-window survey behind these
// picks. Model IDs must match the registry exactly or the flag is silently ignored.
var recommendedModels = map[string]map[string]bool{
// GLM-5.2 (1M window) is Zhipu's newest, injected via additionalModels since
// it predates its models.dev record. GLM-5.1 stays selectable but unstarred.
// GLM-5.2 (1M window) is Zhipu's newest flagship. GLM-5.1 stays selectable
// but unstarred.
"zhipuai": {
"glm-5.2": true,
},
Expand Down Expand Up @@ -504,67 +504,8 @@ var contextLimitOverrides = map[string]map[string]int{
},
}

// glm52Model builds a fresh GLM-5.2 entry. Returns a new object per call so each
// provider gets its own instance (deep-copied again per ModelRegistry).
//
// GLM-5.2 shipped 2026-06-13 to GLM Coding Plan users but isn't on models.dev yet
// (the standalone API / open weights land later), so it can't come through
// registry_generated.go. Spec confirmed from the official Z.ai DevPack config
// ("contextWindow": 1000000, "maxTokens": 131072). The full 1M window requires the
// "glm-5.2[1m]" variant on the Coding Plan endpoint. See internal-doc/model-research.md.
func glm52Model() *RegistryModel {
return &RegistryModel{
ID: "glm-5.2", Name: "GLM-5.2", Family: "glm",
Reasoning: true, ToolCall: true, StructuredOutput: true, Temperature: true,
OpenWeights: true,
ReleaseDate: "2026-06-13", LastUpdated: "2026-06-13",
Modalities: &ModelModalities{Input: []string{"text"}, Output: []string{"text"}},
Limit: &ModelLimit{Context: 1_000_000, Output: 131_072},
DefaultEnabled: true,
}
}

// qwen38MaxPreviewModel builds a fresh Qwen3.8-Max-Preview entry. Returns a new
// object per call so each provider gets its own instance (deep-copied again per
// ModelRegistry).
//
// Qwen3.8-Max-Preview is an early-access flagship exposed only on Alibaba's
// Token Plan endpoints (alibaba-token-plan-cn / alibaba-token-plan). It is not
// on models.dev yet, so it can't come through registry_generated.go. Parameters
// mirror qwen3.7-max for now (same window, pricing shape, reasoning options);
// official specs pending. See internal-doc/model-research.md.
func qwen38MaxPreviewModel() *RegistryModel {
return &RegistryModel{
ID: "qwen3.8-max-preview", Name: "Qwen3.8 Max Preview", Family: "qwen",
Reasoning: true, ToolCall: true, Temperature: true,
Modalities: &ModelModalities{Input: []string{"text"}, Output: []string{"text"}},
Cost: &ModelCost{Input: 2.5, Output: 7.5, CacheRead: 0.5, CacheWrite: 3.125},
Limit: &ModelLimit{Context: 1_000_000, Output: 65_536},
ReasoningOptions: []ReasoningOption{
{Type: "toggle"},
{Type: "budget_tokens", Max: intPtr(262144)},
},
}
}

// additionalModels injects models into EXISTING generated providers when a model is
// released before models.dev publishes it. Applied at init() and MERGED in — an
// entry is skipped if the provider already defines that model id, so once the
// official record lands in registry_generated.go it transparently takes over.
// Key: provider ID → model ID → model. See internal-doc/model-research.md.
var additionalModels = map[string]map[string]*RegistryModel{
"zhipuai": {"glm-5.2": glm52Model()},
"zhipuai-coding-plan": {"glm-5.2": glm52Model()},
"zai": {"glm-5.2": glm52Model()},
"zai-coding-plan": {"glm-5.2": glm52Model()},
// Qwen3.8-Max-Preview is exclusive to Alibaba Token Plan endpoints for now.
"alibaba-token-plan-cn": {"qwen3.8-max-preview": qwen38MaxPreviewModel()},
"alibaba-token-plan": {"qwen3.8-max-preview": qwen38MaxPreviewModel()},
}

func init() {
applyStaticProviders()
applyAdditionalModels()
applyContextLimitOverrides()
applyRecommendedModels()
}
Expand All @@ -573,23 +514,6 @@ func init() {
// built into the registry. They are added to generatedProviders/generatedProviderOrder
// at init time so they behave identically to models.dev providers.
var staticProviders = map[string]*RegistryProvider{
// Alibaba Token Plan endpoints are intentionally separate provider ids: the
// credentials and regional base URLs differ from regular DashScope. Their
// early-access model catalog is injected by additionalModels below.
"alibaba-token-plan-cn": {
ID: "alibaba-token-plan-cn",
Name: "Alibaba Token Plan (China)",
Env: []string{"DASHSCOPE_API_KEY"},
API: "https://token-plan.cn-beijing.maas.aliyuncs.com/compatible-mode/v1",
Models: map[string]*RegistryModel{},
},
"alibaba-token-plan": {
ID: "alibaba-token-plan",
Name: "Alibaba Token Plan",
Env: []string{"DASHSCOPE_API_KEY"},
API: "https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1",
Models: map[string]*RegistryModel{},
},
// Kimi For Coding is Moonshot's subscription coding plan. models.dev carries a
// "kimi-for-coding" record, but its model ids (k2p5/k2p6/k2p7/kimi-k2-thinking)
// are undocumented aliases that the vendor's own /models endpoint does not
Expand Down Expand Up @@ -727,27 +651,6 @@ func applyStaticProviders() {
generatedProviderOrder = append(generatedProviderOrder, staticProviderOrder...)
}

// applyAdditionalModels merges hand-maintained models into existing generated
// providers, skipping any model id the provider already defines (so the official
// models.dev record wins once it lands).
func applyAdditionalModels() {
for provID, models := range additionalModels {
prov, ok := generatedProviders[provID]
if !ok {
continue
}
if prov.Models == nil {
prov.Models = make(map[string]*RegistryModel, len(models))
}
for modelID, m := range models {
if _, exists := prov.Models[modelID]; exists {
continue
}
prov.Models[modelID] = m
}
}
}

// applyContextLimitOverrides patches context windows for built-in models whose
// generated value understates the model's real advertised window.
func applyContextLimitOverrides() {
Expand Down
Loading