mirror of
https://github.com/badlogic/pi-mono.git
synced 2026-11-05 23:37:25 +00:00
fix(ai): treat Kimi K3.1 Pro models as Kimi K3 family
Kimi Coding, Moonshot, and OpenCode Go generated K3.1 Pro and K3.1 Pro Highspeed with Kimi K2 settings because K3 detection matched exact IDs. Detect the K3 family by ID pattern, and Kimi Coding K3-style models by their effort options. Add implied subscription costs and clamp gateway output limits like K3.
This commit is contained in:
Armin Ronacher
parent
bdb151774e
commit
1de2d7adea
2 changed files with 30 additions and 19 deletions
|
|
@ -22,6 +22,7 @@
|
|||
- Fixed streamed tool-call arguments being re-parsed in full on every delta, which made CPU time and memory grow quadratically with the size of a tool call (a 90 KB `write` on Anthropic peaked at about 140 MB of heap; a 500 KB one took minutes). Every provider and the pi-messages client now read each delta once, and keep the raw argument text and their other stream state (`partialJson`, `partialArgs`, `index`, `customInput`, `streamIndex`, Bedrock's encrypted reasoning bytes) off the content blocks, so it no longer reaches partial messages or transcripts of failed streams ([#9265](https://github.com/earendil-works/pi/issues/9265))
|
||||
- Fixed grammar (custom) tool input streaming copying the whole input so far on every delta ([#9265](https://github.com/earendil-works/pi/issues/9265))
|
||||
- Fixed bounded error texts (telemetry span messages, truncated provider error bodies, pi-messages diagnostics, Radius gateway errors) keeping the whole text they were cut from alive ([#9265](https://github.com/earendil-works/pi/issues/9265))
|
||||
- Fixed Kimi K3.1 Pro and K3.1 Pro Highspeed on Kimi Coding, Moonshot, and OpenCode Go using Kimi K2 request settings. They now send `reasoning_effort` (low/high/max), return earlier reasoning, accept tools added mid-conversation, and report Moonshot rates on Kimi Coding. Kimi Coding `k3-256k` now reports Kimi K3 rates.
|
||||
|
||||
## [1.4.0] - 2026-10-21
|
||||
|
||||
|
|
|
|||
|
|
@ -349,7 +349,11 @@ const KIMI_K3_OPENAI_COMPAT: OpenAICompletionsCompat = {
|
|||
thinkingFormat: "openai",
|
||||
supportsReasoningEffort: true,
|
||||
};
|
||||
const OPENROUTER_KIMI_K3_MODEL_IDS = new Set(["moonshotai/kimi-k3", "~moonshotai/kimi-latest"]);
|
||||
// Kimi K3 and its point releases (kimi-k3.1-pro, kimi-k3.1-pro-highspeed) share K3's request
|
||||
// format. https://api.moonshot.ai/v1/models reports effort levels and dynamic tool loading for all.
|
||||
function isKimiK3FamilyId(modelId: string): boolean {
|
||||
return /^kimi-k3(?:\.\d+)?(?:-|$)/.test(modelId);
|
||||
}
|
||||
|
||||
const ANT_LING_RING_THINKING_LEVEL_MAP = {
|
||||
off: null,
|
||||
|
|
@ -928,9 +932,9 @@ function applyOpenAIToolSearchMetadata(model: Model<Api>): void {
|
|||
function applyOpenAICompletionsTranscriptMetadata(model: Model<Api>): void {
|
||||
if (model.api !== "openai-completions") return;
|
||||
const isKimiK3 =
|
||||
(model.provider.startsWith("moonshot") && model.id === "kimi-k3") ||
|
||||
(model.provider.startsWith("moonshot") && isKimiK3FamilyId(model.id)) ||
|
||||
(model.provider === "fireworks" && model.id.includes("kimi-k3")) ||
|
||||
((model.provider === "opencode" || model.provider === "opencode-go") && model.id === "kimi-k3");
|
||||
((model.provider === "opencode" || model.provider === "opencode-go") && isKimiK3FamilyId(model.id));
|
||||
const isMoonshotKimiK2 =
|
||||
model.provider.startsWith("moonshot") &&
|
||||
(model.id === "kimi-k2.6" || model.id === "kimi-k2.7-code" || model.id === "kimi-k2.7-code-highspeed");
|
||||
|
|
@ -1763,13 +1767,12 @@ function processFireworksModels(provider: ModelsDevProvider | undefined): Model<
|
|||
// equivalent Moonshot API rates to estimate the value of subscription usage.
|
||||
const KIMI_CODING_IMPLIED_COSTS: Record<string, Model<Api>["cost"]> = {
|
||||
k3: KIMI_K3_COST,
|
||||
"k3-256k": KIMI_K3_COST,
|
||||
"k3.1-pro": { input: 2, output: 10, cacheRead: 0.2, cacheWrite: 0 },
|
||||
"k3.1-pro-highspeed": { input: 4, output: 20, cacheRead: 0.2, cacheWrite: 0 },
|
||||
"kimi-for-coding": { input: 0.95, output: 4, cacheRead: 0.19, cacheWrite: 0 },
|
||||
"kimi-for-coding-highspeed": { input: 1.9, output: 8, cacheRead: 0.38, cacheWrite: 0 },
|
||||
};
|
||||
// Models whose https://api.kimi.com/coding/v1/models entry reports effort levels and
|
||||
// dynamic tool loading (tools declared in later system messages). Highspeed reports
|
||||
// neither and rejects tool-bearing system messages with a 400.
|
||||
const KIMI_CODING_K3_STYLE_MODEL_IDS = new Set(["k3", "k3-256k", "kimi-for-coding"]);
|
||||
|
||||
// Kimi Coding uses Chat Completions, the same protocol as the official Kimi Code client.
|
||||
function createKimiCodingModels(kimiModels: Record<string, ModelsDevModel>): Model<"openai-completions">[] {
|
||||
|
|
@ -1777,7 +1780,11 @@ function createKimiCodingModels(kimiModels: Record<string, ModelsDevModel>): Mod
|
|||
for (const [modelId, m] of Object.entries(kimiModels)) {
|
||||
if (m.tool_call !== true) continue;
|
||||
|
||||
const isK3Style = KIMI_CODING_K3_STYLE_MODEL_IDS.has(modelId);
|
||||
// K3-style models report effort levels and dynamic tool loading (tools declared in
|
||||
// later system messages) in https://api.kimi.com/coding/v1/models; models.dev mirrors
|
||||
// the effort levels. kimi-for-coding-highspeed reports neither and rejects
|
||||
// tool-bearing system messages with a 400.
|
||||
const isK3Style = m.reasoning_options?.some((option) => option.type === "effort") === true;
|
||||
const impliedCost = KIMI_CODING_IMPLIED_COSTS[modelId];
|
||||
// K3-style models send reasoning_effort, so expose the effort values models.dev lists.
|
||||
const effortLevels = isK3Style && m.reasoning_options ? getEffortThinkingLevelMap(m.reasoning_options) : undefined;
|
||||
|
|
@ -2440,7 +2447,7 @@ async function loadModelsDevData(): Promise<Model<any>[]> {
|
|||
for (const [modelId, m] of Object.entries(moonshotModels[key])) {
|
||||
if (m.tool_call !== true) continue;
|
||||
|
||||
const isKimiK3 = modelId === "kimi-k3";
|
||||
const isKimiK3 = isKimiK3FamilyId(modelId);
|
||||
const compat = { ...(isKimiK3 ? KIMI_K3_OPENAI_COMPAT : KIMI_OPENAI_COMPAT) };
|
||||
models.push({
|
||||
id: modelId,
|
||||
|
|
@ -2451,14 +2458,15 @@ async function loadModelsDevData(): Promise<Model<any>[]> {
|
|||
reasoning: isKimiK3 || m.reasoning === true,
|
||||
input: m.modalities?.input?.includes("image") ? ["text", "image"] : ["text"],
|
||||
// Moonshot does not bill cache writes; models.dev lists the input rate as cache_write for Kimi K3.
|
||||
cost: isKimiK3
|
||||
? { ...KIMI_K3_COST }
|
||||
: {
|
||||
input: m.cost?.input || 0,
|
||||
output: m.cost?.output || 0,
|
||||
cacheRead: m.cost?.cache_read || 0,
|
||||
cacheWrite: m.cost?.cache_write || 0,
|
||||
},
|
||||
cost:
|
||||
modelId === "kimi-k3"
|
||||
? { ...KIMI_K3_COST }
|
||||
: {
|
||||
input: m.cost?.input || 0,
|
||||
output: m.cost?.output || 0,
|
||||
cacheRead: m.cost?.cache_read || 0,
|
||||
cacheWrite: isKimiK3 ? 0 : m.cost?.cache_write || 0,
|
||||
},
|
||||
contextWindow: m.limit?.context || 4096,
|
||||
maxTokens: m.limit?.output || 4096,
|
||||
compat,
|
||||
|
|
@ -2906,8 +2914,10 @@ async function generateModels() {
|
|||
}
|
||||
// Keep Kimi K3's canonical output limit when gateway metadata is missing or incorrect.
|
||||
if (
|
||||
(candidate.provider === "openrouter" && OPENROUTER_KIMI_K3_MODEL_IDS.has(candidate.id)) ||
|
||||
(candidate.provider === "vercel-ai-gateway" && candidate.id === "moonshotai/kimi-k3")
|
||||
(candidate.provider === "openrouter" && candidate.id === "~moonshotai/kimi-latest") ||
|
||||
((candidate.provider === "openrouter" || candidate.provider === "vercel-ai-gateway") &&
|
||||
candidate.id.startsWith("moonshotai/") &&
|
||||
isKimiK3FamilyId(candidate.id.slice("moonshotai/".length)))
|
||||
) {
|
||||
candidate.maxTokens = KIMI_K3_MAX_TOKENS;
|
||||
}
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue