@aliou/pi-neuralwatt 0.14.0 → 0.14.2
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
|
@@ -9,7 +9,35 @@ import { NEURALWATT_MODELS } from "./public-models";
|
|
|
9
9
|
// gated by includeEarlyAccessModels and hardcode entries so they remain
|
|
10
10
|
// available from the offline catalog.
|
|
11
11
|
// Move an entry to public-models.ts once Neuralwatt advertises it publicly.
|
|
12
|
-
export const EARLY_ACCESS_NEURALWATT_MODELS: ProviderModelConfig[] = [
|
|
12
|
+
export const EARLY_ACCESS_NEURALWATT_MODELS: ProviderModelConfig[] = [
|
|
13
|
+
// Qwen3.8 27B FP8 — pre-release dense 27B VL model with MTP speculative
|
|
14
|
+
// decoding, native 262K context, served on 2x H200. Returned by the
|
|
15
|
+
// authenticated /v1/models catalog but absent from the public endpoint.
|
|
16
|
+
// Binary thinking is toggled through `chat_template_kwargs.enable_thinking`
|
|
17
|
+
// (verified: `enable_thinking: false` suppresses reasoning output); the API
|
|
18
|
+
// exposes no `reasoning` block, so the thinking level map is the high-only
|
|
19
|
+
// fallback with `off: null`.
|
|
20
|
+
// https://huggingface.co/Qwen/Qwen3.8-27B-FP8
|
|
21
|
+
{
|
|
22
|
+
id: "Qwen/Qwen3.8-27B-FP8",
|
|
23
|
+
name: "Qwen 3.8 27B FP8",
|
|
24
|
+
reasoning: true,
|
|
25
|
+
input: ["text", "image"],
|
|
26
|
+
cost: { input: 0.45, output: 3.2, cacheRead: 0.25, cacheWrite: 0 },
|
|
27
|
+
contextWindow: 262_128,
|
|
28
|
+
maxTokens: 65_536,
|
|
29
|
+
compat: {
|
|
30
|
+
supportsDeveloperRole: false,
|
|
31
|
+
maxTokensField: "max_tokens",
|
|
32
|
+
requiresReasoningContentOnAssistantMessages: true,
|
|
33
|
+
thinkingFormat: "chat-template",
|
|
34
|
+
chatTemplateKwargs: {
|
|
35
|
+
enable_thinking: { $var: "thinking.enabled" },
|
|
36
|
+
},
|
|
37
|
+
},
|
|
38
|
+
thinkingLevelMap: { ...buildThinkingLevelMap(undefined) },
|
|
39
|
+
},
|
|
40
|
+
];
|
|
13
41
|
|
|
14
42
|
// Per-ID overrides for known early-access models. The authenticated /v1/models
|
|
15
43
|
// endpoint exposes pricing and capabilities, but some Pi-specific behavior
|
|
@@ -65,7 +65,9 @@ export function createNeuralwattProvider(
|
|
|
65
65
|
if (await ctx.env(NEURALWATT_API_KEY_ENV)) {
|
|
66
66
|
return { type: "api_key", source: NEURALWATT_API_KEY_ENV };
|
|
67
67
|
}
|
|
68
|
-
|
|
68
|
+
// Anonymous availability: requests resolve with an empty key, and
|
|
69
|
+
// in aperture proxy mode the gateway owns auth.
|
|
70
|
+
return { type: "api_key", source: "anonymous" };
|
|
69
71
|
},
|
|
70
72
|
resolve: async ({ ctx, credential, signal }) => {
|
|
71
73
|
signal.throwIfAborted();
|