@infernetprotocol/cli 0.1.41 → 0.1.46

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.
@@ -93,44 +93,86 @@ class EventBuffer {
93
93
  // /chat endpoint sometimes doesn't pass a model name). Without this
94
94
  // fallback, every model-unspecified job died with
95
95
  // "ollama backend: no model — set INFERNET_ENGINE_MODEL or pass model in the job"
96
- let enginePromise = null;
97
- function getEngine() {
98
- if (!enginePromise) {
99
- enginePromise = (async () => {
100
- const config = (await loadConfig()) ?? {};
101
- const eng = config.engine ?? {};
102
- const opts = {};
103
- if (eng.backend) opts.backend = eng.backend;
104
- if (eng.model) opts.defaultModel = eng.model;
105
- if (eng.ollamaHost) opts.host = eng.ollamaHost;
106
- // Optional override for the per-request num_thread cap.
107
- // Default (50% of cores) is computed inside the Ollama
108
- // backend; only forward when the operator explicitly set
109
- // an override in config.
110
- if (Number.isFinite(eng.ollama_num_thread) && eng.ollama_num_thread > 0) {
111
- opts.numThread = eng.ollama_num_thread;
112
- }
113
- return createEngine(opts);
114
- })().catch((err) => {
96
+ // One engine per backend, cached. We route PER MODEL: if the requested model
97
+ // is being served by vLLM (:8000), use the vLLM engine; otherwise the default
98
+ // (config backend or auto-select → Ollama). This is what makes vLLM the actual
99
+ // serving path for hf: models while Ollama keeps handling its own tags.
100
+ const engineCache = new Map(); // cacheKey -> Promise<engine>
101
+
102
+ function buildEngine(cacheKey, opts) {
103
+ if (!engineCache.has(cacheKey)) {
104
+ const p = (async () => createEngine(opts))().catch((err) => {
115
105
  // Reset so a transient failure doesn't permanently poison the
116
106
  // daemon — next job will retry initialization.
117
- enginePromise = null;
107
+ engineCache.delete(cacheKey);
118
108
  throw err;
119
109
  });
110
+ engineCache.set(cacheKey, p);
120
111
  }
121
- return enginePromise;
112
+ return engineCache.get(cacheKey);
122
113
  }
123
114
 
124
- export async function shutdownEngine() {
125
- if (!enginePromise) return;
115
+ async function defaultEngine() {
116
+ const config = (await loadConfig()) ?? {};
117
+ const eng = config.engine ?? {};
118
+ const opts = {};
119
+ if (eng.backend) opts.backend = eng.backend;
120
+ if (eng.model) opts.defaultModel = eng.model;
121
+ if (eng.ollamaHost) opts.host = eng.ollamaHost;
122
+ // Optional override for the per-request num_thread cap. Default (50% of
123
+ // cores) is computed inside the Ollama backend; only forward when set.
124
+ if (Number.isFinite(eng.ollama_num_thread) && eng.ollama_num_thread > 0) {
125
+ opts.numThread = eng.ollama_num_thread;
126
+ }
127
+ return buildEngine(`default:${eng.backend ?? 'auto'}`, opts);
128
+ }
129
+
130
+ /** Best-effort: does Ollama (at host) currently serve this exact tag? */
131
+ async function ollamaHasModel(modelName, host) {
126
132
  try {
127
- const engine = await enginePromise;
128
- await engine.shutdown();
133
+ const ctrl = new AbortController();
134
+ const t = setTimeout(() => ctrl.abort(), 800);
135
+ const r = await fetch(`${host}/api/tags`, { signal: ctrl.signal });
136
+ clearTimeout(t);
137
+ if (!r.ok) return false;
138
+ const data = await r.json();
139
+ return (data.models ?? []).some((m) => m.name === modelName || m.model === modelName);
129
140
  } catch {
130
- // best-effort
131
- } finally {
132
- enginePromise = null;
141
+ return false;
142
+ }
143
+ }
144
+
145
+ async function getEngineForModel(modelName) {
146
+ // vLLM is the default serving path. Route PER MODEL so both engines coexist:
147
+ // 1. vLLM is serving this model → vLLM (preferred, high throughput)
148
+ // 2. Ollama has this GGUF tag → Ollama (its library keeps working even
149
+ // while vLLM is up serving another model)
150
+ // 3. otherwise → default (auto-select, which prefers vLLM)
151
+ if (modelName) {
152
+ try {
153
+ const { detectVllmModels, VLLM_HOST } = await import('./vllm.js');
154
+ const vModels = await detectVllmModels();
155
+ if (vModels.includes(modelName)) {
156
+ return buildEngine('vllm', { backend: 'vllm', host: VLLM_HOST, defaultModel: modelName });
157
+ }
158
+ } catch {
159
+ // vLLM module/probe failed — try Ollama, then the default.
160
+ }
161
+ const config = (await loadConfig().catch(() => null)) ?? {};
162
+ const ollamaHost = config.engine?.ollamaHost ?? process.env.OLLAMA_HOST ?? 'http://localhost:11434';
163
+ if (await ollamaHasModel(modelName, ollamaHost)) {
164
+ return buildEngine('ollama', { backend: 'ollama', host: ollamaHost, defaultModel: modelName });
165
+ }
133
166
  }
167
+ return defaultEngine();
168
+ }
169
+
170
+ export async function shutdownEngine() {
171
+ const engines = [...engineCache.values()];
172
+ engineCache.clear();
173
+ await Promise.all(engines.map(async (p) => {
174
+ try { (await p).shutdown(); } catch { /* best-effort */ }
175
+ }));
134
176
  }
135
177
 
136
178
  /**
@@ -167,7 +209,7 @@ export async function executeChatJob({ client, job, node }) {
167
209
  }
168
210
  }
169
211
 
170
- const engine = await getEngine();
212
+ const engine = await getEngineForModel(job.model_name);
171
213
  const buffer = new EventBuffer(client, job.id);
172
214
 
173
215
  const t0 = Date.now();
@@ -2,21 +2,28 @@
2
2
  * Search-driven training-data collector.
3
3
  *
4
4
  * Pipeline:
5
- * 1. Hit valueserp.com /search with a query. Returns ranked URLs.
5
+ * 1. Hit infernetprotocol.com/api/v1/search with a query (Nostr-signed).
6
+ * The control plane proxies to the real provider, holds the upstream
7
+ * key, and enforces a per-pubkey daily quota.
6
8
  * 2. For each URL: fetch HTML, strip nav/script/style, extract main text.
7
9
  * 3. For each page: chunk into Q/A pairs (heading-based) and emit JSONL.
8
10
  *
9
11
  * The output JSONL uses the chatml `messages` shape, ready to feed into
10
12
  * `infernet train run` or any HF/Unsloth/Axolotl pipeline.
11
13
  *
12
- * VALUESERP_API_KEY must be set in env or `~/.config/infernet/config.json`
13
- * under integrations.valueserp.api_key.
14
+ * Self-host escape hatch: set INFERNET_DIRECT_VALUESERP=1 with a
15
+ * VALUESERP_API_KEY in env (or `integrations.valueserp.api_key` in config)
16
+ * to bypass the control plane and call valueserp directly.
14
17
  */
15
18
 
16
19
  import fs from "node:fs/promises";
17
20
  import path from "node:path";
21
+ import { signRequest, AUTH_HEADER } from "@infernetprotocol/auth";
18
22
  import { loadConfig } from "./config.js";
19
23
 
24
+ const SEARCH_PATH = "/api/v1/search";
25
+
26
+ /** Resolve a direct VALUESERP_API_KEY for the escape hatch. */
20
27
  export async function resolveValueSerpKey() {
21
28
  if (process.env.VALUESERP_API_KEY) return process.env.VALUESERP_API_KEY;
22
29
  const cfg = await loadConfig().catch(() => null);
@@ -26,12 +33,88 @@ export async function resolveValueSerpKey() {
26
33
  }
27
34
 
28
35
  /**
29
- * Search the web via ValueSerp and return the ranked list of result URLs.
36
+ * Search via the control plane (default path). Auth is the node's existing
37
+ * Nostr keypair — operators don't need a separate API key.
38
+ *
30
39
  * @param {object} opts
31
40
  * @param {string} opts.query
32
- * @param {string} opts.apiKey
41
+ * @param {string} opts.controlPlaneUrl
42
+ * @param {string} opts.publicKey - 64-hex Nostr pubkey
43
+ * @param {string} opts.privateKey - 64-hex Nostr privkey
33
44
  * @param {number} [opts.num=20]
34
45
  * @param {string} [opts.location]
46
+ * @param {string[]} [opts.domains]
47
+ * @param {number} [opts.timeoutMs=25000]
48
+ */
49
+ export async function infernetSearch({
50
+ query,
51
+ controlPlaneUrl,
52
+ publicKey,
53
+ privateKey,
54
+ num = 20,
55
+ location,
56
+ domains,
57
+ timeoutMs = 25_000
58
+ } = {}) {
59
+ if (!query) throw new Error("search: query required");
60
+ if (!controlPlaneUrl) throw new Error("search: controlPlaneUrl required");
61
+ if (!publicKey || !privateKey) throw new Error("search: Nostr keypair required (run `infernet init`)");
62
+
63
+ const base = controlPlaneUrl.replace(/\/+$/, "");
64
+ const bodyObj = { query, num };
65
+ if (location) bodyObj.location = location;
66
+ if (domains?.length) bodyObj.domains = domains;
67
+ const bodyText = JSON.stringify(bodyObj);
68
+
69
+ const { header } = signRequest({
70
+ method: "POST",
71
+ path: SEARCH_PATH,
72
+ body: bodyText,
73
+ publicKey,
74
+ privateKey
75
+ });
76
+
77
+ const ctrl = new AbortController();
78
+ const t = setTimeout(() => ctrl.abort(), timeoutMs);
79
+ let res;
80
+ try {
81
+ res = await fetch(base + SEARCH_PATH, {
82
+ method: "POST",
83
+ headers: {
84
+ "content-type": "application/json",
85
+ [AUTH_HEADER]: header
86
+ },
87
+ body: bodyText,
88
+ signal: ctrl.signal
89
+ });
90
+ } finally {
91
+ clearTimeout(t);
92
+ }
93
+
94
+ const text = await res.text();
95
+ let payload = null;
96
+ if (text) {
97
+ try { payload = JSON.parse(text); } catch { payload = { raw: text }; }
98
+ }
99
+
100
+ if (!res.ok) {
101
+ const msg = payload?.error ?? `HTTP ${res.status}`;
102
+ const err = new Error(`search failed: ${msg}`);
103
+ err.status = res.status;
104
+ err.body = payload;
105
+ throw err;
106
+ }
107
+ return {
108
+ results: payload?.data ?? [],
109
+ remaining: payload?.remaining,
110
+ limit: payload?.limit,
111
+ resetAt: payload?.reset_at
112
+ };
113
+ }
114
+
115
+ /**
116
+ * Direct ValueSerp call — only used when INFERNET_DIRECT_VALUESERP=1.
117
+ * Kept for self-hosters who don't want to depend on the control plane.
35
118
  */
36
119
  export async function valueSerpSearch({ query, apiKey, num = 20, location } = {}) {
37
120
  if (!query) throw new Error("valueserp: query required");
@@ -142,11 +225,15 @@ export function paragraphsToChatml({ url, title, text, query, minChars = 200, ma
142
225
  * Top-level helper: search + crawl + chunk → write training JSONL.
143
226
  * Used by the `infernet train data` CLI surface.
144
227
  *
145
- * @returns {{ urls: number, examples: number, outPath: string }}
228
+ * Pass `search: { mode: "control-plane", controlPlaneUrl, publicKey, privateKey }`
229
+ * to route through the platform proxy (default), or
230
+ * `search: { mode: "direct", apiKey }` to call the upstream provider yourself.
231
+ *
232
+ * @returns {{ urls: number, examples: number, outPath: string, quota?: object }}
146
233
  */
147
234
  export async function buildTrainingDataset({
148
235
  query,
149
- apiKey,
236
+ search,
150
237
  outPath,
151
238
  num = 20,
152
239
  minChars = 200,
@@ -154,7 +241,29 @@ export async function buildTrainingDataset({
154
241
  domains = null, // optional whitelist (e.g. ["svelte.dev", "github.com"])
155
242
  onProgress = () => {}
156
243
  }) {
157
- const results = await valueSerpSearch({ query, apiKey, num });
244
+ if (!search?.mode) throw new Error("buildTrainingDataset: search.mode is required");
245
+
246
+ let results;
247
+ let quota = null;
248
+ if (search.mode === "direct") {
249
+ results = await valueSerpSearch({ query, apiKey: search.apiKey, num });
250
+ } else if (search.mode === "control-plane") {
251
+ // Server applies the domain filter, but we still pass it so the
252
+ // server can short-circuit irrelevant results before billing them.
253
+ const out = await infernetSearch({
254
+ query,
255
+ num,
256
+ domains,
257
+ controlPlaneUrl: search.controlPlaneUrl,
258
+ publicKey: search.publicKey,
259
+ privateKey: search.privateKey
260
+ });
261
+ results = out.results;
262
+ quota = { remaining: out.remaining, limit: out.limit, resetAt: out.resetAt };
263
+ } else {
264
+ throw new Error(`buildTrainingDataset: unknown search.mode "${search.mode}"`);
265
+ }
266
+
158
267
  const filtered = domains
159
268
  ? results.filter((r) => domains.some((d) => r.domain?.includes(d)))
160
269
  : results;
@@ -179,5 +288,5 @@ export async function buildTrainingDataset({
179
288
  }
180
289
  }
181
290
  await out.close();
182
- return { urls: urlsUsed, examples: totalExamples, outPath };
291
+ return { urls: urlsUsed, examples: totalExamples, outPath, quota };
183
292
  }
package/lib/hf-model.js CHANGED
@@ -7,6 +7,7 @@
7
7
  */
8
8
 
9
9
  import { spawn } from 'node:child_process';
10
+ import { existsSync } from 'node:fs';
10
11
  import { loadConfig } from './config.js';
11
12
 
12
13
  const HF_API = 'https://huggingface.co/api';
@@ -123,68 +124,43 @@ export async function fetchHfModelInfo(repoId, token) {
123
124
  };
124
125
  }
125
126
 
126
- /** Check whether huggingface-cli (Python) is available. */
127
- async function hfCliAvailable() {
128
- return new Promise(resolve => {
129
- const p = spawn('huggingface-cli', ['--help'], { stdio: 'ignore' });
130
- p.on('exit', code => resolve(code === 0));
131
- p.on('error', () => resolve(false));
132
- });
133
- }
134
-
135
- /** Check whether Python huggingface_hub is importable. */
136
- async function hfHubAvailable() {
137
- return new Promise(resolve => {
138
- const p = spawn('python3', ['-c', 'import huggingface_hub'], { stdio: 'ignore' });
139
- p.on('exit', code => resolve(code === 0));
140
- p.on('error', () => resolve(false));
141
- });
127
+ /** The `hf` CLI the installer puts in the vLLM venv. One place, no fallbacks. */
128
+ function hfCli() {
129
+ const home = process.env.HOME || process.env.USERPROFILE || '';
130
+ const base = process.env.INFERNET_HOME || `${home}/.infernet`;
131
+ return `${base}/vllm-venv/bin/hf`;
142
132
  }
143
133
 
144
134
  /**
145
- * Download a model from HuggingFace using huggingface-cli.
146
- * Streams CLI output directly to the terminal (progress bars etc.).
147
- * Returns the local cache path.
135
+ * Download a HuggingFace model via `hf download`. Streams progress and returns
136
+ * the local cache path; surfaces the CLI's stderr on failure.
148
137
  */
149
138
  export async function downloadHfModel(repoId, token, { localDir = null } = {}) {
150
- const hasCli = await hfCliAvailable();
151
- const hasHub = hasCli || await hfHubAvailable();
152
-
153
- if (!hasHub) {
139
+ const cli = hfCli();
140
+ if (!existsSync(cli)) {
154
141
  throw new Error(
155
- 'huggingface_hub not installed.\n' +
156
- 'Fix: pip install -U huggingface_hub\n' +
157
- ' (or: pip3 install -U huggingface_hub)'
142
+ `HuggingFace CLI not found at ${cli}. Re-run the installer on this NVIDIA ` +
143
+ `host — it installs vLLM + huggingface_hub.`
158
144
  );
159
145
  }
160
146
 
161
- // Build the download command
162
147
  const args = ['download', repoId];
163
148
  if (token) args.push('--token', token);
164
149
  if (localDir) args.push('--local-dir', localDir);
165
150
 
166
- return new Promise((resolve, reject) => {
167
- const cmd = hasCli ? 'huggingface-cli' : 'python3';
168
- const finalArgs = hasCli ? args : [
169
- '-c',
170
- `from huggingface_hub import snapshot_download; ` +
171
- `p = snapshot_download(${JSON.stringify(repoId)}` +
172
- `${token ? `, token=${JSON.stringify(token)}` : ''}` +
173
- `${localDir ? `, local_dir=${JSON.stringify(localDir)}` : ''}` +
174
- `); print(p)`
175
- ];
176
-
177
- process.stdout.write(`\nDownloading ${repoId} from HuggingFace...\n`);
178
- if (!hasCli) {
179
- process.stdout.write('(tip: pip install huggingface_hub for a nicer progress bar)\n\n');
180
- }
181
-
182
- const child = spawn(cmd, finalArgs, { stdio: ['ignore', 'inherit', 'inherit'] });
183
- let localPath = localDir;
151
+ process.stdout.write(`\nDownloading ${repoId} from HuggingFace…\n`);
184
152
 
153
+ return new Promise((resolve, reject) => {
154
+ const child = spawn(cli, args, { stdio: ['ignore', 'inherit', 'pipe'] });
155
+ let errTail = '';
156
+ child.stderr.on('data', (d) => {
157
+ const s = d.toString();
158
+ process.stderr.write(s);
159
+ errTail = (errTail + s).slice(-1200);
160
+ });
185
161
  child.on('exit', code => {
186
- if (code === 0) resolve(localPath ?? defaultCachePath(repoId));
187
- else reject(new Error(`huggingface download exited ${code}`));
162
+ if (code === 0) resolve(localDir ?? defaultCachePath(repoId));
163
+ else reject(new Error(`hf download exited ${code}${errTail.trim() ? `:\n${errTail.trim()}` : ''}`));
188
164
  });
189
165
  child.on('error', reject);
190
166
  });
@@ -0,0 +1,81 @@
1
+ /**
2
+ * Shared `~/.infernet/inference/state.json` reader/writer. The file
3
+ * is the single source of truth between `infernet inference …`
4
+ * subcommands and the long-running daemon (which reads it on every
5
+ * heartbeat to assemble specs.rpc / specs.rpc_primary).
6
+ *
7
+ * Shape (all fields optional — the daemon assembles only what's
8
+ * present):
9
+ *
10
+ * {
11
+ * "rpc_slice": { models: [...], host, port, pid, started_at,
12
+ * binary, gpu, vram_gb, ram_gb, max_concurrent },
13
+ * "rpc_primary": { models: [...], gguf_paths: { <model_id>: <path> },
14
+ * version }
15
+ * }
16
+ *
17
+ * Atomic writes via temp-file rename so concurrent subcommands +
18
+ * daemon heartbeats can't see a torn file. Reads tolerate a missing
19
+ * file (returns {}).
20
+ */
21
+
22
+ import fs from 'node:fs';
23
+ import fsp from 'node:fs/promises';
24
+ import path from 'node:path';
25
+
26
+ const STATE_DIR = path.join(process.env.HOME ?? '/tmp', '.infernet', 'inference');
27
+ const STATE_FILE = path.join(STATE_DIR, 'state.json');
28
+
29
+ export function inferenceStateDir() { return STATE_DIR; }
30
+ export function inferenceStateFile() { return STATE_FILE; }
31
+
32
+ export async function readInferenceState() {
33
+ try {
34
+ const raw = await fsp.readFile(STATE_FILE, 'utf8');
35
+ const obj = JSON.parse(raw);
36
+ return obj && typeof obj === 'object' ? obj : {};
37
+ } catch { return {}; }
38
+ }
39
+
40
+ export function readInferenceStateSync() {
41
+ try {
42
+ const raw = fs.readFileSync(STATE_FILE, 'utf8');
43
+ const obj = JSON.parse(raw);
44
+ return obj && typeof obj === 'object' ? obj : {};
45
+ } catch { return {}; }
46
+ }
47
+
48
+ export async function writeInferenceState(next) {
49
+ await fsp.mkdir(STATE_DIR, { recursive: true });
50
+ const tmp = `${STATE_FILE}.${process.pid}.${Date.now()}`;
51
+ await fsp.writeFile(tmp, JSON.stringify(next, null, 2));
52
+ await fsp.rename(tmp, STATE_FILE);
53
+ }
54
+
55
+ /**
56
+ * Read-modify-write helper. The mutator can be sync or async; it
57
+ * receives the current state (always an object) and returns the
58
+ * next state.
59
+ */
60
+ export async function patchInferenceState(mutator) {
61
+ const cur = await readInferenceState();
62
+ const next = await mutator({ ...cur });
63
+ await writeInferenceState(next);
64
+ return next;
65
+ }
66
+
67
+ /**
68
+ * True if `pid` is a live process. `process.kill(pid, 0)` throws
69
+ * ESRCH for dead pids and EPERM for foreign processes — we treat
70
+ * EPERM as "alive but not ours" (unusual but possible after a uid
71
+ * change). Anything else → dead.
72
+ */
73
+ export function isPidAlive(pid) {
74
+ if (!Number.isFinite(pid) || pid <= 0) return false;
75
+ try {
76
+ process.kill(pid, 0);
77
+ return true;
78
+ } catch (err) {
79
+ return err?.code === 'EPERM';
80
+ }
81
+ }
package/lib/prompt.js CHANGED
@@ -17,15 +17,31 @@ import { stdin, stdout } from 'node:process';
17
17
  export async function question(prompt, options = {}) {
18
18
  const { default: defaultValue, secret = false } = options;
19
19
 
20
+ const suffix = defaultValue ? ` [${secret ? '****' : defaultValue}]` : '';
21
+ const full = `${prompt}${suffix}: `;
22
+
23
+ // Non-interactive stdin (piped `curl | sh` install, `--yes` automation,
24
+ // CI, nohup): there's no terminal to read an answer from. Creating a
25
+ // readline with terminal:true and awaiting rl.question() here NEVER
26
+ // resolves at EOF — the await hangs, the event loop drains, and Node
27
+ // exits 0 mid-function, so callers like `init` silently skip saveConfig
28
+ // and write no config at all. Echo the prompt + chosen default and
29
+ // return it immediately instead.
30
+ if (!stdin.isTTY) {
31
+ stdout.write(
32
+ defaultValue !== undefined
33
+ ? `${full}${secret ? '****' : defaultValue} (default, non-interactive)\n`
34
+ : `${full}(non-interactive)\n`
35
+ );
36
+ return defaultValue !== undefined ? defaultValue : '';
37
+ }
38
+
20
39
  const rl = readline.createInterface({
21
40
  input: stdin,
22
41
  output: stdout,
23
42
  terminal: true
24
43
  });
25
44
 
26
- const suffix = defaultValue ? ` [${secret ? '****' : defaultValue}]` : '';
27
- const full = `${prompt}${suffix}: `;
28
-
29
45
  try {
30
46
  let answer;
31
47
  if (secret) {
@@ -41,6 +41,7 @@ export const CATALOG = [
41
41
  { id: "qwen2.5-coder-14b",name:"Qwen2.5 Coder 14B", paramsB: 14, vramMin: 12, ramMin: 32, pullName: "qwen2.5-coder:14b", backend: "ollama", quality: 9, speed: 5, useCases: ["coding","agents"] },
42
42
 
43
43
  // ── 32B (24GB+ VRAM, RTX 4090 / A100 sweet spot) ──
44
+ { id: "qwen3.5-9b-heretic", name: "Qwen3.5 9B Heretic (uncensored)", paramsB: 9, vramMin: 24, ramMin: 32, pullName: "hf:DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-THINKING-HERETIC-UNCENSORED", backend: "vllm", quality: 8, speed: 6, useCases: ["chat","uncensored","vision","agents","study"], uncensored: true },
44
45
  { id: "qwen2.5-32b", name: "Qwen2.5 32B", paramsB: 32, vramMin: 24, ramMin: 48, pullName: "qwen2.5:32b", backend: "ollama", quality: 9, speed: 4, useCases: ["chat","coding","study","agents"], notes: "Single-80GB-GPU sweet spot." },
45
46
  { id: "qwen2.5-coder-32b",name:"Qwen2.5 Coder 32B", paramsB: 32, vramMin: 24, ramMin: 48, pullName: "qwen2.5-coder:32b", backend: "ollama", quality: 10, speed: 4, useCases: ["coding","agents"] },
46
47
 
@@ -49,7 +50,10 @@ export const CATALOG = [
49
50
  { id: "qwen2.5-72b", name: "Qwen2.5 72B", paramsB: 72, vramMin: 40, ramMin: 64, pullName: "qwen2.5:72b", backend: "ollama", quality: 10, speed: 3, useCases: ["chat","coding","study","agents"] },
50
51
 
51
52
  // ── MoE (high VRAM, vLLM only) ──
52
- { id: "mixtral-8x7b", name: "Mixtral 8x7B", paramsB: 47, vramMin: 30, ramMin: 64, pullName: "dolphin-mixtral:8x7b", backend: "ollama", quality: 9, speed: 5, useCases: ["chat","coding","uncensored"], uncensored: true }
53
+ { id: "mixtral-8x7b", name: "Mixtral 8x7B", paramsB: 47, vramMin: 30, ramMin: 64, pullName: "dolphin-mixtral:8x7b", backend: "ollama", quality: 9, speed: 5, useCases: ["chat","coding","uncensored"], uncensored: true },
54
+
55
+ // ── flagship / cluster-tier (multi-GPU or cloud; won't fit a single card) ──
56
+ { id: "glm-5.2-uncensored", name: "GLM-5.2 754B FP8 (uncensored)", paramsB: 754, vramMin: 768, ramMin: 768, pullName: "hf:zandenAI/GLM-5.2-FP8-Uncensored", backend: "vllm", quality: 10, speed: 2, useCases: ["chat","uncensored","cybersecurity","agents"], uncensored: true, gated: true }
53
57
  ];
54
58
 
55
59
  /**