@blockrun/franklin 3.8.21 → 3.8.23

package/dist/agent/compact.js

@@ -427,7 +427,7 @@ function formatCompactSummary(raw) {
 function pickCompactionModel(primaryModel) {
     // Free parent → free compaction (no silent charge)
     if (primaryModel.startsWith('nvidia/') || primaryModel === 'blockrun/free') {
-        return 'nvidia/nemotron-ultra-253b';
+        return 'nvidia/glm-4.7';
     }
     // Use cheapest capable model for summarization to save cost
     // Tier down: opus/pro → sonnet, sonnet → haiku, everything else → flash (cheapest capable)

package/dist/agent/evaluator.js

@@ -168,7 +168,7 @@ export function parseGroundingResponse(raw) {
  *  choice so both quality gates have the same cost profile. Override via
  *  `FRANKLIN_EVALUATOR_MODEL` to experiment with accuracy/cost trade-offs. */
 export function evaluatorModel() {
-    return process.env.FRANKLIN_EVALUATOR_MODEL || 'nvidia/nemotron-ultra-253b';
+    return process.env.FRANKLIN_EVALUATOR_MODEL || 'nvidia/llama-4-maverick';
 }
 // ─── Run grounding check ─────────────────────────────────────────────────
 const MAX_EVAL_TOKENS = 512;

package/dist/agent/intent-prefetch.js

@@ -27,7 +27,9 @@
  */
 import { getStockPrice, getPrice } from '../trading/data.js';
 // ─── Classifier ──────────────────────────────────────────────────────────
-const CLASSIFIER_MODEL = process.env.FRANKLIN_PREFETCH_MODEL || 'nvidia/nemotron-ultra-253b';
+// llama-4-maverick: same rationale as the router classifier — emits plain
+// text under tight max_tokens rather than routing through thinking blocks.
+const CLASSIFIER_MODEL = process.env.FRANKLIN_PREFETCH_MODEL || 'nvidia/llama-4-maverick';
 const CLASSIFIER_TIMEOUT_MS = 2_500;
 const CLASSIFIER_PROMPT = `You extract PREFETCH INTENT from a user message for a CLI agent that has live market-data tools.
 
@@ -98,8 +100,12 @@ export async function classifyIntent(userInput, client) {
     if (process.env.FRANKLIN_NO_PREFETCH === '1')
         return null;
     const trimmed = userInput.trim();
-    // Short inputs (<12 chars) are rarely asking for market data — skip the call entirely.
-    if (trimmed.length < 12)
+    // Only the cheapest gate — skip very short inputs that can't be a real
+    // market question ("hi", "ok", "thanks"). 6 chars covers those while
+    // still letting short-form Chinese / ticker prompts through, e.g.
+    // "BTC 价格" (6), "CRCL 多少" (7). Longer prompts all route to the LLM
+    // classifier, which decides NONE cheaply when not market-related.
+    if (trimmed.length < 6)
         return null;
     const ctrl = new AbortController();
     const timer = setTimeout(() => ctrl.abort(), CLASSIFIER_TIMEOUT_MS);

package/dist/agent/llm.js

@@ -205,12 +205,12 @@ export class ModelClient {
         // Static fallback if router is unavailable. Default to FREE model so
         // users aren't silently charged when their intended model can't resolve.
         const FALLBACKS = {
-            'blockrun/auto': 'nvidia/nemotron-ultra-253b',
-            'blockrun/eco': 'nvidia/nemotron-ultra-253b',
+            'blockrun/auto': 'nvidia/glm-4.7',
+            'blockrun/eco': 'nvidia/glm-4.7',
             'blockrun/premium': 'anthropic/claude-sonnet-4.6',
-            'blockrun/free': 'nvidia/nemotron-ultra-253b',
+            'blockrun/free': 'nvidia/glm-4.7',
         };
-        return FALLBACKS[model] || 'nvidia/nemotron-ultra-253b';
+        return FALLBACKS[model] || 'nvidia/glm-4.7';
     }
     async *streamCompletion(request, signal) {
         // Resolve virtual models before any API call

package/dist/agent/loop.js

@@ -823,7 +823,7 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
                 const hasTools = responseParts.some(p => p.type === 'tool_use');
                 const hasThinking = responseParts.some(p => p.type === 'thinking');
                 if (!hasText && !hasTools && !hasThinking) {
-                    const EMPTY_FALLBACK_MODELS = ['nvidia/qwen3-coder-480b', 'nvidia/nemotron-ultra-253b', 'zai/glm-5.1'];
+                    const EMPTY_FALLBACK_MODELS = ['nvidia/qwen3-coder-480b', 'nvidia/glm-4.7', 'zai/glm-5.1'];
                     const nextModel = EMPTY_FALLBACK_MODELS.find(m => m !== config.model && !turnFailedModels.has(m));
                     if (nextModel && recoveryAttempts < 2) {
                         recoveryAttempts++;
@@ -922,7 +922,7 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
                     if (lastRoutedCategory) {
                         recordOutcome(lastRoutedCategory, config.model, 'payment');
                     }
-                    const FREE_MODELS = ['nvidia/qwen3-coder-480b', 'nvidia/nemotron-ultra-253b', 'nvidia/devstral-2-123b'];
+                    const FREE_MODELS = ['nvidia/glm-4.7', 'nvidia/qwen3-coder-480b', 'nvidia/llama-4-maverick', 'nvidia/qwen3-next-80b-a3b-thinking'];
                     const nextFree = FREE_MODELS.find(m => !turnFailedModels.has(m));
                     if (nextFree) {
                         const oldModel = config.model;

package/dist/agent/verification.js

@@ -113,7 +113,7 @@ export async function runVerification(history, handlers, client, config) {
     ];
     config.onEvent?.({ kind: 'text_delta', text: '\n*Verifying...*\n' });
     // Use cheap model for verification
-    const verificationModel = 'nvidia/nemotron-ultra-253b'; // Free model to keep cost zero
+    const verificationModel = 'nvidia/glm-4.7'; // Free model to keep cost zero
     try {
         // Simple single-turn verification call
         const response = await client.complete({

package/dist/commands/social.js

@@ -159,7 +159,7 @@ async function runCommand(options) {
     const chain = loadChain();
     const apiUrl = API_URLS[chain];
     const appConfig = loadAppConfig();
-    const model = options.model || appConfig['default-model'] || 'nvidia/nemotron-ultra-253b';
+    const model = options.model || appConfig['default-model'] || 'nvidia/glm-4.7';
     console.log(chalk.dim(`  Model: ${model}`));
     console.log('');
     let result;

package/dist/commands/start.js

@@ -124,11 +124,17 @@ export async function startCommand(options) {
         process.exitCode = exitCode;
         return;
     }
-    // Warn when a paid model is active so users know they'll be charged
+    // Warn when a paid model is active so users know they'll be charged.
+    // Set members = BlockRun gateway's current free tier (refreshed 2026-04).
     const FREE_MODELS = new Set([
-        'nvidia/nemotron-ultra-253b',
+        'nvidia/glm-4.7',
+        'nvidia/qwen3-next-80b-a3b-thinking',
         'nvidia/qwen3-coder-480b',
-        'nvidia/devstral-2-123b',
+        'nvidia/mistral-small-4-119b',
+        'nvidia/llama-4-maverick',
+        'nvidia/deepseek-v3.2',
+        'nvidia/gpt-oss-120b',
+        'nvidia/gpt-oss-20b',
         'blockrun/free',
     ]);
     if (!FREE_MODELS.has(model)) {

package/dist/commands/telegram.js

@@ -36,7 +36,7 @@ export async function telegramCommand(opts) {
     // Model: --model flag > config default > free default.
     const model = opts.model ||
         config['default-model'] ||
-        'nvidia/nemotron-ultra-253b';
+        'nvidia/glm-4.7';
     const workingDir = process.cwd();
     const systemInstructions = assembleInstructions(workingDir, model);
     // Resume the most recent session tagged for THIS owner so a process

package/dist/learnings/extractor.js

@@ -9,9 +9,9 @@ import { loadLearnings, mergeLearning, saveLearnings, loadSkills, saveSkill } fr
 // Free models for learning extraction — JSON extraction is simple enough.
 // Ordered by reliability: try the best free model first, fall back to others.
 const EXTRACTION_MODELS = [
-    'nvidia/nemotron-ultra-253b', // Best free model for structured output
+    'nvidia/glm-4.7', // Best free model for structured output
     'nvidia/qwen3-coder-480b', // Strong at JSON tasks
-    'nvidia/devstral-2-123b', // Fallback
+    'nvidia/llama-4-maverick', // Fallback
 ];
 const VALID_CATEGORIES = new Set([
     'language', 'model_preference', 'tool_pattern', 'coding_style',

package/dist/plugin-sdk/workflow.js

@@ -5,7 +5,7 @@
  * Plugins implement Workflow; core orchestrates execution and provides infrastructure.
  */
 export const DEFAULT_MODEL_TIERS = {
-    free: 'nvidia/nemotron-ultra-253b',
-    cheap: 'nvidia/nemotron-ultra-253b', // Was glm-5.1 ($0.001/call). Free by default; opt-in to paid.
+    free: 'nvidia/glm-4.7',
+    cheap: 'nvidia/glm-4.7', // Free by default; opt-in to paid flat-rate via 'zai/glm-5.1'.
     premium: 'anthropic/claude-sonnet-4.6',
 };

package/dist/pricing.js

@@ -8,18 +8,21 @@ export const MODEL_PRICING = {
     'blockrun/eco': { input: 0.2, output: 1.0 },
     'blockrun/premium': { input: 3.0, output: 15.0 },
     'blockrun/free': { input: 0, output: 0 },
-    // FREE - NVIDIA models
+    // FREE — BlockRun gateway free tier (refreshed 2026-04)
+    'nvidia/glm-4.7': { input: 0, output: 0 },
+    'nvidia/qwen3-next-80b-a3b-thinking': { input: 0, output: 0 },
+    'nvidia/qwen3-coder-480b': { input: 0, output: 0 },
+    'nvidia/mistral-small-4-119b': { input: 0, output: 0 },
+    'nvidia/llama-4-maverick': { input: 0, output: 0 },
+    'nvidia/deepseek-v3.2': { input: 0, output: 0 },
     'nvidia/gpt-oss-120b': { input: 0, output: 0 },
     'nvidia/gpt-oss-20b': { input: 0, output: 0 },
+    // Retired (kept at 0 for legacy session-cost records; gateway no longer serves these).
     'nvidia/nemotron-ultra-253b': { input: 0, output: 0 },
+    'nvidia/devstral-2-123b': { input: 0, output: 0 },
     'nvidia/nemotron-3-super-120b': { input: 0, output: 0 },
     'nvidia/nemotron-super-49b': { input: 0, output: 0 },
-    'nvidia/deepseek-v3.2': { input: 0, output: 0 },
     'nvidia/mistral-large-3-675b': { input: 0, output: 0 },
-    'nvidia/qwen3-coder-480b': { input: 0, output: 0 },
-    'nvidia/devstral-2-123b': { input: 0, output: 0 },
-    'nvidia/glm-4.7': { input: 0, output: 0 },
-    'nvidia/llama-4-maverick': { input: 0, output: 0 },
     // Anthropic
     'anthropic/claude-sonnet-4.6': { input: 3.0, output: 15.0 },
     'anthropic/claude-opus-4.7': { input: 5.0, output: 25.0 },

package/dist/proxy/fallback.js

@@ -19,7 +19,7 @@ export const DEFAULT_FALLBACK_CONFIG = {
     chain: [
         'deepseek/deepseek-chat', // Direct fallback — cheap & reliable
         'google/gemini-2.5-flash', // Fast & capable
-        'nvidia/nemotron-ultra-253b', // Free model as ultimate fallback
+        'nvidia/glm-4.7', // Free model as ultimate fallback
     ],
     retryOn: [429, 500, 502, 503, 504, 529],
     maxRetries: 5,

package/dist/proxy/server.js

@@ -94,13 +94,19 @@ const MODEL_SHORTCUTS = {
     // DeepSeek
     deepseek: 'deepseek/deepseek-chat',
     r1: 'deepseek/deepseek-reasoner',
-    // Free models
-    free: 'nvidia/nemotron-ultra-253b',
-    nemotron: 'nvidia/nemotron-ultra-253b',
+    // Free models (gateway free tier — refreshed 2026-04)
+    free: 'nvidia/glm-4.7',
+    glm4: 'nvidia/glm-4.7',
     'deepseek-free': 'nvidia/deepseek-v3.2',
-    devstral: 'nvidia/devstral-2-123b',
     'qwen-coder': 'nvidia/qwen3-coder-480b',
+    'qwen-think': 'nvidia/qwen3-next-80b-a3b-thinking',
     maverick: 'nvidia/llama-4-maverick',
+    'gpt-oss': 'nvidia/gpt-oss-120b',
+    'gpt-oss-small': 'nvidia/gpt-oss-20b',
+    'mistral-small': 'nvidia/mistral-small-4-119b',
+    // Retired-gateway-model aliases (map to closest current).
+    nemotron: 'nvidia/glm-4.7',
+    devstral: 'nvidia/qwen3-coder-480b',
     // Minimax
     minimax: 'minimax/minimax-m2.7',
     // Others

package/dist/router/index.js

@@ -66,12 +66,12 @@ const AUTO_TIERS = {
 };
 const ECO_TIERS = {
     SIMPLE: {
-        primary: 'nvidia/nemotron-ultra-253b',
+        primary: 'nvidia/glm-4.7',
         fallback: ['nvidia/gpt-oss-120b', 'nvidia/deepseek-v3.2'],
     },
     MEDIUM: {
         primary: 'google/gemini-2.5-flash-lite',
-        fallback: ['nvidia/nemotron-ultra-253b', 'nvidia/qwen3-coder-480b'],
+        fallback: ['nvidia/glm-4.7', 'nvidia/qwen3-coder-480b'],
     },
     COMPLEX: {
         primary: 'google/gemini-2.5-flash-lite',
@@ -79,7 +79,7 @@ const ECO_TIERS = {
     },
     REASONING: {
         primary: 'xai/grok-4-1-fast-reasoning',
-        fallback: ['deepseek/deepseek-reasoner', 'nvidia/nemotron-ultra-253b'],
+        fallback: ['deepseek/deepseek-reasoner', 'nvidia/qwen3-next-80b-a3b-thinking'],
     },
 };
 const PREMIUM_TIERS = {
@@ -283,7 +283,10 @@ function classicRouteRequest(prompt, profile) {
 //   - Exposed via async `routeRequestAsync(prompt, profile, classify?)`. Callers
 //     that can't be async (proxy, LLM-client bootstrap) keep using the sync
 //     `routeRequest`, which silently does keyword-only routing.
-const CLASSIFIER_MODEL = process.env.FRANKLIN_ROUTER_MODEL || 'nvidia/nemotron-ultra-253b';
+// llama-4-maverick: clean one-word classification output. glm-4.7 + qwen-
+// thinking emit reasoning into thinking blocks and leave text empty under
+// tight max_tokens — fine for chat, wrong shape for single-word dispatch.
+const CLASSIFIER_MODEL = process.env.FRANKLIN_ROUTER_MODEL || 'nvidia/llama-4-maverick';
 const CLASSIFIER_TIMEOUT_MS = 2_500;
 const CLASSIFIER_SYSTEM = `You classify a user's message into ONE routing tier for a CLI agent. Reply with EXACTLY ONE WORD from the allowed set. No explanation, no punctuation, no quotes.
 
@@ -395,7 +398,7 @@ export function routeRequest(prompt, profile = 'auto') {
     // Free profile — always use free model
     if (profile === 'free') {
         return {
-            model: 'nvidia/nemotron-ultra-253b',
+            model: 'nvidia/glm-4.7',
             tier: 'SIMPLE',
             confidence: 1.0,
             signals: ['free-profile'],
@@ -467,7 +470,7 @@ export function getFallbackChain(tier, profile = 'auto') {
             tierConfigs = PREMIUM_TIERS;
             break;
         case 'free':
-            return ['nvidia/nemotron-ultra-253b'];
+            return ['nvidia/glm-4.7'];
         default:
             tierConfigs = AUTO_TIERS;
     }

package/dist/tools/moa.js

@@ -14,13 +14,14 @@ import { ModelClient } from '../agent/llm.js';
 // ─── Configuration ────────────────────────────────────────────────────────
 /** Reference models — diverse, cheap/free models for parallel queries. */
 const REFERENCE_MODELS = [
-    'nvidia/nemotron-ultra-253b', // Free, strong reasoning
+    'nvidia/glm-4.7', // Free, strong reasoning + coding
+    'nvidia/qwen3-next-80b-a3b-thinking', // Free, explicit reasoning model
     'nvidia/qwen3-coder-480b', // Free, strong coding
     'google/gemini-2.5-flash', // Fast, cheap
     'deepseek/deepseek-chat', // Cheap, good reasoning
 ];
 /** Aggregator model — free by default. Users explicitly pass `aggregator` to upgrade. */
-const AGGREGATOR_MODEL = 'nvidia/nemotron-ultra-253b';
+const AGGREGATOR_MODEL = 'nvidia/glm-4.7';
 /** Max tokens per reference response. */
 const REFERENCE_MAX_TOKENS = 4096;
 /** Max tokens for aggregator. */

package/dist/tools/subagent.js

@@ -18,7 +18,7 @@ async function execute(input, ctx) {
         return { output: 'Error: prompt is required', isError: true };
     }
     // Resolve which model the sub-agent will actually run on
-    const subModel = model || registeredParentModel || 'nvidia/nemotron-ultra-253b';
+    const subModel = model || registeredParentModel || 'nvidia/glm-4.7';
     // Cost gate: if parent is free but sub-agent wants paid, ask user first.
     // Prevents silent charges when the agent decides to spawn a more capable sub-agent.
     if (isFreeModel(registeredParentModel) && !isFreeModel(subModel)) {

package/dist/ui/model-picker.js

@@ -46,13 +46,21 @@ export const MODEL_SHORTCUTS = {
     // DeepSeek
     deepseek: 'deepseek/deepseek-chat',
     r1: 'deepseek/deepseek-reasoner',
-    // Free
-    free: 'nvidia/nemotron-ultra-253b',
-    nemotron: 'nvidia/nemotron-ultra-253b',
+    // Free (BlockRun gateway free tier — refreshed 2026-04)
+    free: 'nvidia/glm-4.7',
+    glm4: 'nvidia/glm-4.7',
     'deepseek-free': 'nvidia/deepseek-v3.2',
-    devstral: 'nvidia/devstral-2-123b',
     'qwen-coder': 'nvidia/qwen3-coder-480b',
+    'qwen-think': 'nvidia/qwen3-next-80b-a3b-thinking',
     maverick: 'nvidia/llama-4-maverick',
+    'gpt-oss': 'nvidia/gpt-oss-120b',
+    'gpt-oss-small': 'nvidia/gpt-oss-20b',
+    'mistral-small': 'nvidia/mistral-small-4-119b',
+    // Backward-compatibility aliases for models the gateway retired.
+    // Map to the closest current free model so old session records + user
+    // muscle memory keep working.
+    nemotron: 'nvidia/glm-4.7',
+    devstral: 'nvidia/qwen3-coder-480b',
     // Others
     minimax: 'minimax/minimax-m2.7',
     glm: 'zai/glm-5.1',
@@ -134,12 +142,14 @@ export const PICKER_CATEGORIES = [
     {
         category: '🆓 Free (no USDC needed)',
         models: [
-            { id: 'nvidia/nemotron-ultra-253b', shortcut: 'free', label: 'Nemotron Ultra 253B', price: 'FREE' },
+            { id: 'nvidia/glm-4.7', shortcut: 'free', label: 'GLM-4.7', price: 'FREE' },
+            { id: 'nvidia/qwen3-next-80b-a3b-thinking', shortcut: 'qwen-think', label: 'Qwen3-Next 80B Thinking', price: 'FREE' },
             { id: 'nvidia/qwen3-coder-480b', shortcut: 'qwen-coder', label: 'Qwen3 Coder 480B', price: 'FREE' },
-            { id: 'nvidia/devstral-2-123b', shortcut: 'devstral', label: 'Devstral 2 123B', price: 'FREE' },
             { id: 'nvidia/llama-4-maverick', shortcut: 'maverick', label: 'Llama 4 Maverick', price: 'FREE' },
             { id: 'nvidia/deepseek-v3.2', shortcut: 'deepseek-free', label: 'DeepSeek V3.2', price: 'FREE' },
-            { id: 'nvidia/gpt-oss-120b', shortcut: 'gpt-oss', label: 'GPT OSS 120B', price: 'FREE' },
+            { id: 'nvidia/gpt-oss-120b', shortcut: 'gpt-oss', label: 'GPT-OSS 120B', price: 'FREE' },
+            { id: 'nvidia/gpt-oss-20b', shortcut: 'gpt-oss-small', label: 'GPT-OSS 20B', price: 'FREE' },
+            { id: 'nvidia/mistral-small-4-119b', shortcut: 'mistral-small', label: 'Mistral Small 4 119B', price: 'FREE' },
         ],
     },
 ];

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/franklin",
-  "version": "3.8.21",
+  "version": "3.8.23",
   "description": "Franklin — The AI agent with a wallet. Spends USDC autonomously to get real work done. Pay per action, no subscriptions.",
   "type": "module",
   "exports": {
@@ -20,6 +20,7 @@
   },
   "files": [
     "dist",
+    "assets",
     "README.md",
     "LICENSE"
   ],