SemiAnalysisAI · functionstackx · Jun 13, 2026 · Jun 13, 2026
@@ -71,6 +71,25 @@ Present what you inferred and get confirmation + category in a single step. Incl
 
 Everything else (`MODEL_OPTIONS`, `DEFAULT_MODELS`, `EXPERIMENTAL_MODELS`, `DEPRECATED_MODELS`, `MODEL_PREFIX_MAPPING`, `getModelLabel()`) is derived automatically.
 
+**`packages/app/src/lib/compare-slug.ts`** (easy to miss — the /compare and /compare-per-dollar pages do NOT derive from `MODEL_CONFIG`):
+
+- `COMPARE_MODEL_SLUGS` — add an entry with `{ slug, displayName, dbKeys, label }`. `displayName` must match the `Model` enum value; `dbKeys` lists the DB buckets to query. Place it per the ordering comment (Chinese-lab flagships first, newer family member leads). Without this entry the model is absent from /compare, /compare-per-dollar, the sitemap, and their OG images.
+- `COMPARE_MODEL_ALIASES` — only if a family-level or older-version slug should 308 to the new entry.
+
+**`packages/app/src/lib/compare-ssr.ts`**:
+
+- `KNOWN_MODELS` — add the display name so `?g_model=` URL overrides validate on compare pages.
+
+**`packages/app/src/app/compare/page.tsx`** and **`packages/app/src/app/compare-per-dollar/page.tsx`**:
+
+- `DESCRIPTION` — these SEO meta strings hardcode a sample model list ("…, Qwen 3.5 397B-A17B, and more"). Add the new model if it should appear in the catalog blurb.
+
+**`packages/app/src/lib/model-architectures.ts`** (optional — powers the per-model architecture diagram on the inference tab):
+
+- `MODEL_ARCHITECTURES` — add a `[Model.X]` entry with verified config.json values. Omitted models simply render no diagram (`getModelArchitecture` returns `undefined`), so this is non-blocking but expected for parity with other models.
+
+`/about` needs no change — its model list derives from `DB_MODEL_TO_DISPLAY` and includes the new key automatically once `models.ts` is updated.
+
 ---
 
 ## Featuring a Day-0 Model

@@ -241,6 +241,43 @@ describe('Model Architecture Diagram', () => {
     });
   });
 
+  describe('Collapsible Transformer Blocks (MoE model - MiniMax M3)', () => {
+    before(() => {
+      // Clear any stale Radix scroll lock from prior Select interactions
+      cy.document().then((doc) => {
+        delete doc.body.dataset.scrollLocked;
+        doc.body.style.removeProperty('pointer-events');
+      });
+      cy.get('[role="combobox"]').filter(':visible').first().click();
+      cy.get('[role="option"]').contains('MiniMax M3').click();
+
+      cy.get('[data-testid="model-architecture-toggle"]').should('be.visible');
+      cy.get('body').then(($body) => {
+        if ($body.find('[data-testid="model-architecture-svg"]:visible').length === 0) {
+          cy.get('[data-testid="model-architecture-toggle"]').click();
+        }
+      });
+      cy.get('[data-testid="model-architecture-svg"]').should('be.visible');
+    });
+
+    it('shows MoE and GQA badges for MiniMax M3', () => {
+      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'MoE');
+      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'GQA');
+      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', '428B');
+    });
+
+    it('GQA attention is NOT expandable (sparse attention rendered as a static block)', () => {
+      cy.get('[data-testid="expand-transformer"]').click({ force: true });
+      cy.get('[data-testid="expand-attention"]').should('not.exist');
+      cy.get('[data-testid="expand-experts"]').should('exist');
+    });
+
+    it('shows MiniMax M3 sparse-attention features', () => {
+      cy.contains('MiniMax Sparse Attention (MSA)').should('be.visible');
+      cy.contains('GQA with QK Norm').should('be.visible');
+    });
+  });
+
   describe('Alternating Attention Blocks (MoE model - gpt-oss 120B)', () => {
     before(() => {
       // Clear any stale Radix scroll lock from prior Select interactions

@@ -12,7 +12,7 @@ import { bucketComparePairsByVendor, formatModelList } from '@/lib/compare-ssr';
 export const dynamic = 'force-dynamic';
 
 const DESCRIPTION =
-  'GPU performance per dollar — head-to-head cost per million tokens across every model and hardware pair we benchmark. Performance normalized by owning-hyperscaler TCO for DeepSeek V4 Pro 1.6T, DeepSeek R1, Kimi K2.5/K2.6/K2.7-Code 1T, GLM 5/5.1, Qwen 3.5 397B-A17B, and more. Pick the cheapest SKU for your workload.';
+  'GPU performance per dollar — head-to-head cost per million tokens across every model and hardware pair we benchmark. Performance normalized by owning-hyperscaler TCO for DeepSeek V4 Pro 1.6T, DeepSeek R1, Kimi K2.5/K2.6/K2.7-Code 1T, MiniMax M3 428B, GLM 5/5.1, Qwen 3.5 397B-A17B, and more. Pick the cheapest SKU for your workload.';
 
 export const metadata: Metadata = {
   title: 'GPU Performance per Dollar',

@@ -13,7 +13,7 @@ import { bucketComparePairsByVendor, formatModelList } from '@/lib/compare-ssr';
 export const dynamic = 'force-dynamic';
 
 const DESCRIPTION =
-  'Browse head-to-head GPU inference benchmark comparisons across every model and hardware pair we test. Latency, throughput, and cost for DeepSeek V4 Pro 1.6T, DeepSeek R1, Kimi K2.5/K2.6/K2.7-Code 1T, GLM 5/5.1, Qwen 3.5 397B-A17B, and more.';
+  'Browse head-to-head GPU inference benchmark comparisons across every model and hardware pair we test. Latency, throughput, and cost for DeepSeek V4 Pro 1.6T, DeepSeek R1, Kimi K2.5/K2.6/K2.7-Code 1T, MiniMax M3 428B, GLM 5/5.1, Qwen 3.5 397B-A17B, and more.';
 
 export const metadata: Metadata = {
   title: 'GPU Comparisons',

@@ -43,6 +43,15 @@ describe('parseCompareSlug — new model-prefixed form', () => {
     expect(parsed?.b).toBe('gb200');
   });
 
+  it('parses the minimax-m3 slug as its own model, distinct from minimax-m27', () => {
+    const parsed = parseCompareSlug('minimax-m3-h100-vs-h200');
+    expect(parsed?.model.slug).toBe('minimax-m3');
+    expect(parsed?.model.dbKeys).toEqual(['minimaxm3']);
+    expect(parsed?.a).toBe('h100');
+    expect(parsed?.b).toBe('h200');
+    expect(parsed?.isAliasModel).toBe(false);
+  });
+
   it('preserves non-canonical GPU order so caller can redirect', () => {
     const parsed = parseCompareSlug('kimi-k26-h200-vs-h100');
     expect(parsed?.a).toBe('h200');
@@ -259,6 +268,11 @@ describe('getCompareModelBySlug', () => {
     expect(getCompareModelBySlug('glm-5')).toBe(GLM_51);
   });
 
+  it('keeps the bare minimax alias on the M2 series, with minimax-m3 canonical', () => {
+    expect(getCompareModelBySlug('minimax')?.slug).toBe('minimax-m27');
+    expect(getCompareModelBySlug('minimax-m3')?.slug).toBe('minimax-m3');
+  });
+
   it('returns null for unknown slugs', () => {
     expect(getCompareModelBySlug('nonexistent')).toBeNull();
   });

@@ -31,10 +31,12 @@ export interface CompareModelSlug {
 }
 
 // Order matches the master /compare and /compare-per-dollar index display:
-// DeepSeek V4 Pro → R1 → Kimi → GLM → MiniMax → Qwen → gpt-oss → Llama 70B.
-// Per product spec — flagship Chinese-developed models first, smaller open
-// US-developed models at the bottom. Qwen sits between MiniMax and gpt-oss to
-// keep the Chinese-lab cluster contiguous before the US transition.
+// DeepSeek V4 Pro → R1 → Kimi → GLM → MiniMax M3 → MiniMax M2 → Qwen →
+// gpt-oss → Llama 70B. Per product spec — flagship Chinese-developed models
+// first, smaller open US-developed models at the bottom. Qwen sits between
+// MiniMax and gpt-oss to keep the Chinese-lab cluster contiguous before the
+// US transition. The two MiniMax entries stay adjacent with the newer M3
+// flagship leading the older M2 series.
 export const COMPARE_MODEL_SLUGS: CompareModelSlug[] = [
   {
     slug: 'deepseek-v4',
@@ -69,6 +71,15 @@ export const COMPARE_MODEL_SLUGS: CompareModelSlug[] = [
     dbKeys: ['glm5.1', 'glm5'],
     label: 'GLM 5/5.1',
   },
+  {
+    slug: 'minimax-m3',
+    displayName: 'MiniMax-M3',
+    // M3 is a new 428B architecture (MiniMax Sparse Attention), not a point
+    // release of the M2 series, so it gets its own slug and dbKey rather than
+    // joining the minimax-m27 group.
+    dbKeys: ['minimaxm3'],
+    label: 'MiniMax M3 428B',
+  },
   {
     slug: 'minimax-m27',
     displayName: 'MiniMax-M2.5',

@@ -68,6 +68,7 @@ export const KNOWN_MODELS = new Set([
   'Qwen-3.5-397B-A17B',
   'Kimi-K2.5',
   'MiniMax-M2.5',
+  'MiniMax-M3',
   'GLM-5',
   'DeepSeek-V4-Pro',
 ]);

@@ -25,6 +25,7 @@ describe('MODEL_ARCHITECTURES', () => {
       Model.GptOss,
       Model.Kimi_K2_5,
       Model.MiniMax_M2_5,
+      Model.MiniMax_M3,
     ];
 
     for (const model of models) {
@@ -276,6 +277,29 @@ describe('getModelArchitecture', () => {
     expect(arch?.denseFFNDim).toBeUndefined();
   });
 
+  it('returns architecture for MiniMax M3 with MoE, sparse attention, and shared expert', () => {
+    const arch = getModelArchitecture(Model.MiniMax_M3);
+    expect(arch).toBeDefined();
+    expect(arch?.totalParams).toBe(428);
+    expect(arch?.activeParams).toBe(23);
+    expect(arch?.architectureType).toBe('moe');
+    expect(arch?.attentionType).toBe('GQA');
+    expect(arch?.attentionExpandable).toBe(false);
+    expect(arch?.numLayers).toBe(60);
+    expect(arch?.hiddenSize).toBe(6144);
+    expect(arch?.numHeads).toBe(64);
+    expect(arch?.numKVHeads).toBe(4);
+    expect(arch?.headDim).toBe(128);
+    expect(arch?.ffnDim).toBe(3072);
+    expect(arch?.numExperts).toBe(129);
+    expect(arch?.activeExperts).toBe(4);
+    expect(arch?.hasSharedExpert).toBe(true);
+    expect(arch?.contextWindow).toBe(1048576);
+    expect(arch?.vocabSize).toBe(200064);
+    expect(arch?.developer).toBe('MiniMax');
+    expect(arch?.sourceUrl).toBe('https://huggingface.co/MiniMaxAI/MiniMax-M3');
+  });
+
   it('returns architecture for gpt-oss 120B with MoE, alternating attention, and sink tokens', () => {
     const arch = getModelArchitecture(Model.GptOss);
     expect(arch).toBeDefined();

@@ -113,6 +113,7 @@ export interface ModelArchitecture {
  * - https://huggingface.co/moonshotai/Kimi-K2.5/blob/main/config.json
  * - https://huggingface.co/openai/gpt-oss-120b/blob/main/config.json
  * - https://huggingface.co/MiniMaxAI/MiniMax-M2/blob/main/config.json
+ * - https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/main/config.json
  */
 export const MODEL_ARCHITECTURES: Partial<Record<Model, ModelArchitecture>> = {
   [Model.DeepSeek_R1]: {
@@ -336,6 +337,39 @@ export const MODEL_ARCHITECTURES: Partial<Record<Model, ModelArchitecture>> = {
     developer: 'MiniMax',
     sourceUrl: 'https://huggingface.co/MiniMaxAI/MiniMax-M2',
   },
+  [Model.MiniMax_M3]: {
+    model: Model.MiniMax_M3,
+    totalParams: 428,
+    activeParams: 23,
+    architectureType: 'moe',
+    // MiniMax Sparse Attention (MSA) is built on a GQA projection layout
+    // (64 Q / 4 KV heads) with sparse KV selection layered on top. Render it as
+    // a static attention block, not the standard GQA Q/K/V drill-down — same
+    // treatment as the M2.5 entry.
+    attentionType: 'GQA',
+    attentionExpandable: false,
+    numLayers: 60,
+    hiddenSize: 6144,
+    numHeads: 64,
+    numKVHeads: 4,
+    headDim: 128,
+    vocabSize: 200064,
+    ffnDim: 3072, // moe_intermediate_size (per-expert FFN)
+    numExperts: 129, // 128 routed + 1 shared
+    activeExperts: 4,
+    hasSharedExpert: true,
+    contextWindow: 1048576, // 1M
+    features: [
+      'MiniMax Sparse Attention (MSA)',
+      'GQA with QK Norm',
+      'Partial RoPE (rotary factor 0.5)',
+      'SwiGLU FFN',
+      'Native Multimodality (text/image/video)',
+      'MoE (128 routed + 1 shared experts, 4 active)',
+    ],
+    developer: 'MiniMax',
+    sourceUrl: 'https://huggingface.co/MiniMaxAI/MiniMax-M3',
+  },
 };
 
 /**

@@ -92,6 +92,8 @@ export const MODEL_TO_KEY: Record<string, string> = {
   'moonshotai/Kimi-K2.7-Code': 'kimik2.7-code',
   // MiniMax-M2.5
   'MiniMaxAI/MiniMax-M2.5': 'minimaxm2.5',
+  // MiniMax-M3 (428B, distinct architecture from the M2 series)
+  'MiniMaxAI/MiniMax-M3': 'minimaxm3',
   // GLM-5
   'zai-org/GLM-5-FP8': 'glm5',
   'amd/GLM-5.1-MXFP4': 'glm5.1',