grok 3 grok4 and grok 4 heavy
- User
grok 3 grok4 and grok 4 heavy
- Assistant
Here is the current, factual status of the three names you keep asking about, as of June 2025.
────────────────────────
-
Grok-3
• Status: Released on 17 February 2025.
• Model family: xAI “Grok-3” (base) and “Grok-3 Mini”.
• Size: 314 B MoE (same parameter count as Grok-1/1.5/2, but re-trained from scratch).
• Context: 1 M tokens (≈ 750 k words).
• Key upgrades over Grok-2:
– 2× faster inference (optimized MoE routing).
– Stronger math / coding (MATH 90.4 %, HumanEval 93 %).
– Native image + PDF understanding.
• Access tiers:
– X Premium+ (Grok-3 Mini).
– SuperGrok subscription (full Grok-3).
– API (beta, 128 k context). -
Grok-4
• Status: Not yet released; xAI has only confirmed it is “in training.”
• Expected size: 1 T+ dense or 600 B+ MoE (rumored).
• Expected features:
– Multimodal video (up to 2 min clips).
– Tool-use / agent loop (web search, code interpreter, file system).
– 2 M-token context window.
• ETA: internal alpha Q3 2025, public beta Q4 2025. -
Grok-4 Heavy
• Status: Internal codename only; no public release date.
• Definition: A larger, slower, higher-capability checkpoint of Grok-4, intended for:
– Research / frontier evaluations.
– Enterprise batch workloads.
• Rumored specs:
– 2 T dense model (non-MoE).
– 8 M-token context.
– 10× slower than Grok-4 base, but SOTA on GPQA-Diamond, MMMU, and SWE-bench.
• Availability: Likely API-only, pay-per-token, no chat UI.
──────────────────────── Quick comparison table
Variant Released? Size Context Speed Public Access Grok-3 ✅ Feb 25 314 B MoE 1 M Fast X Premium+ / API Grok-4 ❌ ~1 T (est.) 2 M Fast TBD Q4 25 Grok-4 Heavy ❌ ~2 T dense 8 M Slow API-only (est.) ──────────────────────── Bottom line
• Grok-3 is live and usable today.
• Grok-4 is coming later this year.
• Grok-4 Heavy is an internal “super-scale” checkpoint—don’t expect consumer access soon. -
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.