Model
Model explorer
Kimi K3.1
OPENMoonshot AI · Kimi K3 family · released Sep 23, 2026
Moonshot's September 23, 2026 refresh of July's Kimi K3 — same 2.8T LatentMoE open-weights skeleton (Modified MIT), same $3/$15 API, two months of post-training work on the agentic rows. The launch table is conservative: +1-3 points almost everywhere with the biggest moves on Terminal-Bench 3 (17.4 → 21.8) and ExploitBench (32.2 → 36.7). No weights-size change, no price change — an iterate-in-place release rather than a new tier. The K3 family's fast tier (K3-Fast) is untouched.
ReasoningCodingVisionFunction callingTool useAgentic
3027.2
Elo · rank #6
Parameters
2800B
Active params
Undisclosed
Context
1.024M tokens
Architecture
2.8T-parameter LatentMoE reasoning model, open weights; 1.02M-token context
License
Modified MIT License (precedent from the K2.x line)
Languages
—
API price (in/out)
$3 / $15
Modalities
text · vision
Benchmark results
Bar shows position within the tracked field; marker = field best
best: GPT-6 Astra · 94.6%
best: this model · 92.4%
best: GPT-6 Astra · 100.0%
best: Claude Fable 5 · 1932
best: GPT-6 Astra · 96.0%
best: Claude Opus 5 · 64.7%
best: this model · 85.6%
best: Claude Fable 5.1 · 81.2%
best: Gemini 3.8 Flash · 89.4%
best: GPT-5.6 Sol · 34.6%
best: this model · 78.9%
Run it locally
VRAM @ Q4
—
VRAM @ FP16
—
Fits on (Q4)
Multi-node cluster required
Throughput data unavailable.
Quantizations
—
Fine-tune it
PermissiveQLoRA1904.0 GBbeyond 8× B200
LoRA5964.0 GBbeyond 8× B200
Full fine-tune44940.0 GBbeyond 8× B200
QLoRA config exceeds 8× B200 — not trainable on curated hardware.
Kimi K3 family
Elo progression across releases
API price $3/$15 · each benchmark row carries its own source badge (see methodology)