Comparing Kimi K3.1 Pro with Kimi K3 is a rare controlled experiment: the same 2.8T base, the same parameter count, with a single architectural change — 24 Gated DSA layers replacing full-attention Gated MLA layers — and everything else coming from substantially larger post-training compute. The score gap between the generations is the net return on post-training — large, but very unevenly distributed.
First, one architectural change: 24 Gated DeepSeek Sparse Attention (Gated DSA) layers replace K3's full-attention Gated MLA layers; the other 69 Kimi Delta Attention layers are unchanged, and total parameters stay 2.8T with 104B active. Second, substantially larger post-training compute, with joint RL across hundreds of domains. Third, the commercial terms changed: K3 carried a Modified MIT license, while K3.1 Pro ships under the custom Kimi K3.1-Pro License (a separate agreement for MaaS revenue over $20M; products with over 100M MAU or over $20M monthly revenue must display the model name). The first two are technical upgrades; the third is a "regression" — from a standard open license to a custom license with commercial conditions.
Artificial Analysis Intelligence Index v4.3.2: K3 scored 44, K3.1 Pro 55, No. 4 overall. An 11-point jump is post-training's hardest independent evidence — same base, same evaluation set, so the gap can only come from post-training. One caution: on the v5.0 re-version K3.1 Pro scores 46, and K3's v5.0 score was not verified in this research round, so 46 must not be compared with 44 across versions. Agent Arena corroborates: net improvement K3 +0.9%, K3.1 Pro +5.1% (Max); Bash Reliability K3 −4.5%, K3.1 Pro +7.9%. Long-horizon coding: Kimi Code Bench 3.0 (internal), K3 55.95%, K3.1 Pro 61.2%.
The vendor's internal finance benchmarks (self-reported): Modeling 50.4 → 57.6, Research 58.8 → 70.4, Business Operation 8.9 → 37 (~4.2×). OfficeQA Pro v2 (finance) 64.4, SOTA per Moonshot; Frontier Finance 58.6. Hallucination rate (Kimi Hallu Bench Hard, claim-level): with tools 23.4% → 8.3%; without tools K3.1 Pro is 5.4% (GPT-6 Astra 4.2%, Opus 5.5 6.8%). These two sets of numbers show where the post-training money went: finance tasks and factual reliability — exactly the agent and office scenarios Kimi targets.
LMArena Text: K3.1 Pro 1488.0, No. 24; K3 1487.4 — a 0.6-point gap, inside the ±13.0 confidence interval, not significant. On general chat, post-training opened almost no gap. The contrast is worth remembering: AA Index +11, finance Business Operation ~4.2×, hallucination down by nearly two-thirds — all task-specific returns; the general-chat return is near zero. Post-training made it directionally stronger, not uniformly stronger.
xdev-bench-hard pass@3: K3 33.3%, the released K3.1 Pro 81.0% — on the surface a 2.4× jump. But 81.0% is the deliberately weakened figure: Moonshot removed cyber-related training data, and the uncleaned checkpoint scored 90.5%. The real post-training gain (33.3% → 90.5%) is larger than the released number shows, while the 81.0% users actually get is the vendor's downgraded version. Both numbers must be stated together in a generational comparison, or the post-training gain will be understated and the released capability overstated.
K3.1 Pro's standard tier is $2/$10, about a third less than Kimi K3 cost at launch; AA's cost per index task is $2.11, basically level with K3's $2.00. A stronger model at a lower price — uncommon in generational upgrades, and explained by post-training costing less at the margin than training a new base. Transition arrangement: Kimi K3 (kimi-k3) is served by K3.1 Pro until December 7, 2026; whether the K3 model ID continues after that date was not verified in this research round.
For agent and finance tasks: yes — AA Index +11, Business Operation ~4.2×, hallucination down by nearly two-thirds. For general chat: the gain is near zero, with LMArena up only 0.6 points, not significant. Also note the license change: K3's Modified MIT became a custom license with commercial conditions.
No. It is a post-trained product of Kimi K3: the same 2.8T base, with the only architectural change being 24 Gated DSA layers replacing full-attention Gated MLA layers; everything else comes from post-training compute.
Kimi K3 (kimi-k3) is served by K3.1 Pro until December 7, 2026; arrangements after that date were not verified in this research round.
Reference: Moonshot — kimi.com tech blog (internal finance benchmarks, hallucination rate, xdev-bench) · Artificial Analysis — Index v4.3.2 (K3 44, K3.1 Pro 55) · LMArena — Text 1488.0 vs 1487.4 (rewritten and organized by ModelBrief; facts per the official blog and the boards)