Kimi K3.1 Leaves a Paper Trail: Kibitz on LMArena, SGLang Hardening PRs, and an Engineering Release in the Making
K3.1 exists today only as a rumor with a strong paper trail: no public endpoint, price, or weight repo, so production code cannot be pinned to it. The same-day SGLang work points to where the next release competes: agent tool-call reliability and serving stability, not headline benchmark numbers.
Kimi K3.1 has not launched, but it is leaving fingerprints. Two anonymous LMArena models named Kibitz and Kalmia appeared on 2026-10-22, and community consensus reads Kibitz as Moonshot AI's alias for the next Kimi model. That same day, a cluster of SGLang pull requests landed: shared-expert output moved onto the output stream, native output validation and terminal failure propagation were added, tool-call parameters switched from whole-buffer buffering to incremental streaming, KDA fused decode became default on AMD gfx950, and FP8/SiTU quantization work went in. These are the fixes a serving stack gets before exposing a new model to external traffic, not while patching a live one.
Public evidence for the release itself is still thin. DeepInfra's full model list contains only K2.6 and K3; SGLang has no kimi_k3_1.py; searches for 'K3.1', 'Kalmia', and 'Kibitz' in the repo return zero. The current K3 is a 2.8T-parameter MoE with 104B activated parameters and a 1M context, priced at $2.85/M input and $14.25/M output on DeepInfra. Its official benchmark card leads Claude Fable 5 on SWE-Marathon (42.0 vs 35.0) and MCPMark-Verified (94.5 vs 87.4), but loses on HLE-Full, CritPt, GDPval-AA v2 Elo, and even Kimi's own Code Bench 2.0. No official comparison includes GPT-6 Astra, whose release date is itself disputed.
If the anonymous runs and PRs are an accurate guide, K3.1 is an engineering-oriented release: deployment stability and tool-call streaming are exactly what the SGLang work is polishing. The version number supports that reading, since K3 already had a K3-0903 refresh in September. Expect the first confirmation signals in DeepInfra's model list and a kimi_k3_1.py merge in SGLang, before any announcement event.
The same-day timing of the SGLang PRs carries more weight than the Arena screenshots, because streaming and failure propagation are prerequisites for exposing a new model, not post-release maintenance.
A reused codename like Kibitz looks like deliberate continuity from Moonshot, but the public record offers no way to rule out a name-pool collision, so the evidence stays circumstantial.
K3.1's likely scope is already visible in its version number: a .1 bump plus inference-stack hardening means an engineering iteration, while a real capability jump would more plausibly land as K4.
The official benchmark card omits GPT-6 Astra entirely, and the closest listed model, GPT-5.6 Sol, still beats K3 on GDPval Elo; treating Astra as the comparison target inflates K3.1 expectations before it is even confirmed.
MXFP4 quantization-aware training changes deployment economics: even 4-bit weights are too large for eight 141GB GPUs, so most users will consume K3.1 through hosted APIs and have little reason to self-host.