Kimi K3: Open Frontier Intelligence
A 2.8-trillion-parameter open-weights model is a model story. What earns this report a sheet is how much of it is about the machine around the model — nine specialist policies distilled back into one, a reward model that writes its own rubric, and a reinforcement-learning loop trained against a swappable harness interface so the policy never overfits to a single one. Read against the four talks this site is built on, it says the harness is no longer only a runtime concern.
Primary source
Kimi K3: Open Frontier IntelligenceMoonshot AI
Paper page — abstract and technical report
Discussion
Hugging Face Journal Club: Kimi K3The Hugging Face research team
Recorded session, 41 min — speakers not identified
The examination — what the discussion did to each claim
- Corroborated
- The discussion agreed with the report.
- Contested
- The discussion pushed back on the report.
- Extended
- The discussion worked out more than the report stated.
- Unresolved
- The report did not say enough to settle it.
A washed band is unsettled — the two sources disagree, or the report does not say enough to decide. Timestamps are positions in the recording; its speakers are not identified, so no quotation is attributed to a name.
Claim 1. Scale, and where it stands
CorroboratedThe record
A 2.8T parameter Mixture-of-Experts model with 104 billion activated parameters, native vision capabilities, and a 1-million-token context window. The report hedges its own result downward: performance “still trails the most powerful proprietary models, namely Claude Fable 5 and GPT-5.6 Sol,” while beating every other model in its suite.
Abstract
The testimony
- 0:39
The … take-home message is that they've trained a model that is probably comparable to Opus 4.8.
Reading
A weights-open model at frontier scale, and a report honest enough to name the two models it does not beat. Nothing in the chain of limitations changes because a model got larger — which is the whole argument.
- 0:39
Claim 2. The architecture is the innovation
ContestedThe record
Built on Kimi Delta Attention and Attention Residuals, “which improve information flow across sequence length and model depth.” Together with Stable LatentMoE, activating 16 of 896 routed experts per token, these “yield an approximately 2.5x improvement in overall scaling efficiency over Kimi K2.”
Abstract
The testimony
- 32:08
I feel like there is nothing really new in a sense … it's more like existing pieces that are done well … there's no innovation is my point.
- 33:24
For many papers you have this kind of unique idea that emerges, and here … they are mostly describing things that already exist. They just push very, very deep the optimizations.
- 34:20
Maybe the frontier work is now on these engineering questions.
Reading
The report's headline is architectural; the room reads it as optimization depth rather than invention. The disagreement is the useful part — if this is what reaching the frontier takes, there is no secret left to find, only accumulated engineering.
Lands onEvolution — how a harness gets better- 32:08
Claim 3. Post-training is nine specialists, merged
ExtendedThe record
“Post-training highlights reinforcement learning across general, agentic, and coding domains and multiple reasoning-effort levels, enabling compositional generalization and robust long-horizon execution.”
Abstract
The testimony
- 1:44
They had this idea of general tasks … general agents which are things like deep research, long-horizon tasks, and then they have coding agents.
- 2:04
On top of that they also included three reasoning efforts … low, high, max. But this now gives you 3x3 — so nine domain experts.
- 2:46
Multi-teacher on-policy distillation, where the teacher is just the experts, and then they're doing everything back into the SFT checkpoint.
- 15:06
All the complexity of RL is really delegated to the individual experts … if you imagine in a team you want to parallelize your efforts, you can just say, you do coding, I do math.
Reading
Sub-agent decomposition, applied to training instead of to inference. Nine bounded specialists, each with a narrower job and narrower context, merged by a process that integrates their results — the same shape the harness runs at runtime, one level down.
Lands onSub-agents — bounded loops with a narrower job- 1:44
FIG. 1 — post-training. One supervised checkpoint, nine reinforcement-learning experts across three domains and three reasoning efforts, distilled back into a single model. Claim 4. Reasoning effort is trained, not prompted
ExtendedThe record
The report claims “multiple reasoning-effort levels” as a post-training result, but states the mechanism only in an appendix.
Abstract; training detail
The testimony
- 8:57
For every task you ascribe some initial budget which is estimated from the SFT model, and then you get minus one reward if your token budget exceeds some threshold.
- 9:36
They first train on the maximum budget and then they gradually anneal … to smaller values to get these expert models.
- 10:02
This seems similar, with the main difference that you're now really discretizing the reasoning … they're actually separate models.
Reading
A token budget enforced by negative reward and annealed downward over training: the model learns what an answer should cost instead of being told. The harness still has to decide which effort level to call and when — that decision stays outside the weights.
Lands onContext Management — protecting the model's attention- 8:57
Claim 5. The reward model writes its own rubric
ExtendedThe record
The abstract names the domains reinforcement learning ran across. It does not describe how reward was computed in the non-verifiable ones.
Abstract; not elaborated
The testimony
- 11:29
The reward model is now an agent itself … the agent looks at the output from the policy, it actually generates a rubric, it scores against the rubric.
- 12:18
Historically when people make rubrics they would have a human say, for the task, these are the rubrics … now the judge itself is generating on the fly.
- 12:44
How do you go beyond just a verifiable reward, if you want some notion of human preference, or fine-grained feedback that is not just binary?
Reading
Verification stops being a fixed check and becomes a generated one. This site's verification layer asks for receipts; here the thing deciding what counts as a receipt is itself a model. That moves the failure mode rather than removing it — and an unexamined rubric is a harder failure to see than a failing test.
Lands onVerification & Observability — the harness asks for receipts- 11:29
Claim 6. Training ran against a swappable harness
CorroboratedThe record
Infrastructure advances include “million-token agentic RL with persistent rollout and sandbox states.”
Abstract; infrastructure
The testimony
- 18:39
They have this unified white box environment which now they can plug in codex, kimi code, hermes, whatever — and then you can basically train on all of them dynamically so that you don't overfit to a single one.
- 19:33
When we implemented the harness support, for now it's only one harness … but we saw that you completely specialize to that harness at some point, especially with small models.
- 20:03
The complexity would be implementing those composable modules … what are the modules that define codex versus hermes versus pi — that would be the biggest work.
Reading
The strongest reason this report belongs in the reading room. The harness stopped being only the system you wrap around a finished model and became a surface the model is trained against — and training against one harness overfits the model to it. Every primitive in the pillar is now a training-time variable as well as a runtime one.
Lands onThe Harness — all eleven primitives- 18:39
FIG. 2 — the white-box interface. The policy never talks to a harness directly; harnesses plug into one interface and are sampled per rollout. Claim 7. Environments were grown, and the report stops there
UnresolvedThe record
The report claims persistent sandbox states at million-token scale. It does not describe how the environments themselves were constructed.
Abstract; undocumented
The testimony
- 20:32
They have this big knowledge graph of different domains … and then you have agents traversing this knowledge graph, doing web search to get papers, blog posts, code, whatever, and then they generate tasks from that exploration.
- 21:06
In principle you can get very large coverage without having to necessarily have lots of humans define all your tasks.
- 21:24
One thing that wasn't clear to me is, once you have the tasks, then how do you put that into the environment?
- 21:45
They haven't really mentioned after synthesis what exactly happens.
Reading
The most interesting claim in the report is the one it documents least. Task synthesis is described in detail; the step from a synthesized task to a running environment is missing, and two readers of the paper could not recover it. Where a source runs out of road is worth recording as carefully as what it establishes.
Lands onExecution Environment — where tool calls become bounded reality- 20:32
Where this reading lands
Read as a model release, this is a bigger Kimi. Read as a harness document, it is the first frontier report where the system around the model is trained as deliberately as the model — nine specialists merged into one, a judge that writes its own criteria, and an environment interface built to be swapped so the policy never learns one harness by heart. The chain of limitations now runs backwards into training.
11The pillar this reading feedsThe Harness