Kimi K3: Where the 2.8T “Open 3T-Class” Model Beats—and Loses to—Fable 5¶

For / Key Points

For: Technical leads deciding whether their development team should evaluate or migrate to a new frontier LLM

Key Points:

Kimi K3 leads Frontend Code Arena, but still trails Fable 5 and GPT-5.6 Sol in overall performance and UX
Its API costs $3 for input and $15 for output, more than three times the previous generation
The weights are not public yet, and three official limitations belong in any migration plan

On July 16, 2026, Moonshot AI released Kimi K3, a 2.8-trillion-parameter model. That day, K3 scored 1,679 in Arena.ai's Frontend Code category and moved ahead of Claude Fable 5 into first place.¹²³ Moonshot nevertheless states that K3's overall performance still trails Fable 5 and GPT-5.6 Sol.¹

This article asks one practical question: which workloads make Kimi K3 a migration candidate, and which ones should wait?

The Verdict: It Wins Frontend, but Loses Overall and on UX¶

K3's win is real, but it is not universal. It leads a human-preference arena for frontend output and also performs well on long-horizon coding, web research, and document understanding in Moonshot's evaluations. Fable 5 or Sol still leads on hard reasoning, broad knowledge work, and user experience.

Decision axis	K3's position	Practical reading
Frontend generation	First in Arena.ai	Prioritize it for a UI pilot
Long-horizon coding	Best SWE Marathon score	A credible autonomous-work candidate
Web and documents	Best BrowseComp and OmniDocBench scores	Promising for research and document processing
Hard reasoning	About 10 points behind Fable 5 on HLE-Full	Do not rush the hardest tasks over
Broad knowledge work	Third on GDPval-AA v2	Too weak a case for one-model standardization
UX	Moonshot acknowledges a gap	Evaluate the actual product experience

The distribution matters more than a single model rank. Before reading the benchmark rows, it helps to understand what sits inside the model.

What Is Inside 2.8 Trillion Parameters¶

K3 is enormous, but it does not activate all 2.8 trillion parameters for every token. It is a sparse mixture-of-experts model that activates 16 of 896 experts, with a one-million-token context window and native visual understanding. Moonshot calls it the world's first “open 3T-class model,” although the weights are not yet available.¹

Item	Kimi K3	Caveat
Total parameters	2.8 trillion	Sparse MoE
Experts	16 of 896 active	Stable LatentMoE
Context	Up to 1M tokens	Kimi Code plans expose 256K or 1M
Input	Moonshot highlights text, image, and video understanding	Accepted formats vary by API surface
Reasoning effort	max only	low and high are planned
API price	$0.30 cached, $3 input, $15 output	Per million tokens
Weights	Promised by July 27	Not public on July 17

The two architectural pillars are Kimi Delta Attention and Attention Residuals. The first targets information flow across long sequences; the second retrieves representations across model depth. Combined with greater sparsity, Moonshot claims roughly 2.5 times the scaling efficiency of Kimi K2.¹

Moonshot also says quantization-aware training starts in the SFT phase, using MXFP4 weights and MXFP8 activations. It recommends supernodes with at least 64 accelerators for inference and plans to contribute a KDA-aware prefix-cache implementation to vLLM. These remain vendor design claims until the technical report can be inspected.

What the 35 Official Benchmark Rows Actually Show¶

Ten representative rows are enough to reveal K3's workload profile. The table below selects ten decision-relevant results from the 35 rows in Moonshot's launch post.¹ Bold marks the best score in each row.

Benchmark	Kimi K3	Claude Fable 5	GPT-5.6 Sol	Claude Opus 4.8
DeepSWE	67.5	70.0	73.0	59.0
Program Bench	77.8	76.8	77.6	71.9
Terminal-Bench 2.1	88.3	84.6	88.8	84.6
SWE Marathon	42.0	35.0	39.0	40.0
FrontierSWE	81.2	86.6	71.3	66.7
BrowseComp	91.2	88.0	90.4	84.3
GDPval-AA v2 (Elo)	1,668	1,760	1,748	1,600
GPQA-Diamond	93.5	92.6	94.1	91.0
HLE-Full	43.5	53.3	44.5	49.8
OmniDocBench	91.1	89.8	85.8	87.9

K3 leads on specification-driven reimplementation, long autonomous coding, web research, and document understanding. It does not lead DeepSWE, FrontierSWE, GDPval, or HLE. “K3 beat Fable 5 on frontend” and “K3 remains behind Fable 5 overall” can both be true.

Independent measurements point in the same direction. Artificial Analysis gives K3 an Intelligence Index score of 57, ranking it fourth out of 189 models, while describing it as slower and more verbose than nearby peers.⁴ In Simon Willison's simple SVG test, K3 used 16,658 output tokens—including 13,241 reasoning tokens—and cost about $0.25.⁵

How to read the benchmark table

K3 is often measured in KimiCode, while competing models use Claude Code, Codex, or another harness depending on the row. Some Fable 5 results include Opus 4.8 fallback. BrowseComp 91.2 also uses context compaction at 300K; with an unmanaged 1M context, K3 scores 90.4. Treat individual rows as vendor claims until a third party reruns them under one harness.

As capability gaps narrow, price becomes the next meaningful difference.

Sonnet-Level Pricing, Not a Budget Model¶

K3 costs $3 per million input tokens and $15 per million output tokens, up more than threefold from K2.6 at $0.95 and $4. That is roughly 3.2 times higher on input and 3.75 times higher on output. Simon Willison notes that the rate now matches Anthropic's Sonnet tier.⁵

Artificial Analysis describes K3 as slow and verbose and reports that its full Intelligence Index run cost $2,690.80.⁴ The real bill depends on reasoning tokens and retries, not just the rate card. A cost-motivated migration therefore needs cost per successful task, measured on the team's own work.

Caching is the strongest counterargument. Moonshot claims a cache hit rate above 90% on coding workloads and prices cache-hit input at $0.30 per million tokens.¹ Repeatedly reusing a large repository context could make the effective price much lower.

“Open Weight” Is Still a Promise¶

On July 17, K3 can be called but cannot be downloaded. Moonshot says it will publish the full model weights by July 27 and release a technical report later.¹ There is no K3 checkpoint in Moonshot's Hugging Face organization, and Artificial Analysis currently classifies the service as a proprietary model.⁴⁹

The evidence available today covers a hosted model's behavior and pricing. Commercial terms, artifact completeness, memory requirements, and self-hosted throughput remain unknown until the license and files arrive. “Open 3T-class” currently describes the release commitment, not the artifact available today.

The launch post's Limitations section is already concrete enough to evaluate.

The Three Limitations Moonshot Documents¶

Moonshot's limitations may matter more to a migration plan than its leaderboard. The launch post explicitly lists sensitivity to thinking history, excessive proactiveness, and a UX gap.¹

Sensitivity to thinking history K3 was trained with preserved thinking history. Quality can become unstable if a harness fails to return the complete history or if a user switches from another model to K3 mid-session.
Excessive proactiveness Long-horizon training can cause K3 to make unexpected decisions when instructions are ambiguous or small problems appear. Explicit boundaries in the system prompt or AGENTS.md are important.
A UX gap Moonshot says K3 is competitive overall but still has a noticeable user-experience gap from Fable 5 and Sol.

This point needs precision. Moonshot does not prohibit third-party harnesses such as Claude Code or Codex, and the Kimi Code documentation includes third-party integration guidance.⁶ The documented risks are switching an existing session and losing the preserved reasoning history.

Teams with established tooling should start a new session, set the intended context limit, preserve history, and validate compatibility on representative tasks. That requirement turns a model comparison into a model-plus-harness comparison.

How to Try It and What to Measure¶

Run the same small, representative task through three access paths. Kimi Code provides the vendor-tested baseline, Kimi API represents a production integration, and OpenRouter provides a low-friction trial path.

Kimi Code
Start a new session and choose K3 with /model. Moderato exposes 256K; Allegretto and higher expose up to 1M.⁶
Kimi API
Use model ID kimi-k3. The API is compatible with OpenAI Chat Completions and the official OpenAI SDKs.⁷
OpenRouter
Use model ID moonshotai/kimi-k3. It currently lists the same $3 and $15 rate.⁸

This is a minimal official-API call through the OpenAI Python SDK.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["MOONSHOT_API_KEY"],
    base_url="https://api.moonshot.ai/v1",
)
response = client.chat.completions.create(
    model="kimi-k3", messages=[{"role": "user", "content": "Introduce yourself in 3 lines"}]
)
print(response.choices[0].message.content)

The first-pass workload decisions fit into five rules.

Frontend and UI generation: prioritize a K3 pilot based on the Arena result
Document understanding and web research: evaluate it, then reproduce the vendor scores on internal data
Hard reasoning and broad agents: keep Fable 5 and Sol in the comparison
Cost reduction: measure dollars per successful task, not tokens alone
Existing harnesses: test a fresh session, history preservation, and permission boundaries together

Summary: Do Not Decide by July 27; Start Verifying on July 27¶

Kimi K3 is a strong example of an announced open-weight model reaching the frontier tier. Its current wins cluster around frontend work, long-horizon coding, research, and document processing, while hard reasoning and UX still show gaps.

Three checks matter around July 27.

The actual weights, license, and scope of the distributed artifacts
The technical evidence behind KDA, AttnRes, and the claimed 2.5× scaling efficiency
Success rate, cost, and correction count under one harness and the team's own tasks

The new implication is that model selection has become harness-contract selection. K3's thinking-history constraint exposes the limits of comparisons based on a model-name swap. Define the work, history, and permissions first; then read the ranking.

Sol vs Fable 5 — Compare the two models K3 is chasing
Kimi K2 Thinking — See the change from the previous generation
Claude Fable 5 release — Fable 5 availability and fallback behavior