Kimi K3 gets creator tests against GPT-5.6 Sol and Fable 5

• Kimi K3 is already live in Kimi's web app: stevibe's web screenshot shows K3 Max and K3 Swarm Max, while Kimi says the model is also available through Kimi Work, Kimi Code, and API.
• The creator test that set the tone was a single-file HTML canvas iPhone prompt; K3 produced the cleaner exploded-view animation in stevibe's side-by-side test against GPT-5.6 Sol.
• K3 looks near-frontier on coding benchmarks, not cleanly ahead of the closed leaders: LLMJunky's benchmark post put it third on DeepSWE, second on Terminal Bench 2.1, second on FrontierSWE, and first on SWE Marathon.
• The open-weight story comes with a hardware wall: stevibe's model taxonomy separated local, open, and closed models, while LLMJunky's hardware reply put a rough lower bound around 1.7 TB of VRAM.
• Hosted access was still rough on day one, with LLMJunky's OpenRouter report saying he hit rate limits every 2 to 3 tool calls.

The official Kimi K3 blog says the model has 2.8T parameters, Kimi Delta Attention, Attention Residuals, native vision, and a 1M-token context window. It also buries two practical details: full weights are scheduled for July 27, 2026, and K3 was trained in preserved thinking history mode, so switching an existing session over to K3 can make quality unstable. OpenRouter's K3 page lists the hosted model at $3 input and $15 output per 1M tokens, with only max reasoning effort available at launch.

Kimi's web selector already exposes K3 Max for chat and agent work, plus K3 Swarm Max for search and batch processing.

The Kimi API quickstart describes K3 as a 2.8T-parameter flagship with native visual understanding and a 1M-token context window. The launch blog says K3 is available on Kimi.com, Kimi Work, Kimi Code, and the Kimi API, with full weights due July 27.

Kimi's availability list is concrete:

• Kimi app on iOS, Android, HarmonyOS, and web.
• Kimi Work desktop app, version 3.1.0 or later, for Windows and Apple silicon Macs.
• Kimi Code in terminal, selected with /model.
• Kimi API with model ID kimi-k3.

The first creator benchmark was not a leaderboard. It was a no-library canvas animation prompt: a 360-degree rotating iPhone that disassembles into an exploded view, pauses, and reassembles.

stevibe said neither output was perfect, but K3 had the cleaner exploded view. stevibe's follow-up framed the remaining flaw as a model-decision problem, not an easy patch.

That test matched the prompt style creative coders actually use: single HTML file, simulated 3D, Apple-ish taste, image reference, no external libraries. shannholmberg's agent workflow post made the adjacent point from a marketing workflow: benchmarks are harder to map onto landing-page agents than direct UI edits.

Higgsfield turned K3 comparisons into a mini film-test battery, with Seedance 2.0 rendering the outputs.

The comparison set covered:

• Futuristic character animation, where higgsfield_ai's split-screen compared K3 with GPT-5.6 Sol.
• Game generation, where higgsfield_ai's Fable 5 comparison put K3 against Fable 5.
• AAA gameplay generation, where higgsfield_ai's gameplay test compared K3 with GPT-5.6 Sol.
• Ink wash animation, where higgsfield_ai's ink wash test used the same K3 versus Sol setup.
• Ice colossus animation, where higgsfield_ai's colossus comparison reused the same subject across models.
• Action choreography, where higgsfield_ai's action-scene test compared K3 with Fable 5.

The notable pattern is not that one model won every clip. The stack itself is becoming routine: language model for structure, video model for render, same prompt turned into a model shootout.

minchoi's thread collected the broadest creator inventory, and it leaned hard toward games.

The examples were easier to scan as a build list:

• Game Boy Advance emulator: minchoi's first clip showed Pokémon Emerald running.
• Creative visuals and functions: minchoi's second example grouped smaller visual demos.
• Full 3D scene: minchoi's scene post called out textures, proper lighting, and detailed objects.
• CS:GO x Portal clone: minchoi's clone post said K3 three-shotted it with 600K tokens.
• Horror House benchmark: minchoi's BridgeBench post claimed K3 beat Fable 5 on that game test.
• Animal Crossing clone: minchoi's sixth example added another playable-style build.
• macOS-style UI: minchoi's UI post said K3 built it in 15 minutes from a single prompt.
• Kernel stack work: minchoi's kernel post said K3 was rewriting and optimizing its own kernel stack.
• Wuxia RPG: minchoi's RPG example added a style-specific game demo.
• Cyberpunk web-swinging game: minchoi's final example closed the thread with another movement-heavy prototype.

kaigani ran a narrower game-generation test and said the Q-Bert Bench result was on par with Claude Fable 5 and at least matching GPT-5.6 Sol Max.

LLMJunky called K3 very good, near parity with the current frontier, and still not better than Fable 5 or GPT-5.6 Sol overall.

The official coding table and LLMJunky's screenshot line up on the most quoted numbers:

• DeepSWE: K3 67.5, behind GPT-5.6 Sol at 73.0 and Fable 5 at 70.0.
• Terminal Bench 2.1: K3 88.3, just behind GPT-5.6 Sol at 88.8.
• FrontierSWE: K3 81.2, behind Fable 5 at 86.6 and ahead of GPT-5.6 Sol at 71.3.
• Program Bench: K3 77.8, slightly ahead of GPT-5.6 Sol at 77.6 and Fable 5 at 76.8.
• SWE Marathon: K3 42.0, ahead of Opus 4.8 at 40.0, GPT-5.6 Sol at 39.0, and Fable 5 at 35.0.

Pricing is part of the reaction. kaigani's pricing screenshot put K3 at $3 per 1M input tokens and $15 per 1M output tokens, compared with $5 and $30 for GPT-5.6 Sol. LLMJunky's price reply called K3 roughly half the price of Sol and close in capability.

Token behavior was the open question. fabianstelzer's Hermes test said the pricing advantage could be eaten by greedy reasoning and latency, while LLMJunky's token-consumption reply argued lower per-token prices matter less when a model burns more tokens.

The cleanest framing came from stevibe: local models run on your hardware, open models run on hardware some people have, and closed models run only inside the company that made them.

K3 sits in the middle bucket. LLMJunky's VRAM reply said a stripped setup might still need about 1.7 TB of VRAM and several hundred thousand dollars, after his original post joked about needing 3.5 TB.

Moonshot's own blog points in the same direction. Its infrastructure section says K3 uses Stable LatentMoE with 16 of 896 experts active and recommends supernode deployments with 64 or more accelerators.

Open weights still changed the ownership conversation. LLMJunky's inference-provider reply argued that providers can solve the hardware problem while users still get rights they do not get from closed models: own, use, and fine-tune the weights.

The first access questions were not abstract. levelsio wanted K3 inside Claude Code because Fable and Opus were blocking his retro-computing work, specifically installing Yahoo! Messenger from 2003 in Windows XP.

Officially, Kimi points developers to Kimi Code and the /model command. In the community thread, LLMJunky's Kimi CLI reply said the Kimi CLI was already pretty good.

OpenRouter was less smooth. LLMJunky's OpenRouter report said K3 was rate-limiting every 2 to 3 tool calls, and his follow-up treated the situation as absurd rather than a small hiccup.

One creator already wrapped the model into a tiny interactive object. gokayfem's Kimi K3 Micro post pointed to a live demo, and the follow-up included both the live URL and GitHub link.