Kimi K3 is the best model ever made (sometimes)

Summary
Software engineer and creator Theo Browne (t3.gg) reviews Moonshot AI’s newly announced Kimi K3 model. He evaluates its benchmark results, technical MoE architecture, subscription and API pricing, real-world agentic software engineering capabilities, and hands-on 3D/frontend design generation.

What is shown

• Benchmark evaluations & charts: Comparison of Kimi K3 against Claude Fable 5, GPT-5.6 Sol, Claude Opus 4.8, and GLM-5.2 across DeepSWE, FrontierSWE, Terminal Bench 2.1, SWE Marathon, BrowseComp, and the Artificial Analysis Intelligence and Omniscience Indices [00:08, 00:37, 07:26, 08:36, 20:56, 23:51].
• Architecture and specs overview: Review of Moonshot’s blog covering the 2.8T parameter architecture, Kimi Delta Attention, Attention Residuals, 16 of 896 active experts in Stable LatentMoE, FP4 weights/FP8 activations, and the recommended 64+ accelerator deployment setup [03:15, 11:18, 19:54, 20:25].
• Code porting test: Testing long-horizon autonomous migration of the legacy ping.gg codebase in a terminal agent harness, completing 120+ subtasks before encountering context window truncation limits [00:48, 12:25, 13:16].
• 3D game creation demo: Prompting Kimi K3 to build “fishslop,” a fully functional 3D aquarium feeding game in Three.js/WebGL with sub control, procedural assets, and audio [16:20, 16:53, 17:46].
• Frontend UI generation & testing: Evaluating Kimi K3 on designing landing page variations for T3 Code [26:30], refactoring the T3 Code desktop UI to pure black theme [28:13], and creating an end-to-end pull request with before/after screenshots [32:12].
• Security auditing run: Running a multi-agent parallel discovery and verification security audit of the Lakebed codebase across 25 verification agents [31:14].
• Pricing & quota breakdown: Whiteboard diagram and subscription dashboard breakdown comparing API token pricing ($0.30 cache hit, $3.00 input, $15.00 output per million) against Kimi Code monthly tiers ($19 to $199/month) and calculating total token expenditure [37:23, 38:17, 38:46].

Claims & numbers

• The presenter states Kimi K3 is a 2.8-trillion parameter model with a 1-million-token context window and native vision capabilities [03:20].
• The presenter notes that at FP8 precision the model would require ~2.8 TB of memory, or ~1.4 TB at FP4 [04:16].
• The presenter shows official text stating full model weights are slated for release on July 27, 2026 [05:22].
• API pricing stated: $0.30/M tokens for cache-hit input, $3.00/M tokens for cache-miss input, and $15.00/M tokens for output [09:26].
• On DeepSWE, Kimi K3 scores 67.5% (trailing GPT-5.6 Sol at 73.0% and Fable 5 at 70.0%) [05:58].
• On FrontierSWE, Kimi K3 scores 81.2%, outpacing GPT-5.6 Sol at 71.3% [06:38].
• On Arena AI’s Frontend Code Arena, Kimi K3 achieves a 76% win rate, ranking #1 [13:45].
• On the Artificial Analysis Intelligence Index, Kimi K3 achieves an index score of 57, ranking 3rd overall behind Fable 5 (60) and GPT-5.6 Sol (59) [20:58].
• The architecture activates 16 out of 896 experts, claiming a 2.5x overall scaling efficiency gain over Kimi K2 [11:31, 12:05].
• Moonshot recommends deploying on supernode configurations with 64 or more accelerators, which the presenter notes costs between $1.5M and $2.6M for 64 H100s [20:32, 20:41].
• Across his test workloads involving ~195 million tokens, the presenter calculated his total estimated cost was approximately $63.25 [38:46].

Notable quotes

• [00:00] “Kimi K3 is here, and it is a huge leap for open-weight models.”
• [18:13] “It’s a good fucking model.”
• [29:12] “I think we finally have a model that’s good at solving real-world UI tasks without having to pay Anthropic massive amounts of money.”

Assessment
This is an authentic, hands-on developer review and empirical evaluation of Kimi K3. The presenter directly runs and demonstrates agentic tasks, terminal workflows, frontend redesigns, and a complete browser 3D game generated live during his testing.

Described by gemini-3.8-flash on 2026-10-12 from the video’s audio and frames.