As of July 25, 2026
Moonshot AI launched Kimi K3 on July 16, 2026, and it is the largest open-weight model announced to date: 2.8 trillion total parameters, a 1-million-token context window, and a #1 ranking on Arena.ai’s Frontend Code Arena that beats Claude Fable 5 and GPT-5.6 Sol at front-end web development. The model is live on Kimi’s app, website, and API today, with open weights promised by July 27, 2026.
But for anyone who buys local hardware to run models, there is a blunt reality under every headline: at 2.8T parameters, even the most aggressive quantization puts Kimi K3 at roughly 700 GB to 1.5 TB of memory. No desk-sized machine on our roster fits it. Not a single DGX Spark (128 GB unified). Not a 4x Spark cluster (512 GB unified). Not a 128 GB Strix Halo box. Not even a workstation with 8x RTX Pro 6000 GPUs totaling 768 GB VRAM (the Q2 raw floor is ~700 GB, leaving ~68 GB before runtime and KV overhead, which is insufficient for a confident fit verdict). This is multi-node datacenter territory.
What Moonshot announced
Kimi K3 is a sparse mixture-of-experts model using a new architecture Moonshot calls Stable LatentMoE. It routes each token through 16 of 896 experts, with an estimated ~50 billion active parameters per token. The exact figure is an estimate because 16 of 896 experts does not account for shared and non-expert parameters (source: Moonshot AI official tech blog, July 16, 2026).
Two architectural innovations come with it:
- Kimi Delta Attention (KDA): Moonshot reports up to 6.3x faster decoding at million-token contexts by hybridizing KDA with MLA and RoPE (source: Kimi_Moonshot X launch post, July 16, 2026).
- Attention Residuals (AttnRes): Moonshot reports roughly 25% higher training efficiency at under 2% additional cost (source: Kimi_Moonshot X launch post, July 16, 2026).
The model ships as a single Kimi K3 variant for both chat and agent tasks (source: Kimi API docs, July 2026). Moonshot applies quantization-aware training from the SFT stage onward using MXFP4 weights with MXFP8 activations for broad hardware compatibility.
The benchmarks that matter
Moonshot published an extensive benchmark table at launch. On the coding benchmarks that drive local-AI buying decisions, K3’s launch scores were:
| Benchmark | K3 score (July 16) | Best competitor |
|---|---|---|
| Terminal-Bench 2.1 | 88.3 | GPT-5.6 Sol at 88.8 |
| DeepSWE (mini-SWE-agent harness) | 67.3 | Claude Fable 5 ~65 |
| FrontierSWE | 81.2 | N/A |
| Program Bench | 77.8 | N/A |
| SWE Marathon | 42.0 | N/A |
| Arena Frontend Code | 1,679 Elo (#1) | Claude Fable 5 at 1,631 |
The Terminal-Bench 2.1 score is the headline number: 88.3 is within half a point of GPT-5.6 Sol (88.8), though Moonshot evaluated K3 with its KimiCode harness while comparator scores used different harnesses per Moonshot’s blog footnote. On Arena.ai’s Frontend Code Arena, K3 scored 1,679 Elo with a 76% pairwise win rate: when shown alongside another model’s output on the same task, human raters preferred K3’s output 76% of the time (sources: Moonshot AI official tech blog, July 16, 2026; Arena.ai on X, July 16, 2026; Arena.ai on X, July 16, 2026).
The DeepSWE score of 67.3 comes from Moonshot’s July 16 launch snapshot using the standardized mini-SWE-agent harness (source: Moonshot AI tech blog footnote, July 16, 2026). Since launch, the DeepSWE v1.1 leaderboard has been updated (July 21, 2026) and now shows K3[max] at 69% plus or minus 5 percentage points, with Claude Fable 5[max] at 70% plus or minus 4 (source: DeepSWE v1.1 leaderboard, July 21, 2026). The two models are within statistical noise on the current leaderboard snapshot.
Pricing and positioning
Kimi K3 costs $3 per million input tokens and $15 per million output tokens on the API, with cache hits at $0.30 per million (source: Kimi API pricing page, July 2026).
K3’s pricing is about 1.5x current Claude Sonnet 5 pricing: Sonnet 5 is running introductory rates of $2/$10 through August 31, and returns to $3/$15 on September 1 (its standard tier, which K3 matches). Against Claude Fable 5 ($10/$50), K3 is about 3.3x cheaper (sources: Anthropic Claude pricing, July 2026; Anthropic Fable 5 page, July 2026). Against DeepSeek V4-Flash ($0.14/$0.28 per million), K3 is about 20x more expensive on input and 50x on output, though that model operates at a completely different scale (284B total, 13B active) (source: DeepSeek API docs, July 2026).
The local-AI reality check
At 2.8T total parameters in FP16/BF16, the full weights land at roughly 5.6 TB. At Q4 (4 bits per parameter, the standard for running large models on local hardware), you are looking at roughly 1.4 TB of memory. At Q2, the most aggressive quantization, the estimate lands around 700 GB.
For context, here is every machine on our site and whether it can hold K3:
| Machine | Memory | Can run K3? |
|---|---|---|
| Mac Studio M4 Max | 64 GB unified | No |
| Mac Studio M3 Ultra | 96 GB unified | No |
| DGX Spark (single) | 128 GB unified | No |
| ASUS Ascent GX10 | 128 GB unified | No |
| Strix Halo box (GMKtec EVO-X2) | 128 GB unified | No |
| 4x DGX Spark cluster | 512 GB unified | No |
| Workstation (8x RTX Pro 6000) | 768 GB GPU VRAM | No (Q2 ~700 GB raw, ~68 GB insufficient runtime headroom) |
| 8x H100 server | 640 GB HBM | No |
| DGX B200 | 1,440 GB HBM | No |
The B200 fails because MXFP4 quantization adds block-scale overhead: one 8-bit scale per 32 4-bit values adds roughly 6.25% to the raw weight memory. For 2.8T parameters at MXFP4, the total is about 1,487.5 GB before KV cache and runtime overhead, which exceeds the B200’s 1,440 GB HBM. Multi-node deployment is required.
No desk-sized box on our roster runs this model. Not even close. The Q2 estimate at ~700 GB is about 5.5x the memory of a single 128 GB capacity box. The 512 GB 4x Spark cluster that runs DeepSeek V4-Flash (284B/13B active) is a natural next-step question, but that is a 284B model, and K3 is 10x larger by total parameters. Moonshot itself recommends deploying K3 on supernode configurations with 64 or more accelerators for inference efficiency, noting that this is a recommended topology rather than a proven K3 deployment (source: Moonshot AI tech blog, July 16, 2026).
The practical takeaway: if you are waiting for the K3 weights to run them on your desk, you will be waiting for hardware that does not exist yet. The model is a cloud-and-datacenter play. The local-AI buyer should keep comparing the machines that actually run the models they can fit: Qwen3.5-122B-A10B on a single Spark or Strix Halo, DeepSeek V4-Flash (284B, 13B active) on a 2x DGX Spark cluster at NVFP4 with FP8 KV cache and TP=2, or a 12B-27B dense model on an RTX 5090 build.
When the weights land
Moonshot has promised full weights by July 27, 2026 (source: Moonshot AI official blog, July 16, 2026). As of July 25, 2026, the Moonshot AI Hugging Face org lists K2-series models only: no K3 checkpoint, no model card, no license text. The previous Kimi models (K2.5, K2.6, K2.7 Code) shipped under a Modified MIT license on their Hugging Face model cards, but the exact K3 license has not been confirmed yet.
What this means for your next build
Kimi K3 is the first open-weight announcement in the 3T-parameter class. Per Moonshot’s published benchmarks, it leads the Frontend Code Arena at 1,679 Elo and scores 88.3 on Terminal-Bench. Its API pricing matches Claude Sonnet 5’s standard tier at $3/$15, though Sonnet 5 is $2/$10 through August 31.
But for the question this site exists to answer, “what hardware do I buy to run this model?”, the answer today is: nothing you can put on a desk.
If you are choosing between a Spark and a Strix Halo box for your next build, K3 does not change that decision. Neither runs it. What runs on that hardware today (Qwen 3.5 122B-A10B on a single Spark or Strix Halo, DeepSeek V4-Flash on a 2x DGX Spark cluster at NVFP4 with FP8 KV TP=2, the Gemma 4 family) is what will run on it when the K3 weights land. Compare your build against the models you can actually run.