Kimi K3 in 10 Minutes: Moonshot AI's 2.8T Open Model, API Setup, Pricing, and Benchmarks

Moonshot AI released Kimi K3 on July 16, 2026, and it is the first open-source model to cross into the 3-trillion-parameter class. At 2.8T parameters with a 1M-token context window and native multimodal understanding, K3 enters the frontier conversation alongside Claude Fable 5 and GPT-5.6 Sol -- but with a critical difference: the weights will be released by July 27, 2026. The video linked below walks through K3's capabilities in 10 minutes; this post covers what you need to know to evaluate and call the model from code today.

Watch the video: Kimi K3 in 10 Minutes

Official Sources#

What Kimi K3 Is#

Kimi K3 is a 2.8-trillion-parameter model built on two new architectural components: Kimi Delta Attention (KDA), a hybrid linear attention mechanism, and Attention Residuals (AttnRes), which selectively retrieves representations across depth. Together they deliver roughly 2.5x the scaling efficiency of the previous-generation Kimi K2.

The model uses a Stable LatentMoE framework, activating 16 out of 896 experts per forward pass. It is natively multimodal -- text, image, and video input all flow through the same model, not a bolted-on vision encoder. Context extends to 1 million tokens, and thinking mode is always enabled with configurable reasoning effort (low, high, max -- default max).

Moonshot AI explicitly positions K3 against Claude Fable 5 and GPT-5.6 Sol. The announcement states that while K3 still trails those two on overall user experience, it is competitive or ahead on several coding benchmarks. For nine of the past twelve months, Kimi models have held the upper bound of open-source model scale -- K3 extends that lead dramatically.

Benchmarks: Where K3 Competes#

K3 was tested across 15+ benchmarks covering coding, agentic productivity, and multimodal understanding. Here are the notable coding results from the official blog:

On long-horizon coding (Terminal-Bench, DeepSWE, Program Bench), K3 leads all tested models. On SWE Marathon it beats Fable 5 and is close to GPT-5.6 Sol. However, PostTrain Bench shows a significant gap behind both Fable 5 and Sol, and the model's own limitations section notes that K3 has a "noticeable gap in user experience" compared to the two giants.

On browse agent tasks (BrowseComp) with 1M-token context and no compaction, K3 hits 87.1% -- competitive territory.

K3's Unique Capabilities#

The K3 announcement includes several case studies that go beyond standard SWE-bench results:

GPU Kernel Optimization. K3, Fable 5, Opus 4.8, and the GPT models were each given up to 24 hours to optimize GPU kernels in an identical sandbox. K3 substantially outperformed Opus 4.8, GPT-5.6 Sol, and GPT-5.5, competing closely with Fable 5. In late development, an early K3 version handled most of the team's kernel optimization work.

GPU Compiler Development. K3 built MiniTriton -- a compact Triton-like compiler with a tile-level IR over MLIR, optimization passes, and PTX codegen -- entirely from scratch. MiniTriton matched or beat Triton on supported workloads and sustained end-to-end nanoGPT training with stable convergence.

Chip Design. In a single 48-hour autonomous run, K3 designed a chip (1.46M standard cells, 0.277 MB SRAM, INT4 MAC array) that closes timing at 100 MHz and sustains 8,700+ tokens/s decode throughput. A model designing a chip to serve a nano model built on its own architecture.

Knowledge Work. K3 produced a 42-year ASIC industry analysis with 2.8k+ web searches, 1.1k+ terminal data pulls, and 11k+ pages across 87 quarterly reports and 99 PDFs -- through 120+ rounds of recursive self-improvement.

API Setup and Pricing#

Kimi K3 uses an OpenAI-compatible API format. The base URL is https://api.moonshot.ai/v1.

# Install the OpenAI SDK
pip install --upgrade 'openai>=1.0'

# Set your API key (get one at https://platform.kimi.ai/console/api-keys)
export MOONSHOT_API_KEY="your-key-here"

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["MOONSHOT_API_KEY"],
    base_url="https://api.moonshot.ai/v1",
)

completion = client.chat.completions.create(
    model="kimi-k3",
    reasoning_effort="max",
    messages=[{"role": "user", "content": "Prove sqrt(2) is irrational."}],
)

print(completion.choices[0].message.content)

Important API details for K3:

• temperature=1.0, top_p=0.95, n=1 are fixed -- omit them from requests
• max_completion_tokens defaults to 131,072 and can go up to 1,048,576
• Reasoning effort uses reasoning_effort at the top level: "low" / "high" / "max" (default "max")
• Vision input requires base64-encoded images or file IDs in ms:// format -- public URLs are not supported
• Context caching is automatic for prefixes over 256 tokens
• For multi-turn conversations, return the complete assistant message unchanged

Pricing (flat rate, no tiering by context length):

The official API achieves a cache hit rate above 90% on coding workloads. K3 requires a minimum $1 top-up to unlock. Full model weights release by July 27, 2026.

Streaming and Thinking Content#

K3 streams both reasoning_content (chain-of-thought) and final content deltas separately:

stream = client.chat.completions.create(
    model="kimi-k3",
    messages=[{"role": "user", "content": "Explain why the sky is blue."}],
    stream=True,
)

for chunk in stream:
    delta = chunk.choices[0].delta
    reasoning = getattr(delta, "reasoning_content", None)
    if reasoning:
        print(reasoning, end="", flush=True)
    if delta.content:
        print(delta.content, end="", flush=True)

Limitations#

The official K3 blog explicitly lists three key limitations:

1. Sensitivity to thinking history. K3 was trained with preserved thinking history. If your agent harness drops or truncates prior reasoning content, output quality degrades sharply. Use a verified-compatible harness like Kimi Code. Do not switch to K3 mid-session from another model.

2. Excessive proactiveness. K3 is optimized for long-horizon challenging tasks, so it may make unexpected decisions when facing minor issues or ambiguous intent. Add explicit behavioral constraints in your system prompt or AGENTS.md if your application needs the agent to stay within defined boundaries.

3. UX gap vs. Fable 5 and Sol. While competitive on benchmarks, K3 still falls short on subjective user experience compared to the two top proprietary models.

When to Use Kimi K3#

Use K3 when:

• You need an open-weight model you can self-host (after July 27)
• Long-horizon coding tasks where 1M context and persistent thinking help
• Cost-sensitive agent fleets (cache-hit input at $0.30/MTok is 10x cheaper than Fable 5's cache-miss rate)
• Tasks blending code with visual reasoning (game dev, frontend, CAD)
• You are already using Kimi Code and want a consistent model stack

Skip K3 (for now) when:

• You need the smoothest developer experience -- Fable 5 and GPT-5.6 Sol are still ahead
• Your agent harness does not preserve thinking history properly
• Production workflows need predictable, bounded behavior (K3's proactiveness can surprise you)
• You are doing PostTrain or similar tasks where K3 trails significantly

Watch the Video#

Kimi K3 in 10 Minutes on the Developers Digest channel walks through the model's capabilities with screen recordings -- showing live demos of the API, Kimi Code integration, and benchmark breakdowns. The video covers the visual flow and pacing that a text post cannot capture.

FAQ#

Is Kimi K3 open source?#

The model weights will be released by July 27, 2026. Moonshot AI is working with inference partners and open-source maintainers to ensure a reliable rollout. Until then, K3 is available exclusively through the Kimi API and Kimi products.

What does Kimi K3 cost?#

$0.30/MTok for cache-hit input, $3.00/MTok for cache-miss input, and $15.00/MTok for output. The API reports cache hit rates above 90% on coding workloads, which means most coding inputs will fall at the $0.30 rate.

How does Kimi K3 compare to Claude Fable 5?#

K3 leads Fable 5 on Terminal-Bench 2.1 (76.1% vs 58.4%) and is competitive on SWE Marathon and DeepSWE. But the official announcement states K3 "exhibits a noticeable gap in user experience" compared to Fable 5, and on PostTrain Bench Fable 5 leads 83.3% to 56.7%. K3 is an open model; Fable 5 is proprietary. The tradeoff is between ownership/self-hosting and polished UX.

Can I use Kimi K3 with Claude Code or Codex?#

K3 is compatible with Claude Code when configured as a backend (select via /model in Kimi Code CLI). The K3 docs specifically note that compatibility with Claude Code is verified. For other agent harnesses, K3's sensitivity to thinking history means you need a harness that preserves full reasoning content between turns.

What is the context window for Kimi K3?#

1 million tokens. The model supports context caching for prompts exceeding 256 tokens, with automatic cache hits for unchanged prefixes. No cache ID or TTL management is required.

Sources#

Continue Reading#

• Claude Fable 5 in 7 Minutes -- the same format for Anthropic's flagship model, covering benchmarks, pricing, and real-world demos
• Best Claude Model After Fable 5 -- if Fable 5 is disabled, here is an honest ranking of alternatives by task
• Agent Fleet Economics: Fable 5 vs Sonnet 5 -- cost analysis for running agent fleets across model tiers, relevant to K3's pricing advantage
• Claude Code vs Codex vs Cursor vs OpenCode -- coding agent comparison; K3 can slot into several of these as a backend model
• Apertus: Sovereign AI -- Europe's Open Model Push -- another open-weight model initiative, for context on the growing open frontier landscape