|
| 1 | +--- |
| 2 | +slug: "kimi-k3-token-station" |
| 3 | +lang: "en" |
| 4 | +title: "Kimi K3 is the first open-source model in the 3-trillion-parameter class. Try it free on Token Station" |
| 5 | +summary: "Moonshot's Kimi K3 has 2.8 trillion parameters and a 1M-token context window, the first open-source model to reach the 3-trillion-parameter class. It's live on Token Station as kimi/kimi-k3 at list price, zero markup." |
| 6 | +category: "tutorial" |
| 7 | +date: "2026-07-25" |
| 8 | +cta: "https://models.bytefuture.ai/intro.html" |
| 9 | +draft: false |
| 10 | +--- |
| 11 | + |
| 12 | +Moonshot AI's Kimi K3 has 2.8 trillion parameters, the first open-source model to reach the 3-trillion-parameter class. Moonshot says a Kimi model has held the frontier for open-source model scale in 9 of the past 12 months. K3 extends that streak by a wide margin. |
| 13 | + |
| 14 | +Scale alone would make K3 notable. What makes it useful is the rest of the spec: a 1-million-token context window, native visual understanding, and an architecture built specifically for long-horizon coding, knowledge work, and reasoning. Full model weights ship by July 27, 2026; the API is live on Moonshot's platform now, and on Token Station today. |
| 15 | + |
| 16 | +## What's new under the hood |
| 17 | + |
| 18 | +K3 is built on Kimi Delta Attention (KDA), a hybrid linear attention mechanism, combined with Attention Residuals (AttnRes) to help information flow through longer sequences and deeper models. On the Mixture-of-Experts side, Moonshot's Stable LatentMoE framework pushes sparsity further: K3 activates just 16 of 896 experts per token. Together with training and data improvements, Moonshot puts K3's overall scaling efficiency at roughly 2.5x its predecessor, Kimi K2. |
| 19 | + |
| 20 | +Two workloads get explicit design attention: |
| 21 | + |
| 22 | +- **Long-horizon coding.** K3 is built to sustain long-running engineering tasks with minimal supervision: understanding large codebases, coordinating terminal tools, and combining software engineering with visual reasoning (using screenshots and visual feedback in frontend work, game development, and CAD). |
| 23 | +- **Knowledge work.** Moonshot reports consistent gains on internal evaluations built from recurring patterns in real user-agent collaboration, beyond what public benchmarks capture. |
| 24 | + |
| 25 | +## Where K3 lands on benchmarks |
| 26 | + |
| 27 | +Moonshot's own framing is direct: K3's overall performance still trails the two most capable proprietary models it compares against, Claude Fable 5 and GPT-5.6 Sol, but it beats Claude Opus 4.8 on several benchmarks. Two published numbers: |
| 28 | + |
| 29 | +- **DeepSWE: 67.3**, using the mini-SWE-agent harness. |
| 30 | +- **BrowseComp: 90.4**, using the full 1M-token context with no context management. |
| 31 | + |
| 32 | +The case studies are where the 1M context and long-horizon design show up in practice. On four NVIDIA Hopper GPU kernel optimization tasks, K3 performed competitively with Fable 5 (using fallback) and outperformed Opus 4.8, GPT-5.6 Sol, and GPT-5.5. In a compiler task, it built a Triton-like compiler (MiniTriton) from scratch that matches or beats Triton and `torch.compile`, and used it to run stable end-to-end nanoGPT training. On an astrophysics research task, it worked through 300+ equations of state and finished in about two hours what the team says would normally take one to two weeks by hand. |
| 33 | + |
| 34 | +## Try Kimi K3 free on Token Station |
| 35 | + |
| 36 | +K3 is live on [Token Station](https://models.bytefuture.ai/intro.html) as `kimi/kimi-k3`, at Moonshot's list price with zero markup: **$3.00 per million input tokens on a cache miss, $0.30 per million on a cache hit, and $15.00 per million output tokens**, with the full 1,048,576-token context window. K3's thinking mode cannot be turned off and defaults to maximum reasoning effort, so budget reasoning tokens as output; set `reasoning_effort` to `low` in your request if you want faster, cheaper responses. |
| 37 | + |
| 38 | +One thing Token Station removes: on Moonshot's own console, K3 is gated behind a minimum $1 top-up before it unlocks. On Token Station, your signup credit unlocks it immediately, no separate Moonshot account or top-up required. |
| 39 | + |
| 40 | +It is free to start. [Register](https://models.bytefuture.ai/signup) and you get $1 in credit, no card needed. Your first top-up then adds up to $50 in bonus credit. Here's where K3 sits next to other models already on Token Station: |
| 41 | + |
| 42 | +| Model | Input / 1M | Output / 1M | Context | |
| 43 | +|---|---|---|---| |
| 44 | +| `kimi/kimi-k3` | $3.00* | $15.00 | 1,048,576 | |
| 45 | +| `kimi/kimi-k2.7-code` | $0.95 | $4.00 | 256K | |
| 46 | +| `glm/glm-5.2` | $1.40 | $4.40 | 1M | |
| 47 | +| `anthropic/claude-opus-4-8` | $5.00 | $25.00 | 1M | |
| 48 | +| `openai/gpt-5.5` | $5.00 | $30.00 | 1M | |
| 49 | +| `anthropic/claude-fable-5` | $10.00 | $50.00 | 1M | |
| 50 | + |
| 51 | +\* Cache miss rate. Repeated context hits the $0.30 cache-hit rate instead, per the pricing above. |
| 52 | + |
| 53 | +Point the coding tools you already use at `kimi/kimi-k3` and run your real work through it. |
| 54 | + |
| 55 | +### Claude Code |
| 56 | + |
| 57 | +Claude Code reads its model and endpoint from environment variables. Route every tier through Token Station to K3: |
| 58 | + |
| 59 | +```bash |
| 60 | +export ANTHROPIC_BASE_URL="https://models.bytefuture.ai" |
| 61 | +export ANTHROPIC_AUTH_TOKEN="gw-YOUR_TOKEN_STATION_KEY" |
| 62 | + |
| 63 | +export ANTHROPIC_DEFAULT_OPUS_MODEL="kimi/kimi-k3" |
| 64 | +export ANTHROPIC_DEFAULT_SONNET_MODEL="kimi/kimi-k3" |
| 65 | +export ANTHROPIC_DEFAULT_HAIKU_MODEL="kimi/kimi-k3" |
| 66 | +export CLAUDE_CODE_SUBAGENT_MODEL="kimi/kimi-k3" |
| 67 | + |
| 68 | +claude |
| 69 | +``` |
| 70 | + |
| 71 | +### Codex |
| 72 | + |
| 73 | +Configure Token Station as the provider and make K3 the model: |
| 74 | + |
| 75 | +```bash |
| 76 | +mkdir -p ~/.codex |
| 77 | +cat > ~/.codex/config.toml <<'EOF' |
| 78 | +model = "kimi/kimi-k3" |
| 79 | +model_provider = "token_station" |
| 80 | +
|
| 81 | +[model_providers.token_station] |
| 82 | +name = "token_station" |
| 83 | +base_url = "https://models.bytefuture.ai/v1" |
| 84 | +env_key = "TOKEN_STATION_API_KEY" |
| 85 | +wire_api = "responses" |
| 86 | +EOF |
| 87 | + |
| 88 | +export TOKEN_STATION_API_KEY="gw-YOUR_TOKEN_STATION_KEY" |
| 89 | +codex |
| 90 | +``` |
| 91 | + |
| 92 | +### OpenClaw |
| 93 | + |
| 94 | +Register Token Station as a provider and set K3 as the default model: |
| 95 | + |
| 96 | +```json |
| 97 | +{ |
| 98 | + "models": { |
| 99 | + "mode": "merge", |
| 100 | + "providers": { |
| 101 | + "token-station": { |
| 102 | + "baseUrl": "https://models.bytefuture.ai/v1", |
| 103 | + "apiKey": "${TOKEN_STATION_API_KEY}", |
| 104 | + "api": "anthropic-messages", |
| 105 | + "models": [ |
| 106 | + { |
| 107 | + "id": "kimi/kimi-k3", |
| 108 | + "name": "Kimi K3 (Token Station)", |
| 109 | + "contextWindow": 1048576, |
| 110 | + "maxTokens": 131072 |
| 111 | + } |
| 112 | + ] |
| 113 | + } |
| 114 | + } |
| 115 | + }, |
| 116 | + "agents": { |
| 117 | + "defaults": { |
| 118 | + "model": { "primary": "token-station/kimi/kimi-k3" } |
| 119 | + } |
| 120 | + } |
| 121 | +} |
| 122 | +``` |
| 123 | + |
| 124 | +## Quirks worth knowing |
| 125 | + |
| 126 | +- **`max_completion_tokens` caps out higher than the default.** It defaults to 131,072 but can be set as high as 1,048,576 for output-heavy tasks. |
| 127 | +- **Vision input needs base64 or an uploaded file ID.** Public URLs are not supported for images or video. Send images inline as base64, or upload either media type via the Files API (`ms://<file-id>`); file upload is recommended for video. |
| 128 | +- **Web search is being updated.** Moonshot does not recommend K3's official web-search tool for production workflows in the near term. |
| 129 | +- **Weights land a couple of days after this API access does.** K3 is open-source, but full weights ship by July 27, 2026. Everything above works today through the hosted API; self-hosting comes shortly after. |
| 130 | + |
| 131 | +One key, the harness you already run, and the largest open-source model shipped to date, at 2.8 trillion parameters. If K3 holds up on your repository, a free signup is all it costs to find out. |
| 132 | + |
| 133 | +Start here: [models.bytefuture.ai](https://models.bytefuture.ai/signup) |
0 commit comments