Kimi k1.5: Complete Specifications, Pricing, API Access & Use Cases (2026)
What is Kimi K1.5?
Kimi k1.5 is Moonshot AI’s multimodal reinforcement-learning large language model, first submitted as a technical report on January 22, 2025, featuring 128K-token RL context scaling and text-vision reasoning, with public Kimi k1.5 API pricing not confirmed from official sources as of June 2026.
The official MoonshotAI repository describes Kimi k1.5 as an "o1-level multi-modal model," while the arXiv report describes it as a multimodal LLM trained with reinforcement learning, long-context scaling, improved policy optimization, and long-to-short reasoning transfer.
Users usually search for Kimi k1.5 to understand whether it is a reasoning model, whether it supports image input, how its 128K context scaling works, and whether it can be called through an API. For neutral comparison research, Kimi k1.5 can be evaluated alongside DeepSeek-R1 specs and API access, Claude 3.5 Sonnet specs and API access, and Qwen2.5-72B-Instruct specs.
What Are Kimi k1.5’s Key Specifications and Pricing?
| Field | Verified Value |
|---|---|
| Provider | Moonshot AI / Kimi Team (as of June 2026). |
| Model Family | Kimi (as of June 2026). |
| Model Type | Multimodal reinforcement-learning LLM / reasoning model (as of June 2026). |
| Release Date | January 22, 2025 technical-report submission; last revised June 3, 2025 (as of June 2026). |
| Context Window | 128K tokens for RL context scaling; public API context window not separately confirmed for Kimi k1.5 (as of June 2026). |
| Input Pricing | Not confirmed from official sources as of June 2026. |
| Cached Input Pricing | Not confirmed from official sources as of June 2026. |
| Output Pricing | Not confirmed from official sources as of June 2026. |
| Pricing Unit | Not confirmed for Kimi k1.5 as of June 2026. |
| Modality Support | Text and vision-language reasoning verified; audio and video support not confirmed for Kimi k1.5 (as of June 2026). |
| Supported Input Types | Text and image-text inputs verified in the technical report (as of June 2026). |
| Supported Output Types | Text output verified by the language-model framing; image, audio, or video output not confirmed (as of June 2026). |
| API Access | Kimi API exists, but Kimi k1.5 is not listed among the currently documented Kimi API model IDs checked (as of June 2026). |
| Model ID | Not confirmed from official sources as of June 2026. |
| Availability | Technical report and GitHub repository verified; public hosted API availability not confirmed for Kimi k1.5 (as of June 2026). |
| Knowledge Cutoff | Not confirmed from official sources as of June 2026. |
| Rate Limits | Not confirmed for Kimi k1.5 as of June 2026. |
| Fine-tuning Support | Not confirmed for Kimi k1.5 as of June 2026. |
| Streaming Support | Not confirmed for Kimi k1.5 as of June 2026. |
| Batch API Support | Not confirmed for Kimi k1.5 as of June 2026. |
| Tool / Function Calling | Not confirmed for Kimi k1.5 as of June 2026. |
| Structured Output / JSON Mode | Not confirmed for Kimi k1.5 as of June 2026. |
| License / Usage Restrictions | Model-weight license not confirmed; GitHub shows no published releases for the Kimi k1.5 repository (as of June 2026). |
Kimi k1.5 is therefore best described as a well-documented research model rather than a fully specified public API product. Its core technical claims are verifiable, but pricing, model ID, rate limits, and production API terms are not.
What Can Kimi k1.5 Do That Makes It Useful in Production?
Kimi k1.5 is relevant for reasoning-heavy evaluation because the report focuses on reinforcement learning over verifiable tasks, including math, coding, and visual reasoning. The model is not presented as a general procurement-ready API product, so production teams should treat it as a technical reference unless Moonshot AI publishes Kimi k1.5-specific API details.
- For math and code reasoning, Kimi k1.5’s report includes benchmarks such as AIME, MATH 500, Codeforces, and LiveCodeBench. These results indicate the intended evaluation domains, but benchmark scores should not be treated as guarantees of correctness in live systems.
- For multimodal reasoning, the report states that Kimi k1.5 is jointly trained on text and vision data and can reason across both modalities. This makes it relevant for image-grounded question answering, chart interpretation research, visual math tasks, and OCR-like reasoning workflows.
- For AI research teams, Kimi k1.5 is also useful as a case study in long-context RL. The report emphasizes long-context scaling, partial rollouts, length penalties, and a framework that avoids relying on Monte Carlo tree search, value functions, and process reward models.
What Are Kimi k1.5’s Supported Modalities?
| Modality | Supported? | Notes |
|---|---|---|
| Text input | Yes | The report covers language data, text prompts, math, code, and general reasoning tasks. |
| Image input | Yes | The report states that the model is jointly trained on text and vision data and includes image-text data. |
| Audio input | Not confirmed | No Kimi k1.5 audio-input support was confirmed in checked official sources. |
| Video input | Not confirmed | No Kimi k1.5 video-input support was confirmed in checked official Kimi k1.5 sources. |
| Text output | Yes | The model is described as an LLM that generates language sequences, reasoning, and final answers. |
| Image output | Not confirmed | No image-generation output capability was confirmed for Kimi k1.5. |
| Audio/video output | Not confirmed | No audio or video generation output was confirmed for Kimi k1.5. |
Where Does Kimi k1.5 Fall Short?
The main limitation is incomplete product documentation. Official Kimi k1.5 sources verify the technical report, GitHub repository, 128K-token RL context scaling, and text-vision reasoning, but they do not confirm public API pricing, a Kimi k1.5 API model ID, rate limits, fine-tuning support, or structured-output support as of June 2026.
- The current Kimi API model list checked for this article includes
kimi-k2.7-code,kimi-k2.7-code-highspeed,kimi-k2.6,kimi-k2.5, and Moonshot V1 models, but not Kimi k1.5. That means developers should not assume Kimi k1.5 is callable through the documented Kimi API without a provider-confirmed model ID. - Kimi k1.5 also has model-specific cost and latency concerns common to long reasoning systems. The report notes an "overthinking" pattern where response length can increase during RL training, and it introduces a length penalty because excessively long reasoning can be costly during training and inference.
- General AI limitations still apply. Kimi k1.5 outputs may be incorrect, incomplete, or overconfident. Legal, medical, financial, cybersecurity, and safety-critical use cases require expert review and independent verification.
What Is Kimi k1.5 Best Used For?
| Use Case | Why Kimi k1.5 May Fit | Important Limitation |
|---|---|---|
| Math reasoning research | The report emphasizes AIME and MATH 500 evaluations. | Benchmark performance does not guarantee real-world correctness. |
| Code reasoning research | Codeforces and LiveCodeBench are included in the evaluation context. | Public API model ID and production latency are not confirmed. |
| Vision-language reasoning | Text-image training and visual reasoning are verified. | Image output, video support, and audio support are not confirmed. |
| Long-context RL study | 128K-token RL context scaling and partial rollout methods are documented. | Architecture and training-scale details remain incomplete. |
| Model comparison pages | Kimi k1.5 can be compared with reasoning and multimodal models. | Missing pricing and API data limit procurement comparison. |
Kimi k1.5 may be useful for teams studying reinforcement-learning reasoning, text-image reasoning, and long-CoT-to-short-CoT transfer. For API-ready procurement comparisons, teams may also review DeepSeek-V3 specs and pricing, GPT-4o mini specs and pricing, and Llama 3.2 Vision specs.
How Does Kimi k1.5 Compare to DeepSeek-R1 and Claude 3.5 Sonnet?
| Comparison Area | Kimi k1.5 | DeepSeek-R1 | Claude 3.5 Sonnet | Scenario Fit |
|---|---|---|---|---|
| Model type | Multimodal RL-trained reasoning LLM. | Reasoning model with API access through deepseek-reasoner in the original release. | General-purpose Claude model for text, coding, and vision workflows. | Kimi and DeepSeek-R1 are closer for reasoning research; Claude is clearer for hosted production use. |
| Context | 128K-token RL context scaling reported; API context not confirmed. | Current DeepSeek docs map deepseek-reasoner compatibility to DeepSeek-V4 Flash thinking mode and note future deprecation. | 200K-token context window at launch. | Claude has clearer public context documentation; Kimi has research-context value. |
| Pricing | Not confirmed for Kimi k1.5. | DeepSeek-R1 release listed cache-hit, cache-miss, and output prices for deepseek-reasoner; current pricing has shifted to V4-era models. | $3 per 1M input tokens and $15 per 1M output tokens at launch. | DeepSeek and Claude are easier to budget from public documentation. |
| Modalities | Text and image input; text output verified. | Reasoning-focused API access; modality support depends on deployment. | Text and vision use cases are supported in Claude 3.5 Sonnet’s product context. | Kimi is useful for multimodal reasoning evaluation; Claude is easier to integrate. |
| Access clarity | Kimi k1.5 API model ID not confirmed. | Original release used deepseek-reasoner; current docs mark legacy model-name compatibility as time-limited. | Available through Anthropic API, Amazon Bedrock, and Google Cloud Vertex AI at launch. | Kimi is less procurement-ready unless API details are separately confirmed. |
This comparison is intentionally scenario-qualified. It does not declare an overall winner because the models differ in documentation maturity, access pattern, modality coverage, and production readiness.
How do I access Kimi K1.5?
Kimi k1.5 can be accessed as a technical report and official GitHub repository. The checked Kimi API documentation confirms a Chat Completions API, API-key authentication, a documented model list, and current model IDs for K2.7 Code, K2.6, K2.5, and Moonshot V1, but it does not confirm a Kimi k1.5 API model ID as of June 2026.
Because the required API details are not confirmed for Kimi k1.5, this page does not include executable Python or curl examples.
FAQs
What is Kimi k1.5’s context window?
Kimi k1.5’s official report states that the RL context window was scaled to 128K tokens. A separate public API context-window specification for Kimi k1.5 was not confirmed from official sources as of June 2026.
How much does Kimi k1.5 cost?
Kimi k1.5 input pricing, cached input pricing, output pricing, and pricing unit were not confirmed from official sources as of June 2026. The checked Kimi API pricing pages list other current Kimi model families, not Kimi k1.5.
Does Kimi k1.5 have an API?
A Kimi k1.5-specific API model ID, endpoint, request schema, and response schema were not confirmed in the checked official Kimi documentation as of June 2026. This article therefore avoids executable API examples.
What is Kimi k1.5 useful for?
Kimi k1.5 is most useful as a reference for multimodal reasoning, math and coding evaluation, long-context RL, and long-CoT research. Production use should account for missing API, pricing, and rate-limit details.
