Kimi K3 Free: 1M Context, Pricing & Hands-On Review (2026)

Moonshot AI released Kimi K3 on July 16, 2026. The model has 2.8 trillion total parameters, a 1-million-token context window, and native visual input for long-horizon coding, knowledge work, and reasoning. Read the official Kimi K3 technical blog.

---

How to Use Kimi K3 Free on Mira

Mira now supports Kimi K3. New accounts receive signup credits, and daily sign-ins add more credits. You can use those credits to try K3 without an API key or local setup. Credit details are available on your account page. Sign up and use Kimi K3 free.

---

What Is Kimi K3?

The parameter count, architecture, context window, and weight-release plan come from the official technical blog. Reasoning behavior, visual inputs, and the API model ID come from the Kimi K3 API quickstart. Moonshot AI has not disclosed the total number of active parameters.

As of July 17, 2026, K3 was described as an open model but was not yet downloadable. Moonshot AI said full weights would arrive by July 27. The same official post recommends deployment on supernodes with 64 or more accelerators, which puts self-hosting closer to institutional infrastructure than a workstation experiment.

Moonshot AI also says K3 uses quantization-aware training from the SFT stage, with MXFP4 weights and MXFP8 activations. Quantile Balancing, Per-Head Muon, SiTU, and Gated MLA are presented as techniques for stable routing and training at 896-expert scale. These remain vendor disclosures until the full technical report is available.

---

What Do the Official Kimi K3 Benchmarks Show?

Moonshot AI published a broad benchmark set for coding, agents, reasoning, knowledge work, and vision. These are the most relevant coding and agent results for a comparison with GLM-5.2:

Source: Kimi K3 official benchmark tables. The harness is not identical across every model and benchmark: Kimi Code, Claude Code, Codex, and other agent frameworks appear in the footnotes, while some competitor results come from official reports or third-party leaderboards. These numbers describe the claimed capability range; they are not a controlled head-to-head test.

Moonshot AI also states that K3 still trails Claude Fable 5 and GPT-5.6 Sol overall. The official evidence supports “frontier-class performance,” not “best model in every task.”

---

What Capabilities Do the Official Cases Demonstrate?

Moonshot AI's launch examples fall into four groups:

• Coding and systems engineering: sustained GPU-kernel optimization and a MiniTriton compiler with its own intermediate representation, optimization passes, and PTX code generation.
• Visual feedback and digital creation: iterative work between code and screenshots for 3D games, simulators, CAD, animation, and video editing.
• Chip design: a 48-hour agent run using open EDA tools to build, optimize, and verify a 4 mm² chip.
• Scientific and knowledge work: reproducing the I-Love-Q relation in computational astrophysics, checking more than 20 papers, evaluating over 300 equations of state, and producing more than 3,000 lines of Python.

The launch post also shows a 42-year ASIC industry study and an analysis of 391 gravitational-wave events. These examples help explain K3's product positioning, but they come from Moonshot AI and are not independent tests.

---

What Does Independent Testing Add?

Artificial Analysis gave Kimi K3 an Intelligence Index score of 57 in its initial July 2026 evaluation and listed it fourth among 189 models as of July 17, 2026. The page measured output speed at about 62 tokens per second and reports that K3 generated roughly 130 million output tokens during the full intelligence evaluation.

Rankings change as new models are added. K3 reached the frontier group while producing unusually long outputs, so real cost depends on task length, agent turns, and cache behavior as well as the list price per million tokens.

Artificial Analysis marked K3 as proprietary while the weights were still unavailable. That does not directly contradict Moonshot AI's plan to release them later: one label described current downloadability, while the other described a future release.

---

Kimi K3 API Pricing

Kimi's official prices separate cache-hit input, cache-miss input, and output:

The RMB prices come from the official Chinese launch article; the USD prices come from the international technical blog. These prices are current as of July 17, 2026. Check the billing page for your account before use.

Moonshot AI says coding workloads on its API exceed a 90% cache-hit rate, but that is a vendor aggregate, not a promise for every application. K3 is not a low-cost continuation of K2: output length and long agent runs can matter more than the headline token price.

---

Kimi K3 vs GLM-5.2

K3's price is an official figure. The GLM-5.2 price above comes from a July 17 third-party pricing summary, not an official GLM price page. Check current pricing before use.

For GLM-5.2's free access, published benchmarks, and a separate Mira research test, see our GLM-5.2 free access and review.

---

Hands-On Test: Kimi K3 Runs a Full Deep Research Workflow

We selected Kimi K3 in Mira and used DP-YAMO-Agent with Deep Research v3.2 to study topology schemes in free-energy perturbation (FEP). The task included literature retrieval, evidence collection, report generation, and citation checks.

K3 delivered a 31.8 KB report plus the complete artifact archive. The report covers single, dual, hybrid, separated, and ATM topology approaches, then connects them to atom mapping, dummy atoms, soft-core potentials, charge changes, ring changes, and software implementations. The run collected 16 numbered references, 44 evidence cards, three contradiction entries, and 206 candidate claims. The final report passed its completeness and citation checks.

We then used GPT-5.4 with the same agent, skill, and research topic to compare report quality, execution time, and token use.

The time figure is the sum of model execution time, not the full wait from starting the task to seeing the result on screen. Input-token totals include cache hits, and providers may count tokens differently. These numbers describe this task rather than every possible workload.

What Differed Between the Two Runs?

• K3 produced the more complete report. It connected topology categories to implementations, dummy atoms, soft-core handling, charge and ring changes, and conditional selection guidance. Its 16 references passed the completeness check.
• GPT-5.4 was faster and used fewer input tokens. It used 23m 55s less execution time and about one-third of K3's input tokens. Its final report contained mismatches between conclusions and citations, so it did not pass the final quality check.
• K3 needed one continuation. Its second long run reached the 120-turn limit. After the user entered “continue,” K3 corrected the remaining issues and packaged the deliverables.

For this task, K3 produced the more complete report, while GPT-5.4 finished with less time and fewer tokens. One task cannot represent every use case, and results will vary with the research question, prompt, and tool setup.

---

Current Kimi K3 Limitations

Moonshot AI lists three limitations in its launch materials:

1. Sensitivity to reasoning history. Agent frameworks should return the full reasoning history; missing history or switching into K3 mid-session can reduce stability.
2. Over-initiative. Because K3 is optimized for hard, long-horizon work, it may resolve ambiguity in ways the user did not intend. Explicit system instructions or project rules help.
3. A remaining experience gap. Moonshot AI says K3 still trails Claude Fable 5 and GPT-5.6 Sol in overall user experience.

The API also launched with max reasoning effort only. Public image URLs are not accepted directly for visual input, and Kimi's official web-search tool was still being updated, so users should not depend on it yet.

---

Who Is Kimi K3 For?

A strong fit for:

• Large codebases, long documents, or multi-stage projects that can use a 1M context window;
• Front-end, game, CAD, video, and scientific workflows that combine code with visual feedback;
• Teams willing to spend more tokens on difficult long-horizon agent tasks;
• Users who want to validate a real research workflow with Mira credits before paying for direct API use.

Use caution when:

• The task is a short answer, simple rewrite, or high-volume lightweight request; max reasoning and long outputs may be inefficient;
• You need local deployment immediately; the weights were not public on July 17;
• Latency matters; Artificial Analysis measured about 62 output tokens per second;
• You need mature reliability data; independent long-term evidence is still limited.

---

FAQ

Is Kimi K3 free?

Mira now supports Kimi K3. New accounts receive signup credits, and daily sign-ins add more credits that can be used to try K3. Credit details are available on the account page.

Do I need an API key to use Kimi K3 free?

Not on Mira. You only need a Moonshot API key when calling Kimi's official API directly.

How large is Kimi K3's context window?

The official API documentation lists 1,048,576 tokens, commonly written as 1M. Product-level access may depend on the relevant plan.

Is Kimi K3 open source?

Moonshot AI described K3 as an open model and planned to publish full weights by July 27, 2026. The weights were not yet downloadable as of July 17, 2026.

How much does the Kimi K3 API cost?

Official prices per million tokens are ¥2 / $0.30 for cache-hit input, ¥20 / $3 for cache-miss input, and ¥100 / $15 for output. Regional billing pages are the final authority.

Is Kimi K3 better than GLM-5.2?

K3 scores higher on several long-horizon coding and agent benchmarks published by Moonshot AI. GLM-5.2 has lower public API pricing in the third-party comparison cited above and already-available MIT weights. The better choice depends on completion quality, total tokens, latency, and deployment needs.

Can I run Kimi K3 locally?

The full weights were not public as of July 17, 2026. Even after release, the official recommendation of a 64-plus-accelerator supernode indicates that K3 is not aimed at ordinary single-machine deployment.

Where can I test Kimi K3 as a research agent?

Use a free Mira account to select Kimi K3, invoke research skills and tools, and keep the workflow history and generated files.

---

Sources

• Kimi K3 official technical blog
• Kimi K3 official Chinese launch article
• Kimi K3 API quickstart
• Artificial Analysis: Kimi K3
• GLM-5.2 official model card and weights
• AIReiter: Kimi K3 pricing comparison

Want to verify K3 on your own research task? Use a registered Mira account and free credits to run it directly. Start with Kimi K3 →