Kimi K3 Release: 2.8T Open-Weight Model with 1M Context, Agentic Capabilities, and Frontier Benchmark Scores

Kimi K3 launches as a 2.8 T parameter open‑weight model

Kimi AI announced that its K3 model is now live. The model has 2.8 trillion total parameters, a 1 million token context window, and native vision support. K3 is positioned as the most capable model in Kimi’s lineup and is marketed for agentic coding and knowledge‑work.

Pricing and cost efficiency

  • Pricing – $3 / 1 M input tokens and $15 / 1 M output tokens (cache reads $0.30). This mirrors Anthropic’s Sonnet pricing and is close to the $2.5 / 1 M token rate of GPT‑5.6 Terra.
  • Reasoning efficiency matters – Users note that models with higher reasoning efficiency (e.g., GPT, Claude Fable) can be cheaper per task even if token prices are similar. K3’s default “max” reasoning mode can consume many more tokens, which may raise effective cost.
  • Cost‑per‑task estimates – Artificial Analysis reports an average task cost of $0.94, comparable to GPT‑5.6 Sol Max ($1.04) and cheaper than Claude Fable (≈$2.75).

Benchmark performance

Benchmark K3 score Rank (higher is better)
GDPval‑AA v2 (44 occupations, 9 industries) 1687 3rd – behind Claude Fable 5 Max and GPT‑5.6 Sol Max
AA‑Briefcase (agentic knowledge‑work) 1527 2nd – behind Claude Fable 5 Max, ahead of GPT‑5.6 Sol Max

These scores place K3 just behind the top proprietary models and ahead of Claude Opus 4.8 Max (1600). The model’s performance is described as “frontier‑level” by Kimi’s own evaluation.

Architectural highlights

  • Mixture‑of‑Experts (MoE) – K3 uses the Stable LatentMoE framework, activating 16 of 896 experts per token. This yields roughly 50 billion active parameters (≈2.5× the scaling efficiency of K2).
  • Parameter count – K3’s 2.8 T total parameters make it the largest open‑weight model publicly listed, surpassing DeepSeek‑V4‑Pro (1.6 T) and Mistral Large 3 (675 B).
  • Reasoning mode – Only a “max” reasoning effort is currently supported; temperature, top‑p, and other sampling knobs are fixed.
  • Vision input – Accepts base64‑encoded images; public URLs are not supported.

Availability and usage limits

  • API – Endpoint: https://platform.kimi.ai/docs/guide/kimi-k3-quickstart. Default max_completion_tokens is 131 072 and can be raised to 1 048 576.
  • Auth – Account creation requires a phone number or Google login; no email‑only sign‑up.
  • Quota – Free tier provides a limited number of tokens; paid plans start at $40/month with a “$100 Kimi Coding” quota that roughly matches Anthropic’s $200 Fable plan.
  • Rate limits – Some users report throttling on the OpenCode Go model and occasional time‑outs on long‑running tool calls.

Community impressions

"K3 pushes the boundary of end‑to‑end knowledge work. On the GDPval‑AA v2 leaderboard it scores 1687, ranking behind only Claude Fable 5 Max and GPT‑5.6 Sol Max." – ekojs (HN comment)

"The model is quite capable and being able to read the reasoning trace is a bonus, though it isn’t the fastest." – avph

"K3 designed a chip in 48 h, achieving 8 700 tokens/s decode throughput in simulation – a clear sign of long‑horizon agentic capability." – lukebechtel

"Pricing is high for a Chinese open‑weight model, but if the performance truly rivals frontier models the cost is justified." – Tiberium

"The model often repeats reasoning steps and can be slower than expected, which may increase real‑world cost." – npn

Open‑weight status and future releases

  • Kimi’s blog (https://www.kimi.com/blog/kimi-k3) states that full model weights will be released by July 27 together with a technical report.
  • Early versions of the announcement omitted the word “open”; the later blog post re‑adds it, confirming the intention to open‑source the weights.
  • Future updates are promised to add low/high effort reasoning modes and expanded tool‑calling support.

Comparison with competing frontier models

Model Parameters Context Pricing (per 1 M tokens) Benchmark rank (AA‑Briefcase)
Kimi K3 2.8 T 1 M $3 in / $15 out 2nd (1527)
Claude Fable 5 Max – – ~ $3 in / $15 out (Sonnet‑level) 1st
GPT‑5.6 Sol Max – – $2.5 in / $12.5 out (Terra) 2nd (≈1495)
Claude Opus 4.8 Max – – $2 in / $10 out 3rd (1600)

K3’s pricing is on par with Sonnet and slightly higher than Terra, while its benchmark scores place it between Fable 5 and Sol Max, making it a competitive option for agents that need large context windows and vision.

Practical takeaways

  • For developers needing long‑context or vision‑augmented agents, K3 offers a unique combination of size, context length, and open‑weight availability.
  • Cost‑sensitive teams should monitor reasoning token consumption; the default max‑effort mode can inflate expenses compared to more efficient models.
  • Researchers gain access to a 2.8 T parameter open model, enabling experiments that were previously limited to closed‑source labs.
  • Enterprises may consider K3 as a viable alternative to Anthropic or OpenAI for high‑throughput coding and knowledge‑work pipelines, provided they accept the current latency and rate‑limit characteristics.

Final assessment

Kimi K3 arrives as a large, open‑weight model with frontier‑level benchmark scores, a 1 M token context, and native vision, priced similarly to top‑tier commercial offerings. While its default max‑reasoning mode can be token‑hungry and its latency is modest, the model’s performance on agentic benchmarks and the upcoming release of its weights make it a noteworthy contender in the rapidly evolving LLM landscape.

Sources

Related