Kimi K3 Is the World's First Open 3T Model - Beats GPT-5.6 Sol on Coding at $3 Per Million Tokens

Kimi K3 launched today from Moonshot AI - the world's first open 2.8 trillion-parameter model, available now on the public API, Kimi.com, and the Kimi Code terminal client. Full weights follow July 27. On GPU kernel optimization benchmarks run over 24 hours on NVIDIA H200 hardware, K3 outperformed GPT-5.6 Sol, Opus 4.8, and GPT-5.5, and matched Fable 5 in most tasks. Moonshot is transparent that K3 still trails Fable 5 and GPT-5.6 Sol on the broader evaluation suite - but for coding-heavy workloads, the benchmark gap is narrower than the pricing gap.

Kimi K3 Beats GPT-5.6 Sol and Opus 4.8 on GPU Kernel Optimization

Moonshot ran a controlled head-to-head. K3, Fable 5, GPT-5.6 Sol, GPT-5.5, and Opus 4.8 each worked independently in identical sandboxes for up to 24 hours, profiling and rewriting four kernel tasks across NVIDIA H200 and GPGPU hardware. K3 matched Fable 5 (which used Opus 4.8 as a fallback on some tasks) and outperformed the OpenAI models across the board. Moonshot used its own KimiCode harness for K3's evaluation - worth noting, since harness choice tends to favor models trained on that specific framework.

Beyond structured benchmarks, K3 built MiniTriton in a single agentic session - a full Triton-like GPU compiler with its own tile-level IR, optimization passes, and PTX code generation. On certain roofline benchmarks, MiniTriton matched or outperformed Triton. K3 also completed a 48-hour chip design run on the Nangate 45nm library: 4mm², 100MHz closing timing, 8,700 tokens-per-second decode throughput in simulation. Neither demo runs under reproducible controlled conditions, but both show how far autonomous long-horizon coding has moved since early 2026.

At $3 Per Million Tokens, Kimi K3 Undercuts Most Frontier Coding Rates

Kimi K3 starts at $0.30/MTok for cache hits, $3 for cache misses, and $15 for output tokens. Cache hits matter here. Moonshot reports cache hit rates above 90% in coding workloads on its official API, which means most tokens in a sustained engineering session land at $0.30 rather than $3. Teams comparing frontier API costs can use the AI API Pricing comparison for current rates across providers - but at that effective coding rate, K3 is among the cheaper options for any model operating at this benchmark tier. K3's 1-million-token context window makes the cache efficiency practical: sessions run long enough to accumulate meaningful savings.

China's Open-Source Labs Have Been Setting Parameter Records for Nine Months

Kimi K3 fits a pattern. DeepSeek disrupted the market in January with an open model that matched closed frontier performance at a fraction of the training cost; GLM-5.2 launched last week as a 1.6-trillion-parameter, MIT-licensed model posting better coding scores than GPT-5.5 at about one-sixth the price. Moonshot says Kimi models have held the upper bound of open-model parameter counts for nine of the past 12 months. None of these labs is treating open-source as a PR move - each release drops weights that research teams and enterprises can run, fine-tune, and deploy without license restrictions.

Weights Land July 27 - but Agent Builders Need to Read the Release Notes First

Full weights release July 27. Moonshot is coordinating with inference partners before dropping them, giving teams 10 days to test the API before the self-hosted path opens. For a broader view of how K3's coding scores stack up against the current frontier, the Grok 4.5 vs Fable 5 vs GPT benchmark breakdown covers the major models head-to-head with pricing.

Two limitations in Moonshot's own release notes matter for anyone building agents on K3. Both are architectural, not bugs. K3 requires harnesses that pass back complete thinking history between turns - if a framework strips or compresses that history, Moonshot warns generation quality becomes "highly unstable." Most popular harnesses like Claude Code handle this correctly, but custom setups built on raw API calls may not. K3's training on long-horizon, challenging tasks also makes it prone to autonomous decision-making when instructions are ambiguous; production deployments will need explicit behavioral constraints in the system prompt or in AGENTS.md to keep the model within defined boundaries.

Gemini 3.5 Pro has no launch date. K3 arriving with frontier-adjacent coding scores, open weights, and a $3/MTok API into the gap Google left is probably better timing than Moonshot planned for. Neither Google nor OpenAI has publicly commented on K3's parameter count record or the July 27 weight release.