Kimi K3 Released: 2.8T Open Source Model with 1M Context — What Coding Teams Pay to Run It

The First Open 3T-Class Model Aimed at Coding Workloads

Moonshot AI announced Kimi K3 on July 17, 2026 — a 2.8 trillion parameter model that will be released with open weights in two weeks. The headline features are native vision, a one million token context window, and two architectural bets the paper calls out explicitly: KDA (Kimi Dynamic Attention) hybrid linear attention, and attention residual (AttnRes) connections that let information flow around expensive attention layers when they aren't needed.

For coding teams, the question isn't whether K3 is technically impressive — it's whether running a 2.8T model yourself is cheaper than paying a closed-source API. The math changes depending on how many tokens you burn per day and whether you already own hardware.

Self-Hosting a 2.8T Model in July 2026

A 2.8T parameter MoE model at BF16 requires roughly 5.6TB of memory just for weights. Even with 4-bit quantization dropping that to about 1.4TB, you're looking at a cluster of eight to sixteen H200 or B200 GPUs to hold the model plus KV cache for a 1M context window. Rented from a hyperscaler, that's $80-$180 per hour when running. Amortized on-prem, expect $1.2M-$2.4M upfront plus power and cooling.

If your team burns 5 billion tokens per month across coding tasks, self-hosting starts to compete around the 20 billion token per month mark. Below that, an open-weight API rental — which several providers will offer within days of Moonshot's weight release — is almost always cheaper.

API Pricing for Comparable Open-Weight Coding Models

Until K3's official hosted pricing lands, here is how the open-weight coding stack looks today, drawn from the current pricing dataset behind our AI cost calculator:

Model	Provider	Input $/M	Output $/M
Kimi K2.7-Code	Moonshot	$0.74	$3.50
DeepSeek V4 Pro	DeepSeek	$0.435	$0.87
Qwen3 Coder	Alibaba	$0.22	$1.80
GLM 4.7 Flash	Zhipu	$0.06	$0.40
Llama 4 Maverick	Meta	$0.15	$0.60
Ring-2.6-1T	inclusionAI	$0.075	$0.625
DeepSeek V4 Flash	DeepSeek	$0.09	$0.18

Kimi K3 will need to price hosted access somewhere between K2.7-Code ($0.74/$3.50) and a full frontier tier to be competitive on both raw capability and cost. Given the pattern for other 2T+ open-weight releases, expect first-day hosted rates from third-party providers around $0.60-$1.20 input and $2.50-$5.00 output per million tokens.

Where the 1M Context Window Actually Saves Money

A million token context sounds like an invitation to overspend. In practice, it saves money on a specific class of tasks: repository-wide code review, large-scale refactoring across many files, and analysis of long production logs. When you fit the whole context in one call, you avoid the retry loops and re-loading that eat tokens in shorter-context workflows.

The trap is filling the window on tasks that don't need it. A 50 line bug fix does not benefit from 950K tokens of surrounding repo context — you'll pay input tokens for material the model never actually uses in its reasoning. Use the full window deliberately, not by default.

Native Vision — Cost Implications for Coding Teams

K3's native vision means image tokens are billed the same as text tokens in most pricing schemes. Screenshots of design mockups, error dialogs, and architecture diagrams no longer need a separate vision model, which historically added 30-50% to per-task costs for teams doing UI work. The savings compound if you already run design-to-code workflows.

Should You Wait for the Weights?

Two weeks from now, when the weights land, three things happen: third-party hosting providers publish rates, benchmark scores get independently verified, and inference optimizations start rolling into vLLM and SGLang. If your budget is below $10K per month on coding models, skip the self-host path entirely and buy hosted K3 access from whoever offers the best rate.

Teams already running $30K+ monthly and doing repository-scale work are the ones who should start planning hardware procurement now. The savings only appear at that scale, and only if you have engineers who can babysit a large inference cluster.

Bottom Line

Kimi K3 is a milestone for open-source, but it doesn't change the fundamental cost math: hosted APIs win under 20 billion tokens per month, self-hosting starts making sense above that if you have the ops capacity, and the models that will actually save you money in 2026 are the sub-$0.20 input ones already available today. Use our cost estimator to see how K3-tier pricing changes your total build cost against models like DeepSeek V4 Flash and GLM 4.7 Flash.