The new open-model stack

Kimi and GLM changed the map. Price the hardware behind them.

The current open-model conversation is sparse, agentic, multimodal, and long-context. Compare Kimi, GLM, Qwen, MiniMax, StepFun, DeepSeek, and gpt-oss by what actually shapes deployment: resident weights, active parameters, context, VRAM, and live GPU cost.

Kimi GLM Qwen MiniMax StepFun DeepSeek
Every entry links to its official model card. Licenses are named explicitly, including MiniMax M3's commercial notice and authorization conditions.
Current catalog
curated open-weight models
Single GPU
models that fit on one card
Sparse by design
mixture-of-experts models
Cheapest Estimate
waiting for live pricing
Cluster ceiling
largest baseline VRAM envelope

Three lanes matter now

The useful split is no longer small versus large. It is practical sparse models, frontier agentic clusters, and differentiated challengers.

01 / Practical

One-card MoE

GLM-4.7-Flash, Qwen3.6, and Kimi Linear put 3B active paths inside much larger models. This is where most teams should begin.

02 / Frontier

Agentic clusters

Kimi K2.7 Code and GLM-5.2 target long-horizon engineering. Their active paths are efficient; their resident weights still demand serious clusters.

03 / Contenders

Different bets

MiniMax M3, Step 3.7 Flash, DeepSeek-V3.2, and gpt-oss each trade license, context, runtime maturity, and memory differently.

Start with the workload, then check the resident weights

Filter by agentic coding, reasoning, long context, or multimodal work. Active parameters hint at token-time compute; the VRAM baseline reflects the much larger set of weights that must stay resident.

Model metadata verified from source cards and vendor docs, then mapped to the tracked GPU catalog. Use the Hugging Face links on each model to inspect the upstream card yourself.

Loading price basis…
Loading the current open-model stack...

Deployment patterns for the models that still fit a bursty workload

The practical flow is usually: pick a model on Hugging Face, use a vLLM-compatible runtime, cache weights aggressively, and choose how much platform management you want to own.

Loading serverless deployment patterns...

Move from model choice to provider and GPU

Compare the lowest tracked vLLM entry points, provider tradeoffs, and model-specific GPU guides without leaving the pricing data.

Compare provider-specific prices for common LLM workloads

These pages start from a provider and narrow its live GPU catalog to vLLM serving, batch inference, fine-tuning, or training-friendly rows.

Model hosting cost by provider

These pages answer the next question after model selection: what does this model cost on a specific GPU cloud?

What the leading open-model labs are publishing now

This feed follows Moonshot AI, Z.ai, Qwen, MiniMax, StepFun, DeepSeek, OpenAI, and Mistral, then screens out merges, repacks, and low-signal variants. We still show release-kind labels because upstream orgs sometimes publish multiple packaging variants around the same core release.

Signal badges are directional reads from metadata and release context, not benchmark leaderboards.

Loading live Hugging Face feed…
Loading latest discovered models...

Quick read on what it takes to host each model

This table is optimized for planning: total versus active parameters, context, memory floor, release timing, and a live cost estimate.

Model Best For Total / Active Context Minimum Setup Released Cheapest Tracked Hosting
Loading model table...

How the hosting estimates are computed

The catalog mixes curated model metadata with live GPU pricing. Use it as a planning tool, then add headroom for your workload.

Loading assumptions...

Hand this runbook to Claude Code or Codex

Open a terminal in the repository where you want the deployment files, start claude or codex, then paste this prompt. It asks the agent to verify sources and stop before it creates billable infrastructure.

Deployment prompt
Use the infrastructure or model context on this page to create a reproducible open-model deployment. Use this guide as the starting context: https://www.getflops.ai/llms. Read the linked model card and provider documentation before choosing hardware or runtime settings. Open every linked primary source and flag any mismatch instead of guessing. Create a deployment folder containing README.md, .env.example with no secrets, a pinned start script or infrastructure manifest, and smoke-test.sh. Make the endpoint OpenAI-compatible where the runtime supports it. Run local/static validation, estimate the billable resources, and stop before provisioning paid infrastructure until I approve.
Guardrails included No secrets in files · verify primary docs · approval before spend