#open-weight-models

Mistral is the only frontier-class vendor whose EU answer is structural: Paris HQ, announced EU inference capacity, open weights. When the premium pays.

Qwen3.8-Max ships as a 2.4T-parameter MoE with 95B active and a 1M context. What the vendor benchmark table shows, and what is still only promised.

An industry letter on open weights grew from 25 to 50 signatories in two days. Behind it sit sanctions threats and quieter federal procurement levers.

Ant Group's Ling-3.0-flash is a 124B MoE with 5.1B params. Ant says it matches its 1T flagship; free on OpenRouter to Aug 3, but weights aren't posted yet.

OpenRouter added 10 frontier and open-weight models in 22 days, prices spanning ~83x on input and ~250x on output. The July 2026 releases worth routing to.

Meituan, the food-delivery super-app, now ships LongCat 2.0 — a 1.6T open-weight (MIT) model trained on Chinese chips. Ant Group and Tencent are doing the same.

Hugging Face says an autonomous AI agent — not a human — ran an end-to-end intrusion of its infrastructure, stealing internal datasets and credentials.

Alibaba previewed Qwen3.8-Max at WAIC Shanghai, claiming 2.4 trillion parameters and second only to Fable 5 — yet shipped zero benchmarks to back it.

Kimi K3's open weights are promised by July 27, not shipped — use the 10-day window to prep hosting for a 2.8T model and check the license before you commit.

Kimi K3's open-weight bet meets Anthropic's closed Fable 5. Vendor benchmarks split 6-8, K3 lists at $3/$15 vs $10/$50, and weights are promised July 27.

Five open-weight moves cluster in one July window — K3, Inkling, M3 Pro, a Mistral MoE teaser and a scheduled DeepSeek V4 — narrowing the gap to a generation.

Kimi K3 brings 2.8T parameters, 1M context, and near-frontier vendor benchmarks, with open weights due July 27, 2026. What the release means for AI buyers.

Mira Murati's Thinking Machines shipped Inkling, a 975B-parameter Apache 2.0 model built to be fine-tuned, not to top benchmarks. The customize-don't-rent bet.

Tencent shipped Hunyuan Hy3, a 295B open-weight MoE, under Apache 2.0 on July 6, 2026. It trails GLM-5.2 on coding but wins on agentic tasks at half the memory.

Pair Fable 5 as the orchestrator brain with GLM-5.2 as open-weight execution muscle. A concrete routing recipe, the cost math, and the single-model risk case.

GLM-5.2 is MIT-licensed and downloadable, but full weights run ~1.5TB and need datacenter GPUs. The hardware reality, quant ladder, and routes that work.

Match 2026's best open-weight coding models to your hardware: Qwen3-Coder-Next, Devstral 2, GLM-5.2 and DeepSeek V4 by VRAM, SWE-bench score and real speed.

US frontier-model gating is pushing global builders toward open-weight models China already dominates. Does Washington's lockdown hand Beijing the edge?

On June 26, 2026 the US gated two frontier models the same day: Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol. Inside the new managed-release era.

The Fable 5 export shutdown showed single-vendor AI can halt your business overnight. A four-step second-source playbook with open-weight failover backups.

DiffusionGemma is Google's first open-weight text diffusion LLM: a 26B MoE under Apache 2.0 hitting 1,100+ tokens/sec on one H100. Where it wins and loses.

OpenRouter added five models in ten days, from Opus 4.8 to MiniMax M3 at $0.30/M input. June 2026 pricing, context windows, and usage rankings.

Open-weight models now run 10-12x cheaper than frontier SaaS. A 2026 framework on the TCO crossover, vendor lock-in, and data sovereignty for agencies.

MiniMax M3 lands at 5-17x lower cost, but Opus 4.8 leads SWE-bench Pro and GPT-5.5 wins Terminal-Bench. A full three-way agentic coding routing matrix.

DeepSeek abandons its no-outside-capital stance in a ~$7.4B maiden round led by Tencent and CATL, valuing it near $59B and reshaping open-weight economics.

Cosmos 3 is the first fully open physical-AI omnimodel: one model reasons, simulates, and predicts robot actions. Inside the two-tower design and how to run it.

MiniMax M3 fuses frontier coding, a 1M-token context window, and native multimodality. Inside its Sparse Attention design, vendor benchmarks, and pricing.

StepFun's Apache-2.0 Step 3.7 Flash pairs a 196B MoE backbone with a 1.8B vision encoder, activating ~11B params per token. The cost case for agentic teams.

When self-hosting open-weight models beats API calls: a cost-crossover model, GPU sizing tables, and a deployment matrix for vLLM, SGLang, and Ollama.

Migrate DeepSeek V3.2 to V4 across open-weight stacks — three reasoning modes, tokenizer change, HCA/CSA attention deltas, KV-cache reduction.

GPU spend, ops headcount, latency, and break-even volume for hosting Llama, Qwen, DeepSeek, and Mistral yourself vs API. With per-token cost curves at 4 scales.

Cross-model quality regression, throughput lift, and VRAM savings at GPTQ-4, AWQ-4, INT8, and FP8 — benchmark data across 6 open-weight models.

Seven serverless inference providers compared on price, latency, model availability, and throughput. 60+ data points across 12 popular models.

Side-by-side input, output, cached, and batch pricing for 30 frontier and open-weight models across 12 providers. Updated April 2026 with 200+ price points.

Q2 2026 gap analysis between open-weight and closed-source frontier models — capability parity, cost economics, and the agency deployment decision tree.