Open weights and cheap inference this week: frameworks, diffusion speed, Flash cuts, and flagship weight timelines
DeepSeek open-sourced production inference framework DeepInfer; Inception, MiniMax, Kimi and Mistral×NVIDIA each moved a fast diffusion model, an agent-oriented Flash price cut, a 3T-class open-weight flagship plan, and an industrial-engineering weight timeline. What matters is license, channel and price differences—not one “open-source boom” or blanket price war.
DeepSeek released DeepInfer on GitHub under Apache-2.0, describing it as the inference framework behind its large-scale production services. This cut offers offline inference first, starting with DeepSeek-V4.1-Flash on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs, and omits full serving pieces such as a KV-cache system, prefill/decode split and load balancing. It is about running DeepSeek’s own models, not shipping a new foundation-model generation.
Inception Labs launched diffusion LLM Mercury 3, claiming parity-or-better versus GPT-6 Luna and Claude Haiku 5.5 on several benches at ~1,000+ tok/s on NVIDIA GPUs (~4–8× those speed-tuned models), with list $/M token prices in the blog. MiniMax formally shipped M3.1 Flash (follow-on to the anonymous Space Bunny that processed 200T+ tokens), stressing agent/coding gains, ~3× online inference speed and ~70% usage-cost cuts at a unified ~$0.15/M input. Both sell “faster/cheaper,” but one is a diffusion API model and the other a vendor Flash generation—benchmarks and prices remain vendor claims.
Moonshot/Kimi announced Kimi K3.1 Pro as its strongest 3T-class open-weight model to date with ~1M context: apps and API available now, full weights planned for 2026-11-04, plus a Highspeed variant ~2–3× faster. Separately, Mistral and NVIDIA described industrial physics-AI workflows around Large 4 (~1T total / ~49B active parameters) with weights planned for end of October. Open-weight timelines, downloadability and industrial-toolchain coupling still need separate checks—not one “everything went open this week” headline.
Sources
- DeepInfer: High-performance, large-scale inference framework for DeepSeek models — DeepSeek
- Diffusion at Scale: Introducing Mercury 3 — Inception Labs
- MiniMax M3.1 Flash: Redefining Agent Performance and Cost — MiniMax
- Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence — Kimi / Moonshot AI
- Mistral Advances AI for Engineering with NVIDIA — Mistral AI