The Asia cadence overtakes the West — Qwen, DeepSeek, Mistral keep shipping while Llama goes quiet
Mistral shipped frontier-class models every six weeks through 2026. Qwen 3.7 Max landed May 20 at top-tier SWE-bench. DeepSeek V4 in April. Llama 5 still unreleased thirteen months after Llama 4. The open-source frontier cadence is now driven by Asia and continental Europe; the US-headquartered OSS frontier has gone quiet.
The Llama 4 launch in April 2025 was Meta's signal that it intended to lead the open-source frontier. The signal did not survive 2025. Thirteen months later there is no Llama 5, the Llama series is in maintenance mode, and the open-weights cadence has been carried by Asia and continental Europe. The shift is structural enough to merit its own narrative.
Mistral Medium 3.5 shipped May 3, the latest in a 2026 cadence that has produced Large 3, Small 4, Ministral 3, Voxtral TTS, Leanstral, Forge, and now Medium 3.5 — one frontier-class release every six weeks. Qwen 3.7 Max landed May 20 (covered AM cycle). DeepSeek V4 in April. Gemma 4 in April. IBM Granite 4.1 family in April. The OSS frontier is shipping at a velocity the Western labs' headline-model cycle doesn't match.
Why the cadence matters more than any single model
Frontier-class models age fast. A model that's top-tier on SWE-Verified in April is mid-tier in July; the field iterates on capability faster than enterprises can validate deployments. The labs that ship every six weeks are constantly at the frontier; the labs that ship every 12-18 months are at the frontier intermittently. Cadence is the structural advantage that compounds.
The self-hosted enterprise deployment surface for OSS models has matured to procurement-default status in the same window. The combination — frontier-class quality, six-week cadence, mature deployment stack — makes the OSS frontier a real competitor to the closed cloud-API stack in 2026, not just an aspirational alternative.
What the Meta retreat actually means
Meta's 8,000 layoffs concurrent with $115-135B capex investment in AI infrastructure (covered AM cycle) is the loudest signal that the company is reorganizing around AI but not around open-weights. The capex is going into proprietary infrastructure and applications, not into a Llama 5 that competes for credit with closed-source labs. The strategic call appears to be: ship apps that use frontier AI, not models that compete at the frontier.
If that's the call, the open-weights frontier is going to belong to Mistral, Alibaba (Qwen), DeepSeek, IBM (Granite), Google (Gemma — uncharacteristically OSS), Microsoft (Phi), and the various long-tail Asian labs. The Western OSS frontier has, in practice, been replaced by a multi-continental OSS frontier where the Asian and European labs ship faster than the US can.
The procurement consequence
For enterprise buyers, the implication is straightforward: an OSS-first procurement strategy is now viable across every model-capability tier. Top-tier (Mistral Medium 3.5, Qwen 3.7 Max, DeepSeek V4): close to closed-frontier quality at 10× lower inference cost. Mid-tier (Gemma 4, Phi-4, Granite 4.1): purpose-built for specific deployment patterns. Long-tail: increasingly specialized models for vertical use cases. The 2025-era objection — "OSS is behind closed by a meaningful margin" — was true through last summer. It's not true now.
The frontier-lab response is one of the most consequential strategic questions of the next eighteen months. Anthropic's bet is Claude Managed Agents with self-hosted sandboxes — keep the orchestration cloud, push the execution local. OpenAI's bet is Stargate-scale infrastructure that no enterprise can match. Whether either is sufficient against an OSS frontier that ships every six weeks is the question every Series F frontier-lab pitch has to answer convincingly.
Codersera — Best Open-Source LLM May 2026 Llama Qwen DeepSeek Gemma Mistral → · Hugging Face — Best Open-Source LLM Models 2026 → · Gosign — Self-hosted Open-Source AI 2026 Enterprise →