AI2 Releases "EMO", a MoE Model with Domain-Specific Routing

Allen AI (AI2) has released a new Mixture-of-Experts (MoE) model called "EMO". It has 1B active / 14B total parameters and was trained on 1 trillion tokens — and what makes it especially notable is its document-level routing mechanism.

While conventional MoE models select experts on a per-token basis, EMO looks at the entire document to assign the appropriate experts. As a result, experts have been observed to cluster by domain, such as healthcare and news, achieving semantic domain partitioning rather than superficial pattern separation.

Routing quality is an issue that directly affects real-world MoE performance, and this document-level semantic routing approach could influence future MoE design.

Local LLM Optimization Is Accelerating Fast — Qwen 35B-A3B Becomes Practical on 12GB VRAM

A detailed benchmark of running Qwen3.6-35B-A3B (a MoE model) on an RTX 3060 12GB was reported in the Reddit LocalLLaMA community.

The key results are as follows:

Achieved about 89 tok/s prompt processing and about 43 tok/s generation at 32K context

GB VRAM acts as a sweet spot for tuning how much of the MoE is offloaded to the GPU

Q8 KV cache can be used with virtually no performance loss

Sufficient speed and context length for practical coding use cases

Because MoE models have few active parameters, they make it easier to reap the benefits of large models even on consumer GPUs with limited VRAM. A 35B-class model running at practical speeds on 12GB VRAM is an achievement that significantly expands the options for local LLMs.

Processing 262K Context at 80+ tok/s with MTP + TurboQuant

Also on LocalLLaMA, results were reported from running Qwen3.6-27B on an RTX 4090, combining MTP (Multi-Token Prediction) with TurboQuant's TBQ4_0 (a lossless 4.25 bpv KV cache).

It achieved generation speeds of 80–87 tok/s at 262K context, with an MTP draft acceptance rate of about 73% — nearly double the 43 tok/s before optimization. The work is available as a llama.cpp fork, so the community can reproduce it.

Being able to process contexts of more than 260K tokens at this speed on local hardware is a major step toward localizing workloads that have previously relied on cloud APIs, such as long-document processing and large-scale codebase completion.

MTP Effectiveness Depends on the Workload — Acceptance Rate Is Key

At the same time, it has also become clear that MTP is not always beneficial. Detailed benchmarks using Gemma 4-26b-a4b show large differences in MTP effectiveness depending on the workload:

Code generation**: 75 tok/s → 114.8 tok/s (1.53x faster, 66% acceptance rate)

Long-form writing**: 75 tok/s → 71.1 tok/s (essentially unchanged, 31% acceptance rate)

JSON output**: 51.3 tok/s → 25.6 tok/s (dropped to 0.5x, 8% acceptance rate)

According to the analysis, once the token acceptance rate falls below 50%, MTP's overhead outweighs its benefits. While MTP is a powerful speedup technique for predictable sequences like code generation, it can backfire for structured output and creative text generation.

These results show that the decision to adopt MTP should be based on real-world use cases.

Pentagon: "Never Again" Will It Depend on a Single AI Provider

A U.S. Department of Defense official has explicitly stated that the Pentagon will "never again depend on a single AI provider." According to reporting by Nextgov, this appears to reflect lessons learned from past overreliance on specific vendors.

The explicit policy of avoiding vendor lock-in and using multiple providers in parallel for AI procurement in the military and national security space could also affect the competitive structure of the AI industry. As competition over large government contracts intensifies, this is an important signal that helps preserve healthy competition for AI companies.

Investors Demand Refund After Pagaya's AI Real Estate Fund Loses 80% of Capital

It has come to light that investors in an AI-driven real estate fund managed by Pagaya, an Israeli AI fintech company, are demanding a refund after 80% of their capital was wiped out. According to Calcalist, the investors claim that the fund's performance fell far short of the AI model's predictions.

The case casts doubt on the risk management of AI-powered financial products, highlighting the fundamental question of how well AI models' predictive accuracy holds up in real market conditions.

---

**References:**

[new MoE from ai2, EMO](https://www.reddit.com/r/LocalLLaMA/comments/1t7kgy4/new_moe_from_ai2_emo/) — Reddit LocalLLaMA

[Qwen 35B-A3B is very usable with 12GB of VRAM](https://www.reddit.com/r/LocalLLaMA/comments/1t7l56a/qwen_35ba3b_is_very_usable_with_12gb_of_vram/) — Reddit LocalLLaMA

[Got MTP + TurboQuant running — Qwen3.6-27B -- 80+ t/s at 262K context on a single RTX 4090](https://www.reddit.com/r/LocalLLaMA/comments/1t7kyju/got_mtp_turboquant_running_qwen3627b_80_ts_at/) — Reddit LocalLLaMA

[MTP is all about acceptance rate](https://www.reddit.com/r/LocalLLaMA/comments/1t7mdrl/mtp_is_all_about_acceptance_rate/) — Reddit LocalLLaMA

[Pentagon will 'never again' rely on a single AI provider, official says](https://www.nextgov.com/artificial-intelligence/2026/05/pentagon-will-never-again-rely-single-ai-provider-official-says/413399/) — Nextgov

[Pagaya investor seeks refund, claiming AI real estate fund wiped 80% of capital](https://www.calcalistech.com/ctechnews/article/3a0p63mof) — Calcalist