Foundation Models: New Benchmarks and Inference Optimizations

What Actually Changed

FusionEval was introduced as a new benchmark for evaluating model fusion capabilities, addressing gaps in existing evaluations for reasoning with long chains of thought and tool use. MoEZip was developed as a routing-aware KV cache compression method specifically for sparse Mixture-of-Experts language models, addressing memory bottlenecks in long-context inference. Kimi K3.1 Pro was launched with competitive pricing at $2.00/million tokens for input and $10/million for output, matching GPT-6.1 Sol and undercutting Claude Opus 5.5. Liquid AI released RIPL, a real-time inference engine achieving 14.4 to 28.4x lower p95 latency than comparable models.

Why It Matters

These developments address critical bottlenecks in foundation model deployment and evaluation. FusionEval fills a significant gap in evaluating model fusion capabilities, which is becoming increasingly important as organizations seek to combine specialized models. MoEZip specifically tackles the memory bottleneck in MoE models, which is crucial for scaling these architectures. Kimi's competitive pricing strategy signals continued price pressure in the market, potentially forcing other providers to adjust their pricing models. Liquid AI's RIPL demonstrates that significant latency improvements are possible with specialized inference engines.

Capability and Benchmarks

FusionEval introduces controlled expert pools containing domain experts independently trained from Qwen3 and Llama base models across math, code, science, instruction following, and tool use. The benchmark provides a unified evaluation suite that can assess capability retention during model fusion. Kimi's own benchmarks show their K3.1 Pro model performing approximately at the level of Claude Opus 5.5, though independent verification is needed. Anthropic reported notable improvements in cybersecurity capabilities for Claude Opus 4 compared to prior models, though specific metrics were not provided. The Route-and-Generate framework was introduced to enable LLMs to adaptively decide when and how much reasoning is required for each input.

Compute, Cost and Pricing

Kimi K3.1 Pro is priced at $2.00/million tokens for input and $10/million for output, matching GPT-6.1 Sol and half the price of Claude Opus 5.5. They also offer a 'Highspeed' version that's 2-3x faster at twice the price, though input token cache hits are priced the same. DeepSeek's DeepInfer is reportedly running 155T tokens per day per cluster, making $22m/day at approximately 80% margin with an 81-day payback period, though these figures come from the company itself and require independent verification. You.com launched an inference router for 12 open-weight models with a 99.9% uptime SLA and no platform fee.

What We Are Watching

We're monitoring the independent verification of Kimi K3.1 Pro's capabilities against its claimed performance parity with Claude Opus 5.5. The adoption of You.com's inference router with its 99.9% uptime SLA could signal a shift toward more reliable API infrastructure. We're tracking how MoEZip's routing-aware KV cache compression affects generation quality in real-world deployments. The promised release of Kimi's full model weights by November 4, 2026, will be an important development for the open-weight ecosystem. We're also watching how the Route-and-Generate framework's reasoning allocation capabilities translate to practical efficiency gains across different task types.