DeepSeek V4 Pro Closes In on GPT-5.2 — 17x Cost Advantage on Agent Benchmarks

Over on Reddit's LocalLLaMA community, word is spreading that DeepSeek V4 Pro has delivered results on par with GPT-5.2 on the agent benchmark "FoodTruck Bench." FoodTruck Bench is a grueling test where a model autonomously runs a food truck business for 30 days using 34 different tools.

The results were astonishing. DeepSeek V4 Pro ranked 4th overall (behind Opus 4.6, GPT-5.2, and Grok 4.3), marking the first time a Chinese-made model has reached the frontier tier. It caught up to the score GPT-5.2 posted in mid-February just 10 weeks later, a sign that the AI performance gap between China and the US is rapidly shrinking.

The price gap is even more dramatic. While GPT-5.2 costs $1.75/M input and $14/M output, DeepSeek V4 Pro charges $0.435/M input and $0.87/M output — roughly 17x cheaper for the same agent workload. It even offers a discount on cache reads, putting it 2nd in cost efficiency behind Gemma 4 31B.

Meanwhile, Xiaomi MiMo v2.5 Pro took 6th place on the same benchmark, giving Chinese models two spots in the top 6.

> First Chinese model to land in the frontier tier on our benchmark. Tied with Grok 4.3 Latest on outcome, within 3% of GPT-5.2's median.

Qwen3.6-27B FP8 Runs 200K Context at 80 TPS on an RTX 5000 PRO 48GB

Also on LocalLLaMA, a user report on running the official FP8 quantization of Qwen3.6-27B on an RTX 5000 PRO 48GB has been drawing attention. They secured a 200K-token context with a BF16 KV cache while recording 60–90 TPS during coding.

The key is a setup that's close to "barely quantized at all." It combines FP8 quantization of the model body with Blackwell hardware acceleration while keeping the KV cache in BF16, so the accumulation of quantization error is kept to a minimum. Reportedly, even long agent sessions are less prone to early compaction or infinite loops.

The configuration is vLLM 0.20.1 + CUDA 12.9, with MTP (Multi-Token Prediction) set to 2 tokens. GPU memory utilization is pushed up to 97.5%.

> I think we might have an answer to all those "what do I buy for $10k?" posts. A pro5k, 64GB RAM, a decent CPU/mobo, and it will run the FP8 quant of 27B.

MTP (Multi-Token Prediction) Comes to llama.cpp — Current State of Supported Models

Multi-Token Prediction (MTP) is a technique that speeds up inference by predicting multiple tokens at once. Support has already progressed in vLLM and other stacks, and it is now expected to land in llama.cpp soon.

A list of models slated for support has been shared, including DeepSeek V3/V3.2/V3.4, Qwen 3.5, GLM 4.5+, MiniMax 2.5+, Step 3.5 Flash, and MiMo v2+. At this stage you still need to download the HuggingFace weights and convert them to GGUF, but MTP adoption in local inference environments looks set to take off in earnest.

US GUARD Act — Age Verification Mandate for AI Chatbots Moves Forward

The US Senate Judiciary Committee has unanimously sent the GUARD Act to the full Senate. The bill would mandate age verification for AI chatbots — nominally in the name of child safety, but in substance it tightens access control over AI platforms.

In the LocalLLaMA community, critical voices dominate, with comments such as "serious concerns about privacy infringement" and "the US is now following the EU's style of regulation" — and some see the value of local AI rising even further.

Retroguard — Cryptographically Verifiable AI Guardrails

A new service called Retroguard was introduced on Hacker News. It offers AI guardrails with cryptographically secure, verifiable robustness. Featuring drop-in integration and an outcome-based pricing model, it presents an approach to controlling AI system outputs that differs from traditional rule-based or heuristic methods.

Ψ-RAG — Efficient RAG via Hierarchical Abstract Trees

Ψ-RAG, presented on Hugging Face Daily Papers, is a new RAG framework built on a hierarchical abstract tree index. It achieves efficient and precise retrieval in logarithmic time, and by combining a multi-granularity agentic retriever with a hybrid retrieval pipeline, it is designed to handle diverse user requests.

The research targets performance gains in complex information retrieval scenarios and is drawing attention as an approach that rethinks the very structure of RAG.

---

**Sources:**

[DeepSeek V4 Pro matches GPT-5.2 on FoodTruck Bench — Reddit](https://www.reddit.com/r/LocalLLaMA/comments/1t47qbw/deepseek_v4_pro_matches_gpt52_on_foodtruck_bench/)

[Qwen3.6 27B FP8 runs with 200k tokens — Reddit](https://www.reddit.com/r/LocalLLaMA/comments/1t46klu/qwen36_27b_fp8_runs_with_200k_tokens_of_bf16_kv/)

[MTP prepares to land in llama.cpp — Reddit](https://www.reddit.com/r/LocalLLaMA/comments/1t46o09/as_mtp_prepares_to_land_in_llamacpp_models_that/)

[US GUARD Act: Age Verification for AI Chatbots — Reddit](https://www.reddit.com/r/LocalLLaMA/comments/1t465yx/us_guard_act_age_verification_for_ai_chatbots/)

[Retroguard – Verifiably secure AI guardrails — Hacker News](https://retroguard.ai)

[OpenAI Raises $4B for The Deployment Company — Hacker News](https://officechai.com/ai/openai-raises-4-billion-for-the-deployment-company-to-help-businesses-leverage-ai/)

[Hierarchical Abstract Tree for Cross-Document RAG — Hugging Face Papers](https://huggingface.co/papers/2605.00529)