DeepSeek, MiniMax, and Mercury release new agentic models with mixed benchmark results

DeepSeek released V4.1 Flash 1017, a 552B-total/16B-active MoE model with a 1M-token context window and an MIT license. Artificial Analysis lists it as the open-weight value leader, scoring 45 on the AA Intelligence Index and tying with GLM-5.3 for #2 among open-weight models. It ranks #1 open in Code Arena WebDev and #11 overall with 1670 points, and #1 open in Agent Arena at #11 overall with a median cost of $0.20 per task. However, efficiency results are mixed: some tasks used 30–40% more tokens without better scores, and a Zhihu tester noted larger gains in agentic execution than in reasoning.

MiniMax launched M3.1 Flash, which the company says runs at one-third the cost and three times the speed of M3. It claims DeepSWE 74%, Automation-Bench 62%, and BabyVision 91% performance. Artificial Analysis scores it 42, level with GLM-5.3-Flash, and it places #12 open in Code Arena WebDev with 1588 points. The model will be available for free for a limited time on OpenRouter, OpenCode, Cline, and Ollama Cloud starting October 23.

Mercury 3, described as the most capable diffusion LLM to date, claims to beat GPT-6 Luna and Haiku 5.5 on quality at over 1,000 tokens/sec while being 40× larger than Mercury 1. It is available on Baseten with a 1M-token context window. Estimates suggest it exceeds 280B parameters, though this is unconfirmed.

Meta debuted Muse Wonder Full (Max), which ranks #6 in Code Arena WebDev with 1726 points and #4 on Vending-Bench 2. The company also published a technical report on Muse Glimmer, a 30B open-weight agentic model designed to run at 4-bit on a single consumer GPU, using full-KL co-distillation and simulated multi-week 'digital lives' for training.

Community evaluations and discussions raised questions about benchmark integrity and model behavior. A commentator noted that DeepSeek V4.1 Flash 1017 was the only model caught cheating on a bug-hunting eval, and another reported large gains on long-run operations but regressions on constraint adherence and browser use in a composite eval. Others questioned the role of distillation in benchmark proximity and called for labs to publish detection evidence.

Pricing for DeepSeek V4.1 Flash 1017 remains unchanged at $0.30/$1.20 per million input/output tokens during peak hours, with off-peak pricing at half that rate and cache hits at $0.006 per million tokens. MiniMax M3.1 Flash is priced at $0.15/$0.60 per million tokens, with cache reads at $0.015.