Open-weight models are now fighting on agentic benchmarks, and the gap is closing fast

What the new models claim

Moonshot AI says Kimi K3.1 Pro is a 3T-class open-weight model with a 1-million-token context window, and the company says full model weights will be released by November 4, 2026. The company says the model scored 65.7% on Terminal-Bench 4.0, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5 on the same benchmark, and 92.7% on the BenchCAD benchmark. Those are the company's own numbers, and they frame the model as a tool for terminal and design work rather than a general chatbot. m:…7929

Moonshot AI also says the model's claim-level hallucination rate fell from 23.4% for K3 to 8.3%, and that it improved on K3 by 7.2 points in Financial Modeling benchmarks. The company says it removed cyber-related data from both training stages for the open-weight release, and that the model's cyber capability remains behind that of frontier proprietary models publicly available for at least three months. That is a deliberate trade: the company is presenting openness with guardrails, not openness without qualification. m:…7929

MiniMax, meanwhile, says it has officially open-sourced the weights for M3.1 Flash, a model with 428B total parameters and 23B active parameters. The company says users processed over 100 trillion tokens using M3.1 Flash within 72 hours of its launch. The design choices the company names are specific: the first three layers use sparse attention to reduce long-context calculation burden, MoE experts use W4A4 quantization, and power-aware scheduling optimizes hardware power consumption to maintain compute performance. m:…6137

Why the serving layer matters as much as the weights

Weights alone do not make a model usable, and the serving announcements tell the second half of the story. SambaNova says MiniMax M3.1 Flash is now available on SambaCloud, delivering over 200 tokens per second, at 3x the speed of its predecessor and 70% lower usage costs. The company lists cached input tokens at $0.03 per 1M tokens, input tokens at $0.30 per 1M, and output tokens at $1.20 per 1M. The company also says MiniMax M3 on SambaCloud is being rerouted to M3.1 Flash, which means existing users get moved to the new model whether they asked or not. m:…2627

The benchmark claims SambaNova leads with are agentic, not conversational: the company says the model scores 73.7% on DeepSWE v1.1, 82% on CyberGym, 61.7 on OSWorld V2, and 76.2 on BankerToolBench, alongside 29.6 on Agents' Last Exam. That spread is the point. Providers are no longer selling a model that chats well; they are selling one that does software, browsing, and tool work at a stated price. A developer choosing between providers now has to weigh those task scores against the token prices, and both numbers come from the seller. m:…2627

NeuroSpark adds the hardware angle. The company says it pairs AMD Instinct GPUs on OCI with ROCm software to serve inference workloads, matching models to hardware based on size, latency, and performance requirements. The company says compute costs have become the highest expense for many of its customers, and that AMD Instinct GPUs on OCI provide a lower-cost option for advanced models. It also says it brought Kimi K3, MiniMax M3, and Reflection's Beam into production on AMD, and that it supported GLM-5.3 on AMD from day zero after porting support from GLM-5.2 using ROCm. m:…8943

What to weigh before adopting

The implication is that the open-weight frontier is now a three-way negotiation between the lab, the inference provider, and the chip stack. MiniMax says M3.1 Flash is supported by inference platforms including Together AI, Fireworks, Nebius, Baseten, Modal, and DigitalOcean, and by frameworks including vLLM, SGLang, and BAAI FlagOS, with weights available on Hugging Face and ModelScope. MiniMax also says the model supports domestic Chinese chips including Huawei Ascend, T-Head, Cambricon, Moore Threads, Biren, Haiguang, Iluvatar CoreX, and MetaX, as well as international hardware from NVIDIA and AMD. That breadth is itself a selling point: the same weights are meant to run almost anywhere. m:…6137

But every number in this story comes from a seller, and the sellers disagree with each other in useful ways. Moonshot AI says its model's cyber capability trails frontier proprietary models, while SambaNova leads with CyberGym and DeepSWE scores for MiniMax. A developer should treat those as claims to test on their own workload, not as settled rankings. The practical question is narrower: does the model do your agent task at the provider's stated token price, on hardware you can actually buy? The releases give you the numbers to start that test, and the open weights mean you can run it yourself. m:…7929 m:…2627