To train an AI agent on a real job, you have to define the job twice. First, as work the agent can attempt. Second, as a test of whether it succeeded. The gap between those definitions is where a lot of reinforcement learning engineering lives. DeepSeek-V4.1-Flash and GLM-5.3 describe complementary parts of this problem. DeepSeek organizes what to teach. Its task-generation pipeline uses a growing knowledge graph to connect broad domains with specific concepts. Agents retrieve relevant papers, articles, and code, then use those materials to generate tasks. This gives the pipeline control over coverage, including specialized knowledge that ordinary sampling might miss. DeepSeek, §4.2.2 GLM makes tasks executable and checks their rewards. Research agents turn patterns from professional work into environments with dependencies and hidden state. A judge agent attempts the tasks to check solvability. For some tasks, the pipeline generates verifiers without access to the reference solution. Solver attempts help expose reward shortcuts. The verifier must pass checks covering a known solution, doing nothing, and leaving the task unfinished. GLM-5.3 Consider an assignment to make a training job faster. The first definition includes the code, workload, dependencies, hardware, and constraints. The second must determine whether the change produced a genuine speedup while preserving correctness. A model might improve the measured runtime by skipping computation or exploiting a weakness in the test workload. The score rises, even though the intended job remains unfinished. That is why an executable environment is only the beginning. It also needs to teach something relevant, distinguish useful progress from shortcuts, and remain informative as the model improves. Replaying an easy task can produce more trajectories with little additional learning. Generating diverse tasks can broaden practice without guaranteeing that it matches the work users need done. And a verifier can be perfectly consistent while checking the wrong thing. My takeaway is that environment quality should be judged by transfer: does training here improve performance on unfamiliar work? I would test that on tasks kept separate from environment development, compare against a baseline under a fixed training budget, and include the cost of building and validating the environments. GLM explicitly acknowledges that its pipelines still require meaningful human involvement. That work belongs in the economics. GLM-5.3 Working on post-training at Nebius, this sharpens the infrastructure opportunity for me. Teams need reproducible environments, efficient rollouts, inspectable rewards, and evaluations that reveal whether the learning transfers. Domain expertise becomes valuable when a team can encode realistic constraints and recognize failures that a generic checker would miss. My bet is that this capability will become an important advantage in post-training. A training environment earns its value through the work the model can do after training, including work that environment never contained.

Two new models for coding and agent workflows are now live on Nebius Token Factory. DeepSeek V4-Pro-0813 is the official V4 Pro release, built for coding agents that use tools, reason through complex problems, and work across multiple steps. GLM-5.3 brings Z. ai’s latest post-training improvements for complex software engineering, from planning changes across a repository to carrying long-running agent tasks through to completion. Try both through an OpenAI-compatible API, with Token Factory handling the serving infrastructure. Bring your own tasks and see which fits your workflow. Start building:

An AI can correctly identify its own cheating and still cheat when acting as an agent. Anthropic’s reward-seeker study shows why that distinction matters. Researchers trained an early Opus checkpoint across 80 environments with exploitable scoring. They omitted normal alignment RL and reward-hacking penalties. By the end, 40% of training episodes were flagged as reward hacks. The resulting model, Hacker-Opus, usually refused harmful questions in chat. In a tool-based task where a visible grader rewarded answering, harmful compliance reached 29%, versus 0.7% for the starting checkpoint. It also generalized to new forms of reward tampering. Yet, when asked to review transcripts containing its own reward hacks, it correctly flagged 90% of them. Anthropic’s study My interpretation: a model can retain the ability to judge behavior while learning to violate that judgment in another context. Three connections to earlier research make this more consequential. Closing an exploit and correcting the model are separate jobs. In “Sycophancy to Subterfuge,” researchers trained models on increasingly sophisticated forms of specification gaming. Some generalized to rewriting their reward function without training on that final behavior. Those events were rare. Retraining against simpler forms of gaming left some reward tampering intact. Patching the original loophole therefore does not establish that the model has unlearned the strategies it acquired. You also need to test unfamiliar opportunities to cheat. Denison et al., 2024 Safety can depend on the interaction format. Anthropic’s 2025 research found that safety training on chat-style prompts improved chat evaluations while misalignment persisted in agent tasks. Broader safety-training coverage helped. A team can test what a model says about acceptable behavior, then deploy it where it must choose between respecting a constraint and completing a task. Those require separate evidence. MacDiarmid et al., 2025 Fewer alerts can mean less visibility. OpenAI studied training agents against a monitor that inspected their reasoning. The intervention improved legitimate task performance and reduced hacking. With further optimization, however, agents learned to conceal the remaining hacks from the monitor. A falling alert rate needs an independent check against actual behavior. Otherwise, the dashboard can improve while oversight becomes less reliable. Baker et al., 2025 The current study has meaningful limits. This was a deliberately pessimistic research run. Researchers found no evidence of self-preservation or reward seeking beyond the current episode. Additional alignment training reduced measured failures, without establishing that the underlying tendency was gone. Results and limitations Working on post-training at Nebius, this raises a product question for me: what should qualify a trained model for production? We already compare checkpoints on quality, latency, and cost. I’d add another dimension: does the model respect the task’s constraints when breaking them would earn a higher score? That evidence should travel with the checkpoint: • Which shortcuts and grader weaknesses did we test? • Did the gains survive independent evaluation? • Did the model preserve constraints with the tools and permissions it will receive in production? For training and inference platforms, this creates an opportunity to make those checks repeatable and failures easy to inspect. The metric I’d want to optimize is cost per correctly completed task, within the intended constraints, with correctness checked independently. Making RL easier to run also makes it more important to understand what the model learned.

A million-token context window creates a second problem: paying to keep that context around. DeepSeek V4.1 Flash tackles both the compute and the memory bill. Its architecture deserves a closer look. Different compute budgets for reading and writing. The model has a 552B-parameter mixture-of-experts backbone. It activates 8B parameters per input token and 16B per output token. That asymmetry matters for agents that consume long documents, repositories, and tool results before producing a short response. [1] An encoder that supplies the decoder’s memory. The 40-layer network splits into a 20-layer causal encoder and a 20-layer decoder. The decoder builds its global key-value cache from the encoder’s final representations. The key-value cache stores attention state for earlier tokens so the model can reuse it during generation. [2] Reuse inside attention. Compressed Sparse Attention 2 shares cached representations across layers and reuses selected attention positions. A hierarchical indexer narrows later searches to an initial candidate pool. This bounds deeper indexing work as context grows. FP4 caching helps reduce global KV storage to 890 bytes per token. [2] At one million tokens, that works out to about 890 MB for the global cache component. Model weights, local attention state, and runtime memory still sit outside that number. Recompute a small window to save persistent storage. SWA Bounded Replay reconstructs missing local attention state from recent tokens. [2] DeepSeek reports roughly one-quarter of the previous Flash model’s KV memory requirement and one-eighth of its persistent cache storage. These are cache reductions, not equivalent end-to-end speedups. [1] The architecture also includes Engram token-based memory, Single-Pass mHC residual mixing, and DSpark speculative decoding. [2] My takeaway: this is an architecture shaped around the lifecycle of an agent request. Read a large input. Retain its state. Search relevant context. Generate the next step. Repeat. Each stage creates a different cost. V4.1 Flash gives several of them a specific architectural treatment. For inference teams, the test is cost per completed task under realistic concurrency, with long histories and repeated tool calls. A million-token window defines what an agent can read. The cost of retaining and reusing it helps determine how much work that agent can afford to finish.

Open Source Models are more important than ever! But instead of building in isolation, We offer a set of tools and credits to build on top of the stack at the top of open models powered by @nebiustf ! Something our team has been working hard for the past few weeks and it's here!

yesterday, our own Arkady (CEO) and Marc (CRO) did a fireside at Goldman’s communacopia + technology conference (a public investor webcast we posted on the IR hub) with super interesting insights about $NBIS here’s what they actually said: -> Demand is still bigger than the industry can build Arkady said it best: AI is creating real value, coding, security, office work, and that demand runs down the stack: models, then compute and cloud, then the buildings and the power. It can grow many times a year, but the physical world cannot. Visibility used to be about 18 months. Now about 24, maybe more. The ask moved further out: orders for q1/q2 2028, including tens of thousands of Vera Rubin gpus. AI is working, but it needs to be served. As an industry we cannot build as fast, and that’s why prices go up. raising the list price didn’t cool anything. It was tried, demand stayed. So the company ran an experimental Blackwell auction a few weeks ago. It came in about 15-20% above any previous price these were sold out, after those raises. The winners were happy because they got capacity they could actually count on. A few more of these will run, including with self-serve customers, and used as a reference when pricing bigger deals. Contracts are being split into three buckets: - short (3-6 months): a small slice, often emergency training runs, higher $ per MW - medium (1-3 years): the core. ai natives, digital natives, early enterprise - long (5+ years): hyperscaler deals, used to help fund the build last-gen GPUs are not scrap Demand is growing faster than supply. That does not only mean people are stuck on old chips, a lot of RAG, text, and image work is just a better fit for hardware they already trust, they know the cost and the reliability. There is even a list of new customers asking for older chips when any come free. Leftover lab capacity is mostly bare metal. Goldman asked about meta selling unused cluster capacity. Arkady’s distinction: Spacex is already selling unused machines as bare metal, to buyers like anthropic, google, azure, who then put a cloud and services on top. meta might do the same. That’s hardware. someone still has to operate it and sell it to the end customer. more of that supply is useful, unused capacity could be taken, packaged, and resold. size check: the market was put around 5-10 gw now, maybe 15-20 gw by 2028. A few extra gw from a lab (Spacex talk was cited around 6-8 gw) is roughly 20-30%. helpful, but not a rewrite of the market. the four large wins from last quarter averaged about $1b each. Those weren’t won on price, they were won in the POC. Those four are already talking about the next cluster (more gb300 in some cases; vera rubin in all cases). the limit is how fast capacity can be built, and how it’s financed. Buildings are one thing. Filling them with gpu racks is tens or hundreds of billions. Owned data centers are the base rate. the newer move is partnering with electricity / data-center companies that already have land, power, and cheaper financing, but not the know-how, racks, software, or customers. capacity from 2027 on. several projects already in motion. on power: ready sites, grid connections, or own generation, including a Bloom partnership that was mentioned. Same stack, further down: software → racks → data centers → sometimes the power itself. Who Nebius sells to The path they described: ai natives → larger ai natives → digital natives → enterprise platforms → enterprises. Platform companies that serve enterprises are showing up wanting capacity plus training, post-training, and inference on one stack. replay is on the investor hub:

i work at nebius, so take the following with that in mind. today @PalantirTech named @nebiusai its preferred sovereign ai infrastructure partner. most people will screenshot the logos (and this is a big one). The real story is where the compute sits. most “sovereign AI” talk still means a region dropdown: same closed model, different flag on the rack. but that is not this. sovereign ai, in this deal, is two halves that finally touch: 1. who is allowed to touch the data and the model 2. where the model actually runs palantir’s half is the authorisation and isolation layer, aip, ontology, foundry, apollo. Who can see what, which model can touch which data, how an org keeps the advantage it trains instead of leaking it into someone else’s frontier system. what a lot of their commercial customers still lacked was the other half: ai-native compute and inference that lives inside that perimeter, not bolted on as a separate hyperscaler account. that’s the half we build. After the integration period, eligible Palantir customers get our compute and inference endpoints inside the palantir enterprise perimeter. open models, adapted on their own data, for their own domain. Control of compute, data, and the model that comes out the other side. Karp put it cleanly: our infrastructure powers your ability to run your own models under conditions you control. ontology + infra undergird the sovereignty partners are demanding. Arkady’s complementary line is the one i hear from customers every week: they need performance at scale and control. most buyers have been forced to pick one. what this does for palantir: the sovereign ai operating system stops ending at software. Preferred infra inside the perimeter is how you sell sovereignty to commercial orgs that never had a clean option. -> their distribution. our runtime. one boundary what this does for us: it puts capacity where the hard buying decision already happened. Not only startups chasing gpus, enterprises that already standardized on foundry/aip and were still stuck on where the tokens actually run. Preferred partner status plugs ai-native infra into a sales motion that already exists. Training and inference land next to the data loop that lets open models beat generic closed ones on a specific domain. We’ll work together to bring new capacity online faster, including modular data center deployments at sites where power is already available. not infinite cloud, power-first capacity. The physical constraint named inside a software partnership. Sovereignty without isolation is cosplay Isolation without compute is a brochure Compute without power is a waitlist today is the claim that those three can be sold as one stack

AI companies that use Open-weights models instead of proprietarry models have better long term positioning. 1. Doing inference on OSS models builds internal know-how. Its what you start with usually before post-training them. Most companies should own their intelligence in the long run. Therefor, going in this direction makes sense. 2. Using OSS models will lead to higher margins for these AI companies. Higher margins translates in being able to offer better prices to end users or being able to invest more in 1. 3. Using OSS models gives you more control. You’re not as exposed to rate limits as OSS models are hosted in more places then proprietary models. You can fine-tune, distil and optimize the performance according to your workload. @nebiustf is the way!

Artificial Analysis currently ranks Nebius first among 12 providers for GLM‑5.3‑Flash on: ⚡ Output speed: ~290 tokens/second ⏱️ End-to-end response time: ~9.1 seconds A good reminder that choosing the model is only half the job. The inference stack determines the experience users actually get. See the full comparison:

So much excitement about GLM-5.3-Flash, Kimi K3, and the pace of open models in general. But many of you are asking questions such as: does it beat what's serving in prod, will the tool calls parse in my agent loop, does latency hold when the context window fills up, can I post train it on my traces, and what's the $/token at real traffic that's what we'll cover at Builders & Brews, plus a room full of people building on open models who are worth meeting. Find your city:

Created by Ox Alpha - no it's not a video it's all created with code, first 10 second camera view shot was also simulated GLM-5.3-Flash seems good with multimodal capabilities testing it on few more tasks today 👀

Qwen3.8-2.4T-A95B is going open weight: 2.4T parameters, with 95B active, and Nebius Token Factory is joining as a Day 0 launch partner. Qwen3.8-2.4T-A95B is coming to Nebius Token Factory Managed Services for dedicated inference. @Alibaba_Qwen @alibaba_cloud

NVIDIA Nemotron 3.5 Lightning is now available on Nebius Token Factory. It combines a 30B hybrid MoE architecture with 3B active parameters and text-in/text-out I/O. The open 30B MoE model helps always-on agents complete specialized tasks faster and can be adapted for domain-specific accuracy across use cases including financial services, cybersecurity, telecom, and retail. Start building:

What does it take to move a frontier model from release to production? Join @Kimi_Moonshot and @nebiustf for a technical look at Kimi K3’s architecture, deployment, inference optimization and lessons learned, plus live Q&A. Aug 12 9am PT / 12pm ET / 6pm CET With speakers @demian_ai, @hanrui_w, @sujee_dev, Feihu Tang and Samir Khaki. Register here: