As AI solves complex maths puzzles, mathematicians face uncertain future - South China Morning Post
CSU Uses Artificial Intelligence to Improve Hail Forecasting - North Forty News
Artificial intelligence use doubles among rural Nebraskans - Nebraska Public Media
Import AI 479: Spacewar; low-latency tasks versus LLMs; Genie 4
Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe.
LLMs are very bad at low-latency tasks:
…Realtime-Bench says it’ll be a while till AI systems can play games or do online auctions for you…
Researchers have built Realtime-Bench, “a suite of carefully curated dynamic tasks where delayed actions directly affect success”. The test helps evaluate how well modern, general-purpose AI systems can perform tasks where decisions need to be made in real-time.
The benchmark consists of 50 interactive game and web tasks, split across 30 web tasks in four categories (travel, e-commerce, service, and security), and 20 game tasks (arcade, puzzle, strategy, and simulation).
Who did the research: The benchmark was built by people at the University of Texas at Austin, Ohio State University, University of Illinois at Urbana-Champaign, NVIDIA, Northwestern University, and Steadyworks.
What it measures: The test is designed to probe for three qualities: temporal state awareness (”RT-Perceive”), where models can use observations to track changes in the world; adaptive deliberation (”RT-Deliberate”), where models alter their reasoning length and depth according to available time; and time-aware action selection (”RT-Act”) where models need to appropriately time their actions for the situation.
“For example, an agent purchasing a ticket must track the sale status and assess whether observed availability is still current (RT-Perceive), limit further deliberation when the purchase window is closing (RT-Deliberate), and choose between waiting for sales to open and purchasing while tickets remain available (RT-Act).”
They have three different conditions: Wall clock, which is designed to test out performance under real-world latency. Paused clock, which lets you arbitrarily pause things so you can get a sense of baseline capability, and Token clock, which “advances the environment at a fixed rate per generation token. An action is applied at the simulated time reached when it is committed, so consuming more tokens delays execution regardless of actual inference latency”.
How well do models do? The authors tested out models from OpenAI (GPT-6-Astra, GPT-5.6 Sol, GPT-Realtime-2, OpenAI Decisions), Anthropic (Claude Fable 5.1, Claude Opus 5), Google (Gemini 3.8 Flash, Gemini 3.1 Flash Live, Gemini 3.8 Live), Jev 1.13, as well as Chinese models from Deepseek (V4.1) and Moonshot AI (Kimi-K3).
GPT-6 Astra (high) got the highest pass rate on the ‘wall clock’ evaluation setting, with 36%.
By comparison, with a paused clock, models with high reasoning get closer to 70% or even 80% success.
Why this matters: This benchmark is an effective proxy for where AI agents will break - the low scores here point to tasks where AI agents cannot yet be expected to perform well.
Check out the leaderboard: Realtime-Bench (official website).
Read the paper: Realtime-Bench: Can Frontier Models Solve Real-time Interactive Problems? (arXiv).
***
The future of spacewar, courtesy of the Defense Intelligence Agency:
…Anti-satellite tech, space nukes, and more…
Working in AI can be stressful so sometimes to relax myself I like to read about other topics to get a sense for what horrors lie elsewhere. I take the same approach with depression, where if I feel sad I will try and listen to music by people sadder than me in an attempt to cheer myself up (e.g., Radiohead, Neil Young, and in very dark moments Bill Callahan / Smog, etc).
So with that in mind I recently read “Challenges to Security in Space”, by the Defense Intelligence Agency.
What’s going on in space: As with conflict on Earth, space is defined by conflict between the US versus China, Russia, Iran, and North Korea, with China the most technologically sophisticated, and Russia the most seemingly gung ho.
“China and Russia are developing a formidable spectrum of offensive capabilities to degrade or destroy space services of other countries, from kinetic destruction to sophisticated jamming and cyber operations,” the DIA writes. “China has more than tripled its space launches per year in the past decade, with 22 launches in 2016 and 93 launches in 2025… Beijing judges space is central to the United States and its allies’ operations. Beijing also considers space an offensive military domain, which has shaped integral components of PLA military planning and campaigns.”
Spacewar: The general vibe you get from reading this 100+ page report is there is a huge amount of ugly, frightening stuff happening in space, or being done to things in space by things on the ground. Some example quotes:
“By the late 2020s, China may field space-based directed energy systems and higher-powered ground-based directed energy systems capable of threatening satellites in higher orbits”
“China is developing space-based technology, including on-orbit counterspace weapons, capable of holding space systems at risk”
“Russia’s Peresvet, a ground-based laser system, is designed to dazzle enemy satellites’ optical systems to mask the movement of Russia’s strategic missile systems”
“Russia has launched several satellites over the last few years which conducted test activities on orbit for technologies that could enable attacks on other countries’ satellites and result in temporary or permanent damage”
China has a couple of spaceplanes, akin to the US’s hyper-secret X-37B vehicle, and they appear to be getting more sophisticated. In 2023, one of China’s aforementioned vehicles, “released and interacted with an object.”
Satellite-on-satellite fights: It’s obvious throughout the report that we’re in a world where rival nations have satellites that try to go up and mess with other satellites from other people.
“In 2019, Russia launched satellite Cosmos 2542, which released a second satellite, Cosmos 2543. Cosmos 2543 then began following a U.S. national security satellite and approached it multiple times at a distance close enough to create potentially dangerous operating conditions,” the DIA writes. “Russia is developing kinetic orbital systems to permanently disable, or “kill,” satellites. Moscow’s Burevestnik is a developmental, co-orbital air-launched ASAT weapon for killing spacecraft in LEO”.
“In February 2022, Russia launched a satellite called Cosmos, which Moscow claimed at the time was testing the effects of space radiation on electronic systems. This satellite was launched into an unusual and isolated orbit not used by any other spacecraft”.
Space nukes: Russia is experimenting with nukes in space. “A nuclear detonation in space would create an electromagnetic pulse, which would immediately affect satellites within range and could impact civilian infrastructure,” the DIA writes. “A nuclear detonation in space could be tailored to impac
Meet TokenEconomizer: Microsoft’s Plugin That Cuts Coding Agent Costs by Up to 30.9%
A team of researchers from Microsoft Research and UC Berkeley have released TokenEconomizer, an open-source cost controller for coding agents. TokenEconomizer is a plugin that learns from past sessions in a repository to cut waste in later ones. It works with Claude Code, Codex, GitHub Copilot CLI and Pi. It does not touch model weights or agent code. On 277 SWE-Bench tasks, it lowered mean cost by 18.4% to 30.9% across 5 agent-model settings.
TL;DR
- Package: A standard-library Python package and agent plugins, with 3 switchable mechanisms.
- Runs on: Any machine running Claude Code, Codex, Copilot CLI or Pi with Python 3.10+. No GPU and no API key of its own.
- Performance: 18.4% to 30.9% lower mean SWE-Bench cost, and cheaper in all 30 repository-setting cells.
- Best: 30.9% lower cost on Claude Code with Opus 4.8 (33.2% when combined with RTK).
- Worst: Only 6.4% to 9.3% savings on ReposVul audits; Pass@1 drops 1.0 point on Opus 4.8.
Why do coding agents waste tokens?
Every agent turn re-bills the system prompt, every loaded tool schema and the full history. The research team profiled 60 SWE-Bench tasks across 6 repositories using Claude Code with Opus 4.8. They found 3 recurring sources of waste:
- Repeated discovery: once 5 tasks precede the current one, 30% of tool calls repeat an earlier action.
- Unused tools: 23 tool schemas load every turn. 16 are never called, and 3 cover 98% of calls.
- Failure loops: 10 failed runs were 17% of tasks but 32% of total spend.
They also tested 3 fixed open-source strategies: RTK, Serena and Aider’s RepoMap. None was cheaper everywhere. All 3 cut cost on Django and raised it on SymPy.
How does TokenEconomizer work?
TokenEconomizer is an external controller that treats the agent as a black box. It keeps a state per repository, built from completed traces. Before each task, it sets 3 policies through the prompt, the tool settings and the event stream.
1. Micro-skills
An LLM proposer turns trace digests into 1-line facts, such as the working test command. Each fact carries regex detectors that check later traces without a model call. A fact enters the prompt after 3 confirmations on tasks with non-overlapping edited files. Its need rate must also stay at or above 0.4. One contradiction counts as 5 confirmations, so wrong facts retire quickly.
2. Tool pruning
Shell, read, write and edit tools are always kept. Other tools are masked after 3 consecutive unused tasks, or when an LLM judge marks them irrelevant. The full catalog returns every 5 tasks for a re-audit. On Django with Claude Code, the first request on task 2 fell from 43,351 to 16,295 prompt tokens.
3. Failure-loop intervention
The controller learns a turn budget: 1.25 times the 90th percentile of successful run lengths. Cheap rule checks flag repeated failures or dense errors. An LLM judge then decides if the agent is stuck, at 0.7 confidence or higher. A stuck agent gets 1 redirect, then a stop if the loop recurs.
What do the SWE-Bench results show?
| Setting | Cost vs default | Pass@1 default | Pass@1 TokenEconomizer |
|---|---|---|---|
| Claude Code + Sonnet 5 | -23.6% | 74.7 | 76.0 |
| Claude Code + Opus 4.8 | -30.9% | 74.0 | 73.0 |
| Codex + GPT-5.6 Terra | -24.7% | 73.8 | 74.7 |
| Codex + GPT-5.6 Sol | -26.1% | 80.1 | 81.7 |
| Copilot CLI + Sonnet 5 | -18.4% | 75.0 | 77.6 |
Source: arXiv 2610.22732, Table 1. Costs include all controller calls. 3 seeds per setting.
Pass@1 improved in 4 of 5 settings. Lower reasoning effort saved cost but cut Pass@1 by 1.8 to 10.3 points. RTK alone raised cost in 12 of 30 repository-setting cells, and RepoMap in 15.
Savings build over time. Most repositories cost more at first, but all were cheaper by the sixth task. Alone on Claude Code, tool pruning saved 13.2% to 17.7%, micro-skills 11.8% to 14.3%, and the loop guard 5.4% to 7.1%.
Results also held outside SWE-Bench. Costs fell 16.0% to 21.2% on Web-Bench and 16.5% to 19.6% on CredData. ReposVul saw only 6.4% to 9.3%, because shorter runs leave less to remove.
How does it compare with other token-saving tools?
| Feature | TokenEconomizer | RTK | Serena | Aider RepoMap |
|---|---|---|---|---|
| Approach | Learns per-repository policies from past sessions | Compresses shell command output | MCP server with symbol-level code tools | Ranked map of repository symbols in the prompt |
| Learns from earlier tasks | Yes | No | No | No |
| License | Apache 2.0 | Apache 2.0 CLI; BUSL-1.1 engine | GPL-3.0-or-later (MIT for SolidLSP) | Apache 2.0 |
| Agents | Claude Code, Codex, Copilot CLI, Pi | 18 AI coding tools | Any MCP client | Aider |
| Mean SWE-Bench cost change (paper, 5 settings) | -18.4% to -30.9% | -8.9% to +4.5% | Not in main runs (raised pooled cost in profiling) | -6.1% to +6.5% |
| Cells with higher cost (of 30) | 0 | 12 | Not disclosed | 15 |
Cost rows: TokenEconomizer paper. Licenses and claims: each project’s repository or docs.
How do you install TokenEconomizer?
The plugin needs one of the 4 agents and Python 3.10 or newer. On Claude Code, clone the repo, add it as a marketplace, then run claude plugin install tokeneconomizer@tokeneconomizer. Learned data stays local in ~/.tokeneconomizer. Tool p
Nace AI Open-Sources Drex Max v1: A 27B Decision Model That Runs Locally on Apple Silicon
Nace AI has released Drex Max v1 as open weights. Drex Max is its most capable decision model, and it now runs on your own hardware. The release ships bf16 weights, a Q8_0 GGUF, an affine Q4 checkpoint and forks of llama.cpp and Ollama. For developers, Drex Max open weights mean routing, classification and policy checks can stay on a local machine.
TL;DR
- Size: Dense 27B parameters, 262,144-token window (16,384 by default), up to 255 options per question.
- Runs on: NVIDIA GPU with Python (H100 bf16, or L40S 48 GB in 8-bit). Apple Silicon Macs with 24 GB+ (affine Q4) or 48 GB (Q8_0 GGUF). CPU via llama.cpp and Ollama.
- Performance: 67.52 on the public Decision Index 0.4, 6.27 points ahead of Drex v2.0 (self-reported).
- Best: 97.3 skill on InfographicVQA. Tools area reaches 81.9. A top-tier decision model you can now run offline on a Mac.
What is Drex Max?
Drex Max is a decision model from Nace.AI that answers typed questions about a given context. You pass a state (text, JSON or images) plus named questions. The model returns a probability for every option in one forward pass. Nothing is generated.
Questions come in 3 types: choice among named options, yes/no, and an ordinal score. That makes it a fit for agent routing, ticket triage and guardrail checks. The hosted version is presented at routing, classification and policy checks.
How does Drex Max work?
The model card describes a dense 27B backbone with 64 layers. 48 are Gated DeltaNet linear-attention layers that run left to right. The other 16 are full-attention layers, placed every fourth layer, and they are non-causal. Every token can attend to the whole sequence.
Images pass through a native vision tower and are visible to every question in a request. A 255-way readout head then turns the last hidden state into option probabilities. The config lists a qwen3_5 model type with a 262,144-position window.
&&&How does Drex Max perform on benchmarks?
All figures below are self-reported by Nace AI,
- Decision Index 0.4 (public text): 67.52. Drex v2.0 scores 61.25 and Jev 1.13.0 scores 57.96.
- Area skill: Tools 81.9, Language 72.2, Retrieval 68.8, Knowledge and Reasoning 57.1, Arts 53.4.
- Public vision index: 73.92 across 11 benchmarks and 12,210 rows.
- Artificial Analysis Decision Index: 64, which Nace says ranks 4th of 18.
- JevBench: 92.2% (213/231). Decision Bench v4: 94.30% (1010/1071).
The results dataset covers 140,620 requests. Unanswered requests count as wrong.
How do you run Drex Max locally?
Nace documents 5 runtimes. All of them serve the same POST /v1/systemone API.
| Runtime | Hardware | Checkpoint | Images |
| Python | NVIDIA GPU | bf16 (~49 GB) or 8-bit (~26 GB) | Yes |
| llama.cpp fork | Apple Metal, CPU | Q8_0 GGUF (27.25 GB) | No |
| Ollama fork | Apple Metal, CPU | Q8_0 GGUF | No |
| Kapitsa v0.2.25+ | Apple Silicon | Q8_0 GGUF or affine Q4 | No |
| MLX | Apple Silicon | affine Q4 (14.45 GB) | No |
The affine Q4 checkpoint targets Macs with 24 GB or more. The Q8_0 GGUF needs 48 GB. On a 48 GB Mac, Nace recommends –context 2048 for Kapitsa.
Ollama users set num_gpu 40 to offload 40 layers to Metal. Setting num_gpu 0 runs it on the CPU. The Ollama fork needs a CAPABILITY decision line in the Modelfile.
Quantization holds up in Nace’s tests. On a 48 GB Apple M5 Pro, the Q8_0 GGUF matched H100 bf16 within 0.02 per probability. The affine Q4 stayed within 0.0329 of Q8_0, with no choice flips.
Drex Max vs other decision models
| Drex Max v1 | Drex v2.0 | Drex v1.5 | Jev 1.13.0 | |
| Parameters | 27B dense | 35B MoE, ~3B active | 9B | Not published |
| Max context | 262,144 | 131,072 | 131,072 | Not disclosed |
| Inputs | Text, JSON, images | Text (vision tower dropped) | Text | Not disclosed |
| Weights | Open, Nace.AI Open RAIL-M | Open, Nace.AI Open RAIL-M | Open, Nace.AI Open RAIL-M | Closed, hosted |
| Smallest local file | Affine Q4, 24 GB+ Mac | Q8_0, 36.9 GB | Q8_0, ~9.5 GB | Not applicable |
| Public Decision Index | 67.52 (v0.4) |
NVIDIA Releases Auto Ontology: An Open Semantic Layer That Grounds AI Agent SQL
NVIDIA has released Auto Ontology, an open library under the Apache 2.0 license. NVIDIA Auto Ontology is a semantic layer framework that builds a context layer from enterprise metadata. Agents use that layer to write SQL against certified business definitions. The repository states the problem plainly: a fast revenue answer is worthless if the number is wrong.
TL;DR
- Library: Default LLM is
nvidia/nemotron-3-nano-30b-a3b, with Nemotron embedding and reranking models. - Runs on: Docker Compose for a local stack, or Kubernetes via Helm. Inference uses hosted NVIDIA NIM endpoints or self-hosted NIMs.
- Performance: Strong engineering, thin public evidence.
- Best: Ontology creation time cut by 90% in NVIDIA testing (self-reported).
- Worst: 0 published text-to-SQL accuracy scores. The eval repo ships a harness, not results.
- Bottom line: An open, MCP-ready context layer that any agent harness can query. But you must run the benchmarks yourself to trust it.
What is NVIDIA Auto Ontology?
It is a context layer that sits between an agent and the data warehouse. It tells the agent where data lives and what fields mean. It also says whether to use a field, how to use it, and where it came from.
Here are the five jobs:
- Discover: generate ontologies from metadata and connect existing semantic models.
- Ground: give agents trusted meaning for metrics, entities and relationships.
- Validate: check generated queries against certified definitions and lineage.
- Expose: serve context through MCP, CLI and API.
- Learn: fold certified corrections back into the ontology.
NVIDIA’s developer page frames the output as a starting ontology. Business experts still test, refine and certify it.
How does Auto Ontology work?
The stack has 4 core services. A Next.js 16 frontend handles the web UI and authentication on port 3000. A FastAPI backend with LangChain runs chat, NL-to-SQL and catalog APIs on port 3001. An ingestion worker writes embeddings on port 3002. Postgres with pgvector stores the catalog, the ontology and the vectors.
Source databases can be Databricks, Postgres, Snowflake or DuckDB. Models are set through environment “triplets” of endpoint, key and model. Separate roles exist for reasoning, non-reasoning, embedding and reranking. Any OpenAI-compatible endpoint works.
&&What does the MCP server expose?
The MCP server docs describe 9 read-only tools. ask_question returns the answer, the SQL and the rows as separate fields. Results cap at 100 rows. check_readiness confirms a compiled semantic layer and a live database connection exist.
There is no tool for browsing raw schemas. NVIDIA chose that on purpose, so agents reach data through the semantic layer. The underlying OpenAPI spec publishes 87 operations. Sign-in runs through OAuth, and every call carries the signed-in user’s permissions.
How is it benchmarked?
NVIDIA ships a separate harness, auto-ontology-eval. It runs 2 arms in NeMo Gym: a schema-only LLM control and the ontology-grounded agent. Scoring is deterministic execution match, with no LLM judge.
Bundled datasets include BIRD Dev (1,534 questions, 11 databases) and Mini-Dev (500 questions). FDABench-Lite adds 169 questions, and BEAVER defaults to a 100-question sample.
The harness repo adds 2 limitations. Optional few-shot exemplars and value anchors go to the ontology arm only. A delta measured with them covers those inputs, not grounding alone. Its tested configuration also routes both arms through a Claude model on NVIDIA’s inference gateway.
How does it compare with other semantic layers?
| Feature | NVIDIA Auto Ontology | Wren AI | Cube Core | dbt MetricFlow |
|---|---|---|---|---|
| License | Apache 2.0 | Apache 2.0 core (open core) | Apache 2.0 backend, MIT client | Apache 2.0 (v0.209.0+) |
| How the layer is built | Generated from metadata, then certified | Agent-assisted context files in Git | Defined in code | Defined in code in a dbt project |
| Agent interface | MCP, CLI, REST API | MCP, CLI, SDKs | SQL, REST, GraphQL APIs | Python library and CLI |
| Data sources | Databricks, Postgres, Snowflake, DuckDB | 22+ | All SQL data sources | Via dbt adapters |
| Default models | Nemotron via NVIDIA NIM | Not disclosed | Not applicable | Not applicable |
Deployment and availability
Local setup is docker compose up -d --build after filling in .env. Kubernetes users get a Helm chart and a DEPLOYMENT.md guide. NVIDIA team also publishes the backend as an NGC container. An NVIDIA API key from build.nvidia.com covers the default models.
Key Takeaways
- Auto Ontology generates a starting semantic layer from enterprise metadata.
- Agents query it through 9 read-only MCP tools, plus CLI and API.
- NVIDIA reports 90% faster ontology creation in its own testing.
- No text-to-SQL accuracy scores are published yet.
- Defaults to Nemotron on NIM, but accepts any OpenAI-compatible endpoint.
Deep Principle Releases Hydrogen-1.0: A 753B Open-Weight Model for Scientific Discovery
Deep Principle has released Hydrogen-1.0, an open-weight scientific model for chemistry, materials science and life sciences. It is built by post-training Z.ai’s GLM-5.3 on domain data, laboratory records and agent discovery trajectories. For developers, it offers open weights that report a FrontierScience-Research win over GPT-6-Astra. The weights ship under the MIT License.
TL;DR
- Size: 753B parameters (MoE, ~40B active per the GLM-5.3 base), 1M-token context, 128K max output.
- Runs on: 8 GPUs with SGLang (tested); NVFP4 for NVIDIA B200/B300, MXFP4 for AMD MI350X/MI355X.
- Performance: Beats GPT-6-Astra on FrontierScience-Research; roughly matches Kimi-K3 on GPQA-Diamond.
- Best: 38.3% on FrontierScience-Research vs 33.3% for GPT-6-Astra (self-reported).
- Worst: Trails GPT-6-Astra (max thinking) by 3.33 points on NMR structure elucidation.
What is Hydrogen-1.0?
Hydrogen-1.0 is a scientific research assistant model from Deep Principle, a Hangzhou-based AI-for-materials company. The model card describes it as a GLM-5.3 derivative for scientific problem solving. It targets question framing, experiment planning, results analysis and open-ended discovery. Deep Principle team states that proprietary models produced recent AI science results, but their discovery capabilities are not open. Hydrogen-1.0 is its attempt at a reproducible, inspectable alternative.
How was Hydrogen-1.0 trained?
The pipeline combines supervised fine-tuning (SFT) with reinforcement learning (RL). Data comes from 4 sources:
- Domain knowledge: chemistry, materials science and life sciences data.
- Research scenarios: task-specific feedback from Deep Principle scientists.
- Laboratory records: full-cycle electronic lab notebook records captured through Mira.
- Discovery trajectories: sequences of choices and feedback, including failed attempts.
Infrastructure builds on Megatron and SGLang across NVIDIA CUDA and AMD ROCm. The team keeps dynamic sparse attention causal across context-parallel shards. Muon optimizer state is moved between CPU, GPU and disk per backend. RL uses resumable rollouts, and each token keeps its original behavior log-probability for importance correction. A layout adapter converts training checkpoints into BF16 and 128×128 blockwise FP8 inference weights. Full datasets and recipes are not released.
How does Hydrogen-1.0 perform on benchmarks?
On public benchmarks, Deep Principle reports Hydrogen-1.0 outperforms GPT-6-Astra on FrontierScience-Research. The tweet thread from one of the founders’ report 93.0% on GPQA-Diamond. That improves on GLM-5.3 and roughly matches Kimi-K3, per the blog.
Internal research-task evals compare it with GPT-6-Astra at maximum thinking:
- Scientific computing: Hydrogen-1.0 leads by 5.13 percentage points.
- Forward reaction prediction: Hydrogen-1.0 leads by 4.76 percentage points.
- NMR structure elucidation: GPT-6-Astra leads by 3.33 points.
In transition metal complex design, models chased a larger HOMO-LUMO gap over 5 rounds. By round 5, Hydrogen-1.0 reached a mean best-so-far gap of 4.318 eV. GPT-6-Astra reached 3.689 eV and Kimi-K3 reached 3.353 eV. On 117 experiment-planning questions, Hydrogen-1.0 scored above Kimi-K3. GPT-6-Astra and Opus-5.5 were excluded after refusing more than 50% of requests.
&What can it do in a real lab workflow?
The technical post shows 4 demos. Hydrogen-1.0 writes Python to build a polyethylene melt and runs GROMACS. It detects a topology error, repairs it, and analyzes the corrected 40 ns trajectory. It grows an Ireland-Claisen reaction network of 27 structures through 30 plausible reactions. It ranks 7 candidate monoformylation products of 2-bromonaphthalene from a 1H NMR trace. It also writes a JSON protocol for a 12-vial Suzuki coupling study. A virtual lab compilation step catches an aspiration without a fresh tip before execution.
How do you deploy Hydrogen-1.0?
Weights are on Hugging Face and ModelScope. Deep Principle team tested BF16 serving with SGLang on 8 GPUs using tensor parallelism of 8. NVFP4 targets NVIDIA B200 and B300. MXFP4 targets AMD Instinct MI350X and MI355X. The server exposes an OpenAI-compatible API with a reasoning_effort setting.
Hydrogen-1.0 vs GLM-5.3, Kimi K3 and GPT-6-Astra
| Feature | Hydrogen-1.0 | GLM-5.3 | Kimi K3 | GPT-6-Astra |
|---|---|---|---|---|
| Developer | Deep Principle | Z.ai | Moonshot AI | OpenAI |
| Total params | 753B | 753B (HF) | 2.8T | Not disclosed |
| Active params | ~40B (inherited) | 40B (LM Studio) | 16 of 896 experts | Not disclosed |
| Context | 1M | 1M | 1M | 1.05M |
| Weights / license | Open, MIT | Open, GLM-5.3 License | Open | Proprietary |
| Modality | Text | Text | Text + vision | Not disclosed here |
| Science focus | Chemistry, materials, life sciences | Coding, agents | General | General |
| FrontierScience-Research | 38.3% | Not disclosed | Not disclosed | 33.3% (per Deep Principle) |
| GPQA-Diamond | 93.0% | Not disclosed | Not disclosed | 96.0% (OpenAI-reported) |
| API price (in/out per 1M)< |
Zyphra AI Introduces Zyphra Hierarchical Attention (ZHA): A Path to True Subquadratic Attention
Zyphra has released Zyphra Hierarchical Attention (ZHA), a sparse attention method built around a tree-structured indexer. ZHA targets a known gap in sparse attention: the indexer that picks tokens still scans the whole context. Zyphra’s technical paper shows ZHA can be distilled into frozen pretrained models with about 4B tokens. On Qwen3.6-35B-A3B it stays close to dense attention on reasoning, agentic coding and 256K retrieval benchmarks.
TL;DR
- Size: It is a method that adds 81,920 trainable parameters to Qwen3.6-35B-A3B (about 0.00024%). Tested up to 256K context.
- Runs on: Retrofits frozen pretrained models after a 4B-token KL warmup at 32K context. Speedups come from a modeled H100 BF16 roofline. No optimized kernels released. An FP4 router variant was also tested.
- Performance: Within a few points of dense attention on 7 reasoning, coding and retrieval benchmarks.
- Best: LongBench v2 hits 63.4% (deep tree) vs 58.0% dense, using about 1.1% of keys.
- Worst: On NoPE-based Nemotron-3.5-Lightning-30B-A3B, deep-tree RULER-13 drops 25.3 points below dense.
What is Zyphra Hierarchical Attention (ZHA)?
ZHA is a sparse attention method from Zyphra that finds relevant keys by searching a tree. It keeps the full KV cache. The tree only decides which cached keys each query reads.
Fixed-state methods like SSMs and sliding windows saturate at long contexts. Fine-grained sparse methods like DeepSeek Sparse Attention score every past token with an indexer. The final attention is sparse, but the indexer still grows with the full context.
Near 1M tokens, cited work shows indexing takes 60 to 90% of layer time. ZHA makes the search itself sparse.
How does ZHA’s tree indexer work?
Building the tree
ZHA splits keys into contiguous chunks and compresses each chunk into a summary key. Summaries are grouped and compressed again, level by level. With chunk size 4, 6 levels cover 4,096 tokens per top node.
Summaries come from a small set of learned reference queries. Each reference query gives a tangent-plane estimate of a chunk’s log attention mass. The max over these planes approximates how much attention any future query would give the chunk.
How does top-p routing pick tokens?
At decode time, a query scores the top-level nodes. It keeps the smallest set holding a fraction p of estimated attention mass. It then expands only those children, repeating until it reaches raw KV entries.
Top-p makes the budget adaptive. A single-fact lookup touches few branches. A query that needs many scattered facts expands more. The first token is always kept as an attention sink.
How does ZHA perform against dense attention and MSA?
Zyphra team converted Qwen3.6-35B-A3B with 2 trees: deep (chunk 4, 6 levels) and wide (chunk 256, 1 level). Both use 16 reference queries and p = 0.95. It also trained a MiniMax Sparse Attention (MSA) indexer on 40B tokens, 10x ZHA’s budget.
| Benchmark (Qwen3.6-35B-A3B) | Dense | ZHA deep | ZHA wide | MSA |
|---|---|---|---|---|
| AIME 2026 | 93.1 | 91.2 | 90.8 | 91.7 |
| SWE-bench Verified | 67.4 | 67.4 | 68.4 | 63.6 |
| Terminal-Bench 2.1 | 49.9 | 46.7 | 53.3 | 46.1 |
| MMLU-Pro | 84.9 | 85.3 | 85.2 | 85.3 |
| GPQA Diamond | 84.8 | 85.0 | 85.0 | 83.1 |
| RULER-13 (256K) | 96.0 | 93.3 | 96.2 | 77.1 |
| LongBench v2 (thinking) | 58.0 | 63.4 | 62.4 | 49.3 |
Source: arXiv 2610.42112, Figure 5. All numbers self-reported by Zyphra.
The biggest gap is on retrieval. MSA drops to 77.1% on RULER-13, while both ZHA trees stay above 93%. On LongBench v2, the deep tree and MSA read similar key shares (1.1% vs 1%). The deep tree scores 14.1 points higher.
Across 28 tree geometries, RULER accuracy held steady when attending to 10 to 30% of keys. At fixed compression, deeper trees kept accuracy at higher sparsity. A 4-parameter model predicts RULER accuracy from geometry with 1.44-point RMSE.