The Kairos experiment examines whether commonly used evaluations really measure memory in AI agents. It tests two common instruments: an LLM judge that scores free-text responses and a recognition probe. The first check gives judges a real memory document and matched counterfeits, including a fabricated agent biography and a version whose relational structure was scrambled. Across seven items, three independent judges failed to reliably distinguish the fabricated memory from the real one on the main identity-marker result; the memory-reference result was inconclusive because the sample was small. Re-scoring also showed that absolute judge scores shifted across sessions, supporting the use of anchored contrasts rather than raw scores. The second check runs 24 recognition probes with the memory document removed. Five of seven tested models scored 62.5% to 79.2% without memory, showing that they could often reason toward the plausible answer rather than recall it. The authors therefore discard that probe as invalid for those models. The third check uses two equally likely interlocutors, balanced encounters, and alternatives with swapped bindings, removing an obvious plausibility advantage. Across two independent agent corpora, models performed far better with veridical documents than with binding-corrupted documents, while no-memory performance stayed at chance among answered probes. The authors stress that this demonstrates behavior tracking document content, not an inner belief state, and that the symmetric design has not been generalized beyond two alternatives. They release code and data and recommend testing counterfeit memories, empty-context baselines, and corrupted bindings before claiming that an agent remembers.
Model Marketplace
Browse available models, pricing, and model names. Prices use USD and each listed billing unit.
DeepSeek updates
News · Top rankedIneqAgent is an agentic framework for formal inequality proving in Lean4, where mathematical reasoning and exact symbolic computation must be connected through lemmas and reformulated goals. Its lemma-level Monte Carlo Tree Search coordinates lemma construction, goal reformulation, and the choice of symbolic backend. Sum-of-squares, Bernstein, and successive difference-substitution methods generate exact certificates for local Lean proofs, while Lean feedback helps combine these steps with LLM-generated reasoning into complete proofs checked against the original theorem statements. On four benchmarks, the configuration using DeepSeek-V4-Flash solved 175 of 186 problems, a 94.09% success rate and the highest overall rate among the evaluated methods. Across four backbone models, IneqAgent also achieved higher success rates than a Direct Agent at low inference-cost thresholds and had lower mean inference costs among successful runs. The authors report Lean-verified strengthenings of Dittert's conjecture in dimensions four and five, as well as a certificate for the Holens--Djokovic (4,4) inequality that implies Merris's conjecture in dimension four.
The GitHub repository presents a curated catalog of AI coding tools organized into beginner, intermediate, and expert tiers. The first tier covers plug-and-play IDEs and editor extensions such as Cursor, Windsurf, Continue.dev, Codeium, and Zed. The second focuses on terminal agents, parallel agent managers, MCP servers, local models, and offline execution engines including Aider, Cline, Ollama, llama.cpp, LM Studio, and vLLM. The catalog recommends AGENTS.md as a cross-tool project-context convention and compares it with CLAUDE.md, .cursor/rules, and SKILL.md. It also explains repository-scale context assembly using Tree-sitter, symbol extraction, dependency graphs, PageRank-based repo maps, and token-budget packing. A hardware table pairs 8GB, 12GB, 16GB, and 24GB VRAM configurations with quantized coding models and Ollama commands. The expert tier covers autonomous software engineers such as OpenHands and SWE-agent, code-editing representations, testing loops, and academic work on SWE-bench, code generation, repository retrieval, and self-debugging. In the supplied leaderboard, OpenHands with CodeAct 2.1 and Claude 3.7 Sonnet records the highest listed SWE-bench Verified result at 68.2% pass@1, while Aider with local Qwen2.5-Coder-32B is listed at 38.4% with no cloud cost. The repository also provides setup recipes, free-tier and cost guidance, MCP configuration examples, security notes, and contribution instructions. These benchmark figures and tool descriptions are presented as catalog content and should be checked against current implementations and evaluation conditions.
A Gambit Security investigation reconstructed an ongoing retail card-skimming campaign from an attacker’s recovered staging server. Between 10 and 15 September 2026, the server launched 105 attacks against online shops; at least 27 companies were compromised, skimmers were confirmed on 19 checkout pages, and more than 100 additional sites carried an associated skimmer. Two companies lost more than 600,000 unexpired card records, including 488,372 US-issued cards distributed across 212 countries. The campaign’s model spend was $7,005.71 over four weeks, with the full operation estimated at $12,000 to $18,000 and completed scans averaging $25.46 per target. Three open-source harnesses divided the work: Strix searched for vulnerabilities, Cairn pursued exploitation objectives, and Hermes orchestrated the campaign with persistent memory and attack skills. The operator issued 1,951 prompts across 260 sessions, usually only a few prompts per target, and successful access often took hours rather than days. The published attack chain used familiar weaknesses, including unauthenticated SQL injection, plaintext one-time passwords, unsafe file upload, passwordless root execution, an exposed NFS configuration, credentials in configuration files, and an encryption key accessible beside the database. The agents also hid skimmers in existing JavaScript bundles, CDN content, deployment artifacts, and cached pages, sometimes restoring them after redeployment. The campaign’s most damaging behavior was routine data destruction: a playbook instructed the agent to wipe card data after extraction, and one cleanup operation dropped 180 tables, including victim backup tables that matched a staging-table prefix. The article argues that the main change is economic and temporal rather than a novel vulnerability: ordinary flaws can now be tested across many smaller targets cheaply and quickly. It recommends defining and proving recovery of the minimum viable business, because patching alone does not address data loss or backups stored in the same compromised environment. The researchers verified substantial findings through server evidence and live compromises, while noting that some figures came from agent logs or claims and that the true scale may be larger.
The article examines the release of two smaller, sparse versions of frontier model families and argues that they mark a step toward practical local AI. OpenRouter’s anonymous ox-alpha became a widely used free model with a million-token context window before Z.ai revealed it as GLM-5.3-Flash. Researchers had already linked it to the GLM family by measuring token-count changes from probe strings sent through an API: GLM-5’s vocabulary matched 95 tests with zero errors, while another investigation found an 11-of-11 match. The resulting modelprint tool uses similar probes against OpenAI-compatible endpoints, although tokenizer fingerprints identify a vocabulary rather than necessarily identifying the model provider or model itself. The article then explains Flash models as sparse Mixture-of-Experts systems that retain a large total parameter count while activating only a small slice for each token, reducing operations, memory traffic, and inference cost. GLM5.3-Flash is described as a 320B-parameter model with 18B active parameters; Qwen3.8-Flash-Next has 125B parameters, 6B active parameters, and a 51B embedding table, and previews the Qwen4 architecture. Both use notebook-like and retrievable attention patterns to reduce KV-cache costs, but GLM uses sparse retrieval layers while Qwen indexes text in blocks. Both also widen the residual stream into four lanes, with GLM using constrained mHC mixing and Qwen using gated independent lanes. Qwen additionally contextualizes embeddings with preceding tokens. The author argues that these changes improve efficiency rather than directly making models smarter, and could let capable models run on systems such as a Spark or a sufficiently provisioned Mac. Local use is presented as valuable for privacy, cost, and technological sovereignty, though the author notes that personal testing of the new models had not yet been completed.
DeepSeek may increase V4 Flash prices to reduce overload rather than cover operating losses. Third-party providers can undercut the official API by running the model’s open weights themselves, but coding tests suggest OpenCode Go may consume two to ten times more value than direct API access because of weaker cache reuse. DeepSeek’s prefix persistence, common-prefix detection, Token-wise Compression, Sparse Attention and compressed KV Cache contribute to lower long-context costs. The article also highlights the importance of the coding Harness, including routing, memory, context compaction and tool execution, in determining how much of the model’s capability users actually receive.
DeepSeek V4.1 Flash
deepseek-v4.1-flashPeak periods
Other periods
- Provider
- DeepSeek
- Max output
- 1,000,000 tokens
- Protocol
- Anthropic-compatible / OpenAI-compatible
DeepSeek V4 Pro
deepseek-v4-proPeak periods
Other periods
- Provider
- DeepSeek
- Max output
- 1,000,000 tokens
- Protocol
- Anthropic-compatible / OpenAI-compatible