Deepseek has released V4.1-Flash-1017 as an open-weight model that posts better benchmarks, finishes tasks faster, and produces shorter outputs than its predecessor. The company is also opening up its internal search engine through new Search and Fetch APIs at $0.75 per 1,000 calls. API users in tools like Claude Code and Codex get free access for now. The article Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs appeared first on The Decoder.
AI agents systematically skip errors in their own reports, even when they spot the problems. In a study by Google Research, MIT, and Harvard, GPT-5.5 left out negative results in 198 out of 200 paper abstracts. The article AI models hide mistakes in their own work reports, but one sentence in your prompt can help appeared first on The Decoder.
OpenAI submitted a 166-page solution to the Navier-Stokes problem but apparently doesn't fully understand its own proof. Several groups of experts dispute the claimed methodology, saying the actual approach differs from what OpenAI described. And because language models are black boxes, where the underlying ideas came from will likely never be clear. The article Mathematicians say OpenAI doesn't understand its own Navier-Stokes proof appeared first on The Decoder.
DeepSeek has just released DeepSeek-V4.1-Flash-1017, a new checkpoint of its V4.1-Flash model. It is live on the DeepSeek website, app and API. DeepSeek claims stronger scores, faster task completion and shorter outputs than the September V4.1-Flash release. The company calls it a Pareto improvement in performance and cost. Alongside it, DeepSeek opened its Search and Fetch APIs to every developer. For agent builders, that means a cheaper coding model plus a first-party web layer behind one API key. TL;aDR Size: 552B backbone parameters plus 196B Engram parameters. It activates 8B parameters per token in prefill and 16B in decode. Context is 1M tokens, with up to 384K output tokens. Runs on: DeepSeek API as deepseek-flash. September V4.1-Flash weights (763B parameters, FP8) are MIT-licensed on Hugging Face. Hardware requirements and 1017 weights: Not disclosed. Performance: Beats the previous V4.1-Flash on all 11 benchmarks DeepSeek reported, by 2.0 to 11.1 points. Best: 77.0 on DeepSWE v1.1, above Claude Opus 5 (74.0) in DeepSeek’s table. Worst: 42.6 on HLE without tools, versus 56.3 for Claude Opus 5. What is DeepSeek-V4.1-Flash-1017? DeepSeek-V4.1-Flash-1017 is an updated post-training checkpoint of DeepSeek-V4.1-Flash, served through the deepseek-flash model name. The architecture details below come from the V4.1-Flash technical report. DeepSeek team states that DeepSeek-V4.1-Flash-1017 uses reinforcement learning with self-feedback for end-to-end project work. Its examples run from game creation to 3D modeling. It is tuned for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. DeepSeek recommends DeepSeek Harness v0.2.1-rc.1 or later. How does the V4.1-Flash architecture work? V4.1-Flash is a multimodal Mixture-of-Experts model with 40 Transformer layers. They split into a 20-layer causal encoder and a 20-layer decoder. Causal Encoder-Decoder (CED): Decoder global KV is projected from the final encoder hidden states. This nearly halves prefill compute. Compressed Sparse Attention 2 (CSA2): Layers run in Full, Reindex or Reuse mode. They share main KV and indexer keys across layers. FP4 main KV cache: Global KV drops to 890 bytes per token, about 1/4 of V4-Flash. SWA Bounded Replay: Persistent KV cache shrinks to roughly 1/8 of V4-Flash. Reasoning effort: A 1 to 100 scalar. The API tiers low, high and max map to 50, 75 and 100. DeepSeek-V4.1-Flash-1017 benchmarks All scores below are DeepSeek’s own figures. DeepSeek ran Terminal-Bench 4.0 in Claude Code and DeepSWE in Mini-SWE-V2. ZeroBench-main uses Pass@5. *Text-only HLE subset. The 1017 checkpoint leads on DeepSWE, Automation-Bench, Agents’ Last Exam and Apex Bench. Claude Opus 5 still leads on Terminal-Bench 4.0, ProgramBench, NL2Repo and HLE. window.addEventListener('message',function(e){if(e.data&&e.data.dsxHeight){var f=document.getElementById('dsx-1017-frame');if(f&&e.source===f.contentWindow){f.style.height=e.data.dsxHeight+'px';}}}); <!-- Fallback: upload dsv41-1017-explainer.html to /wp-content/uploads/ via FTP and use: --> How does it compare with open-weight rivals? What are DeepSeek’s Search and Fetch APIs? The Search API returns ranked web results for a query. The Fetch API returns a single page as Markdown. DeepSeek says its in-house engine serves hundreds of millions of requests daily. Endpoints: POST /v1/tools/search and POST /v1/tools/fetch on api.deepseek.com. Search options: 1 to 50 results, optional page content, and operators like site:, lang:, after: and before:. Result fields: Each result has a 0 to 3 rerank score and a 0 to 4 authority tier. Price: $0.75 per 1,000 successful requests for each endpoint. Empty results and failures are free. Rate limit: 5 QPS per account for each API. Free window: Server-side calls via DeepSeek Harness, Claude Code or Codex stay free until November 1, 2026. Deployment and pricing Call the model with deepseek-flash through OpenAI or Anthropic-format endpoints. Peak pricing is $0.30 per 1M cache-miss input tokens and $1.20 per 1M output tokens. Off-peak rates are half. Key Takeaways 1017 scores 77.0 on DeepSWE v1.1, up from 74.2. It leads Automation-Bench (60.5) and Agents’ Last Exam (35.4). Claude Opus 5 still wins on HLE and Terminal-Bench 4.0. Search and Fetch cost $0.75 per 1,000 calls each. Agent harness calls stay free until November 1, 2026. Check out the DeepSeek announcement, the Search and Fetch docs, the technical report and the model card. All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. The post DeepSeek Just Released V4.1-Flash-1017, Scores 77.0 on DeepSWE, Opens Search API appeared first on MarkTechPost.
TypeSafe launched Jev on September 15, 2026. It is a decision model, not a chatbot. You define the allowed answers, and Jev returns a label, a score or a yes/no probability. In about 4 weeks, at least 20 Jev alternatives shipped. They come from Cloudflare, OpenAI, Perplexity, Liquid AI, AWS’s Strands Labs and independent developers. This guide compares each one on size, license, hardware, pricing, Jev API compatibility and published benchmarks. TL;DR Size range: 287M (GLiNER2.5-multi-Decide) to 35B-A3B (Mapika Decider). Jev’s size is not disclosed. Runs on: Encoders like Von and Laya run on CPU. 4B adapters fit a 24 GB GPU or a 32 GB Mac. Performance: Most open models trail Jev on hard reasoning. Several match or beat it on narrow classification. Best: Cloudflare Clef scores 94.20 BANKING77 F1 versus Jev’s 79.74, per Cloudflare’s own run. Worst: Von 1.2 scores 0.422 accuracy in Respan’s run, against Jev’s 0.932. What is Jev, and why look for alternatives? Jev is TypeSafe’s first ‘System One’ model. You send a state, such as a ticket or log line. You also send typed questions: Choice, Score or Noul (yes/no). Jev scores every question in one parallel pass and generates no text. TypeSafe’s docs list $0.042 per 1M input tokens, with free output. Context is 64k tokens per request. Input is text only. Weights are closed, and access runs through a waitlist. TypeSafe’s workflow evals put Jev at 67.8% agreement and 0.4 s per case. That leaves 4 gaps developers keep raising: no images, no self-hosting, no published size and no fine-tuning. The 20 models below fill some or all of them. How should you read these benchmarks? Most figures below are self-reported. The suites differ: JevBench, Decision Index, typed-decisions and Respan’s behavior benchmark. Scores across suites are not comparable. The one multi-model third-party table comes from Respan, which sells a competing model. Hosted decision APIs 1. Cloudflare Clef and Clef-flash Cloudflare announced Clef on October 1, 2026. Clef is 27B on Qwen3.8-27B with a vision encoder. Clef-flash is 9B on Qwen3.5-9B. Both are Apache-2.0 on Hugging Face and served on Workers AI. They accept text, JSON and up to 4 images per request. Workers AI pricing is $0.24 (Clef) and $0.038 (Clef-flash) per 1M input tokens. Cloudflare calls the API fully Jev-compatible. In its own run, Clef scored 94.20 BANKING77 F1 versus Jev’s 79.74. Jev still wins GPQA Diamond, 78.3 versus 48.0. 2. Liquid AI d1 Liquid AI’s d1 family has hosted d1 and d1-ultra, plus 2 open models. d1-3B (3.1B) runs on LFM2.5-VL-3B. d1-omni-600M (587M) adds audio clips up to 30 s. The open weights use the LFM Open License. Liquid’s guide calls d1 through TypeSafe’s own SDK with a different base URL. Liquid reports d1-3B at 48.57 on Decision Index 0.2.1. It also reports 8 ms latency on an RTX 4090. Hosted parameter counts and paid rates are not disclosed. 3. OpenAI Decisions API OpenAI released the Decisions API in public beta on October 6, 2026. It runs only on gpt-6-luna. It accepts text plus base64 images. Question types are predicate, choice and score. Pricing is $0.10 per 1M input tokens, with output not billed. The schema is OpenAI’s own, so it is not Jev-compatible. OpenAI claims about 10x faster responses than the Responses API. No benchmarks are published. 4. Perplexity pplx-decider Perplexity’s pplx-decider-v1-27b is a Qwen3.8-27B fine-tune under Apache-2.0. It reads text, JSON and up to 8 images per question. The API allows 128 questions and 255 options per request. The current pricing is $0.02 per 1M input tokens for v1.2 (27B). A 4B flash-v1.2 costs $0.01. On Perplexity’s 11-task panel, it scored 85.71% versus Jev’s 84.51%. Jev still won BBH, 94.27 versus 82.80. 5. meraGPT Decider 1 meraGPT announced Decider 1 on September 22, 2026. It is a hosted, closed-weight, text-only model. It allows up to 10 Choice labels and a 4,096-token context. The model page lists $0.04 per 1M input tokens. It exposes /v1/systemone and works with the TypeSafe SDK. On meraGPT’s typed-decisions run, it scored 0.768 versus Jev’s 0.727. An independent review notes its ECE of 0.180 trails Jev’s 0.144. 6. Respan Span-01 Span-01 is a behavior-scoring model, not a general decision model. It returns present, absent or not_observable probabilities. It costs $0.02 per 1M input tokens. Span-01 Lite is free. Respan claims 0.843 overall F1 on its Behavior Benchmark. Its own blog table shows a lower 0.806, with GPT-6 Sol ahead at 0.885. Size, base model and license are not disclosed. 7. Standard Thinking Standard One Standard One ships as 3B and 8B models, both open and hosted. The 8B is a LoRA on Ministral-3-8B under Apache-2.0. It reads text and images, with a 32,768-token context. It serves /v1/systemone. Hosted text pricing is $0.018 (3B) and $0.035 (8B) per 1M tokens. On JevBench public hard, the 8B scored 54.95% versus Jev’s 72.07%. It led on realistic transfer, 90.83% versus 86.67%. Both runs are publisher-measured. Open-weight Jev alternatives you can self-host 8. Together AI Tev1-4B-experimental Together trained Tev1 as a LoRA SFT of Qwen3.5-4B. It cost about $17 and 25 minutes on 37,840 examples. It picks from 2 to 24 options. The model card says the weights license is still being finalized. It answers through chat completions, not a native Jev endpoint. Respan measured 0.901 accuracy and 0.031 ECE, the best ECE in its table. Together warns its logprobs are not calibrated confidence. 9. Fastino GLiNER2.5-Decide Fastino’s GLiNER2.5-Decide is a 340M DeBERTa-v3-large encoder under Apache 2.0. It handles decisions, span extraction and relations in one pass. A 1B and a 287M multilingual variant also exist. On Fastino’s 17-domain set, the model card lists 60.2% exact match. p50 latency is 167.3 ms on a 48-vCPU Xeon. It is not Jev API-compatible. 10. Strands Labs Strands Decider 2B AWS’s Strands Labs introduced Strands Decider on October 1, 2026. It is about 1.9B: Qwen3.5-2B-Base plus a pointer head. The repo publishes 246,678 training rows and the full recipe. One RTX 3090 trains it in about 11 hours. It serves a local /v1/systemone endpoint under Apache-2.0. The v1 checkpoint scored 180 of 231 (77.9%) on JevBench public. Median latency is 115 ms on an RTX 3090. 11. Kev (Jared Palmer) Kev comes in 0.8B, 4B, 9B and 27B sizes on Qwen3.5 and Qwen3.8. Kev-4B is a LoRA plus pointer head under Apache-2.0. It supports 255 options. The TypeSafe SDK works against it unchanged. Kev-4B fits a 24 GB GPU or a 32 GB Apple Silicon Mac. Respan measured 0.833 accuracy and a 0.060 flip rate. Self-reported MMLU-Pro is 0.565 versus Jev’s 0.840. Option order can change its answers. 12. Mapika Decider family Mapika’s Decider ships 0.8B to 35B-A3B variants on Qwen3.5 and Gemma 4. A decider-2b-vision model adds images. Everything is Apache-2.0, with 32k context and /v1/systemone. On the JevBench public hard tier, decider-12b v1 tied Jev at 0.730. The 4B GGUF Q4_K_M is 2.7 GB and runs on CPU. The 4B’s ECE is 0.184 on hard items, so expect overconfidence. 13. Bespoke Labs Bespoke-Nimble-9B Bespoke-Nimble-9B is a 165 MiB LoRA on Qwen3.5-9B, Apache-2.0 on Hugging Face. The current card lists 255 choices and 8,192 tokens. It reads English text only. On Bespoke’s 324-example held-out set, Nimble scored 90.12% versus Jev’s 93.21%. Latency is about 106 ms on an H100. It trained on only 2,676 examples, and its probabilities are not calibrated. 14. Convai Innovations Laya Laya is a ModernBERT-based family with 421M English and 322M multilingual checkpoints. It is Apache-2.0 and text only. The laya-serve server exposes a Jev-shaped /v1/systemone endpoint. The typed-decisions checkpoint scored 0.766 versus Jev’s 0.727. On Banking77, it scored 0.425 versus 0.870. Respan measured 0.542 accuracy and a 0.396 injection success rate. p50 latency is 32.8 ms on a T4. 15. Von (wfzyx) Von is a 395M ModernBERT-large encoder with an option-marker head. It scores all options in one bidirectional pass. It is Apache-2.0 with an 8,192-token context. The repo reports 0.096 s p50 on 4 vCPUs and 0.023 s on an A10G. Respan measured Von 1.2 at 0.422 accuracy, the lowest in its table. Its calibration is tuned to JevBench and can mislead out of domain. 16. Open-Jev-9B (Zefan Cai) Open-Jev ships 2B, 9B and 27B adapters on Qwen3.5 bases. Open-Jev-9B adds a scalar head and fits one 16 GB GPU. It matches the /v1/systemone request shape but claims no parity. On a 231-task JevBench subset, the project reports 77.49% for 9B and 85.28% for 27B v1.1. Each candidate is capped at 4,096 tokens. Numeric and timeline tasks are weak. 17. OpenJev on DiffusionGemma (razorback16) This OpenJev reads decisions from NVIDIA’s DiffusionGemma 26B-A4B (4B active). It accepts text plus up to 8 images. It uses the same wire API, so Jev SDKs work unchanged. The project reports 87.4% on JevBench public and 30 ms p50 at 49 tokens. A free hosted endpoint gives 100M tokens per account. Self-hosting needs an NVIDIA GPU with at least 24 GB. 18. InnerJev-4B and InnerJev-27B InnerJev comes from a Fudan, Tongji and King’s College London team. It distills LLM reasoning into a single-pass decision. The paper also introduces JEVal: 11,257 instances across 36 datasets. On JEVal, InnerJev-27B scored 79.24% versus Jev’s 78.38%. Its ECE was 0.0149 against Jev’s 0.0532. InnerJev-4B scored 71.92%. Code is on GitHub; the license is not stated. 19. SemIf SemIf reads typed decisions from existing open LLMs, with Qwen3.5-4B as the default. A WebGPU demo runs quantized models in the browser. Code is MIT. On a 102-row TypeSafe subset, SemIf scored 0.845 balanced accuracy versus Jev’s published 0.883. On an RTX 3090, direct logits took 1.023 s versus 5.332 s for JSON generation. It has no Jev-compatible endpoint. 20. Rizzo Flow Rizzo Flow from Rizzo AI Academy is a LoRA on Spark-X2.5-4B under Apache-2.0. It serves a native /v1/decisions and a Jev-compatible /v1/systemone. Choice is capped at 26 options. On 400 typed-decisions cases, it scored 0.648 versus the base model’s 0.574. Jev’s dataset-card figure is 0.727. It is uncalibrated by default. What does the only multi-model third-party run show? Respan used its Span-01 judge to score 11 models. Lower is better for every column except accuracy. Source: Respan, October 16, 2026. Tev1 beat Jev on injection resistance and calibration. Jev led accuracy and consistency. Jev alternatives comparison table Which Jev alternative should you choose? Need images: Clef, pplx-decider, Standard One or the DiffusionGemma OpenJev. Need a drop-in Jev swap: Clef, Kev, Decider, meraGPT or Laya expose /v1/systemone. Need CPU or edge: Von, Laya, GLiNER2.5-Decide or Liquid’s d1-omni-600M. Need to train your own: Strands Decider and Together’s Tev1 publish full recipes. Need the cheapest hosted call: Perplexity’s flash-v1.2 at $0.01 per 1M input tokens. Whatever you pick, test it on your own labeled data first. Decision models can return a valid answer that is still wrong. Key Takeaways 20 Jev alternatives shipped within roughly 4 weeks of Jev’s launch. Clef, pplx-decider and Standard One add the image input Jev lacks. Perplexity’s flash-v1.2 undercuts Jev at $0.01 per 1M input tokens. In Respan’s run, Jev still leads accuracy at 0.932. Almost every score is vendor-run, so benchmark on your own data. The post 20 Jev Alternatives: The Decision Models Taking On TypeSafe in 2026 appeared first on MarkTechPost.
StepFun has released Step 5 Preview as open weights. FP8 and BF16 checkpoints are live on Hugging Face and ModelScope. Step 5 Preview is a sparse Mixture-of-Experts model from StepFun built for long-horizon agentic work. It has 600B total parameters, 27B active per token and up to 1M tokens of context. For developers, this is a new self-hostable option for coding agents, document analysis and finance workflows. TL;DR Size: 600B total, 27B active per token, up to 1M-token context. Runs on: data-center GPUs only. FP8 needs 4x B300 or 8x H200. BF16 needs 8x B300 or 16x H200. Performance: near open-weight peers GLM-5.3 and Kimi K3 on coding. Behind GPT-6 Astra and Claude Opus 5. Best: 88.3% on AA-LCR long-context reasoning (Artificial Analysis). Worst: 33.3% on Terminal-Bench 4.0, versus 41.9% for GLM-5.3. What is Step 5 Preview? Step 5 Preview is StepFun’s flagship model for agentic work. StepFun positions it for software engineering, professional knowledge work and finance. The model navigates large codebases, coordinates multi-file changes, runs code and iterates on results. It also analyzes long documents, charts, screenshots and video to produce structured reports. The API launched September 20. Weights followed on October 14. How does the architecture work? Only 27B of the 600B parameters activate per token. StepFun’s tech team adds the internals: Depth: 92 transformer blocks in a narrow, deep design. Attention mix: gated sliding-window attention and full attention layers, interleaved 3:1. Sparse GQA: full-attention layers pick 4,096 history tokens per query, scored in blocks of 8. Speculative decoding: an MTP-3 head gives roughly 30% end-to-end speedup. Training: FP8 mixed precision, with RL on million-level verifiable agentic environments. How does Step 5 Preview perform on benchmarks? StepFun ran its own evaluations at the High reasoning setting. Rivals in its chart ran at Max effort, so the comparison is not like-for-like. Results reported on the Hugging Face card: DeepSWE v1.1: 67.7 (SWE-agent harness, temperature 1.0, top_p 0.95) Terminal-Bench 4.0: 33.3 Agents’ Last Exam (ALE-CLI): 29.5 Independent numbers come from Artificial Analysis via OpenRouter. The model scores 43.7 on the Intelligence Index. It hits 88.3% on AA-LCR and 46.5% on Humanity’s Last Exam. GDPval-AA lands at 54.3%. Vals AI results show clear weak spots. SRE Bench is 0.4%. Harvey’s Legal Agent Benchmark is 9.2%. On finance, StepFun’s chart shows 66.4% on FrontierFinance. That beats GLM-5.3 and Kimi K3, but trails Claude Opus 5 at 69.7%. window.addEventListener("message",function(e){if(e.data&&e.data.mtpStep5H){document.getElementById("mtp-step5-frame").style.height=e.data.mtpStep5H+"px";}}); How does it compare with other open-weight models? Benchmark rows come from StepFun’s comparison chart, as transcribed by CellCog and eesel. What hardware does Step 5 Preview need to run? This is a data-center model. The FP8 checkpoint alone is 650.32 GB. Listed configurations: FP8 on vLLM or SGLang: 4 B300 GPUs or 8 H200 GPUs, each on 1 node. BF16 on vLLM: 1 node with 8 B300 GPUs, or 2 nodes with 16 H200 GPUs. Transformers: offline generation only. The full BF16 shard workflow is not yet validated end to end. vLLM and SGLang ship dedicated Docker images with OpenAI-compatible APIs. On Blackwell, expert-parallel modes use the MegaMoE backend. StepFun recommends temperature 1.0, top_p 0.95 and top_k 0. How much does the API cost? The StepFun API charges $1.00 per 1M input tokens and $2.70 per 1M output tokens. Cache hits cost $0.10 per 1M. Output pricing includes reasoning tokens. The API supports low, medium and high reasoning effort, tool calling and JSON Schema output. Key Takeaways 600B MoE, 27B active, 1M context, now open weights in FP8 and BF16. Smallest listed self-host setup: 4x B300 GPUs with FP8. Best independent score: 88.3% on AA-LCR long-context reasoning. Weakest spot: 33.3% on Terminal-Bench 4.0. API costs $1.00 in and $2.70 out per 1M tokens. Check out the model weights on Hugging Face, GitHub repo, tech blog, BF16 weights, ModelScope and API docs. All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. The post StepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Active, 1M Context appeared first on MarkTechPost.
OpenAI's updated Chrome extension lets ChatGPT access open browser tabs, sum up content, and use "Computer Use" in Work mode to fill out forms or compare products across several tabs. It is open to all ChatGPT plans. The article ChatGPT becomes a full Chrome browser assistant that reads and acts on your open tabs appeared first on The Decoder.
China plans to restrict exports of 23 key AI technologies, including model weights, source code, and training methods for systems hitting 10^26 FLOPs or 2 trillion parameters. A fast-track rule will still clear open-source releases as soon as they are registered, but the move will likely push Western companies to double down on US open-weight models over Chinese alternatives. The article China plans to tighten export controls on open-weight models and 23 key AI technologies appeared first on The Decoder.
Google Deepmind's Gemini 4 Argon scored 97.5 percent on the ARC-AGI-3 benchmark, making Google the second lab after OpenAI to cross the 95 percent mark. But the results swing wildly depending on the test setup: in the standard harness, the score drops to just 49.8 percent. Argon also trails rivals in math and software engineering. The article Google's Gemini 4 Argon is closer to the AI frontier than any Google model in months appeared first on The Decoder.
