Google has open-sourced TorchTPU, a native PyTorch backend for Cloud TPUs. It was co-developed with the PyTorch team at Meta and announced at PyTorch Conference 2026 on October 20, 2026. Existing PyTorch code targets a TPU through a standard torch.device("tpu"). TPUs now sit beside GPUs and CPUs as a native PyTorch device. Google team states that TorchTPU replaces PyTorch/XLA and recommends it for all new workloads. TL;DR Size: A software stack, not a model. 4 open repos: TorchTPU, vLLM-TorchTPU, SGLang-TorchTPU, TorchTitan. TorchTPU is v0.2.0 on PyPI, Python 3.11+, Apache 2.0. Runs on: Google Cloud TPUs. Tested on generations from v5e, scaled to 8,192 Ironwood (TPU7x) chips. Serving supports FP8, MXFP4 and NVFP4 checkpoints. Performance: Eager mode for debugging, torch.compile for production-level throughput. Best: 52% MFU training Qwen 3 8B with TorchTitan on Ironwood (self-reported). Worst: Out-of-box eager took 11.49s per step in a fine-tuning case, versus 2.78s fully tuned. What is TorchTPU? TorchTPU is an open-source PyTorch backend from Google and Meta that runs PyTorch models natively on TPUs. It plugs in through PyTorch’s PrivateUse1 interface. Tensors are ordinary PyTorch tensors, with no subclasses or wrappers. It runs one process per device and preserves tensor views and in-place mutations. The predecessor, PyTorch/XLA, recorded lazy-tensor graphs. Developers often added code to mark compilable regions. PyTorch/XLA also only supported pure SPMD code. TorchTPU also handles divergent MPMD execution. How does TorchTPU execute PyTorch code? TorchTPU ships 3 eager modes: Strict (default): dispatches each operator asynchronously with standard PyTorch semantics. Debug: runs synchronously with tensor bounds checking. DeferAndFuse: fuses runs of eager operators to cut HBM traffic. The TorchTPU paper says it can be 50% or more faster than Strict. For peak speed, torch.compile(backend="tpu") captures the graph with TorchDynamo. It lowers through StableHLO to OpenXLA, which emits TPU machine code. Compiled executables are cached across processes, hosts and job restarts. Custom Pallas kernels are callable as standard PyTorch functions. Prebuilt Tokamax kernels and Helion tile kernels also work. Distribution uses the tpu_dist process group. It supports DDP, FSDP2 and DTensor, along with a topology-aware mesh for TPU interconnects. How fast is TorchTPU? All figures below are self-reported by Google and partners. Training scale: TorchTitan reached up to 52% MFU on Qwen 3 8B. Google reports near-linear scaling up to 8,192 Ironwood TPUs. Tuning ladder: In a TorchTitan fine-tuning case, step time fell from 11.49s to 2.78s. DeferAndFuse alone gave 1.15x. Full torch.compile gave 3.80x. Model coverage: Google continuously tests about 5,000 Hugging Face models in eager and compiled modes. Fleet serving: Zenteiq grew daily output from 1.47B to 5.97B tokens per 1,000 TPU v6e chips. That gain came from llm-d routing on GKE, not TorchTPU alone. #mtp-ttpu-wrap iframe{width:100%!important;border:0!important;display:block!important;background:transparent!important}#mtp-ttpu-wrap p:empty,#mtp-ttpu-wrap hr{display:none!important} window.addEventListener("message",function(e){if(e.data&&e.data.mtpTorchTPU){var f=document.getElementById("mtp-ttpu-frame");if(f&&e.source===f.contentWindow){f.style.height=e.data.h+"px";}}}); What integrations ship with TorchTPU? Hugging Face: Transformers v5 deploys to TPU with device_map="tpu". Diffusers supports eager, torch.compile and tensor parallelism. vLLM-TorchTPU: built with Inferact. Keeps upstream schedulers and API servers, with custom Pallas kernels. SGLang-TorchTPU: built with RadixArk. Keeps RadixAttention prefix reuse. Serving features: FP8 KV caches, FP8, MXFP4 and NVFP4 checkpoints. Speculative decoding includes MTP and EAGLE-3. TorchTitan: recipes for Llama 3 70B/405B, Qwen 3 8B/32B, Qwen 3 235B MoE and FLUX.1 12B. Also: llm-d 1.0, Ray, TorchRec, VeRL (preview) and LiteRT-Torch for on-device export. How does TorchTPU compare with PyTorch/XLA, TorchNeuron and JAX? How do you install TorchTPU? Install the package on a TPU VM, then pick the tpu device: Copy CodeCopiedUse a different Browser# pip install torch_tpu import torch device = torch.device("tpu") x = torch.randn((4, 4), device=device) The repo’s README pins torch_tpu==0.2.0 with the CPU PyTorch wheel index. Without it, pip pulls the CUDA build of torch. Full guides live at torchtpu.readthedocs.io. Key Takeaways TorchTPU is open source under Apache 2.0, co-built with Meta. It replaces PyTorch/XLA with eager-first, PrivateUse1-based execution. torch.compile(backend="tpu") lowers through StableHLO and XLA. TorchTitan hit 52% MFU on Qwen 3 8B, scaling to 8,192 Ironwood chips. vLLM, SGLang and Hugging Face now run on TPU through standard PyTorch. Check out the GitHub repo, paper, docs and Google Cloud announcement. 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 Google Open-Sources TorchTPU, a Native PyTorch Backend for Cloud TPUs appeared first on MarkTechPost.
DeepSeek has open-sourced DeepInfer, the inference framework that runs its production services. The DeepSeek DeepInfer inference framework is the engine behind DeepSeek-V4.1-Flash. It ships under the Apache 2.0 license. DeepSeek also released deepseek-recipe, its MIT-licensed token encoding and decoding layer. For developers, this is a rare look at a frontier lab’s real serving stack. TL;DR Size / scope: 1 supported model (DeepSeek-V4.1-Flash), 1M-token context window per the model card. Runs on: NVIDIA SM100 (Blackwell) GPUs and Huawei Ascend 950 NPUs. Benchmarks used 4 GPUs for prefill and 64 for decode. Performance: Production-grade throughput; DeepSeek says 1 cluster serves 155 trillion tokens per day. Best: Above 25,000 output tokens/s/GPU at 256 requests per rank (EP64, 131,072 context). Worst: Offline inference only. KV cache system, P/D disaggregation and load balancing are not included. What is DeepInfer? DeepInfer is a high-performance inference engine from DeepSeek for its own models. According to it’s GitHub README, it powers DeepSeek’s large-scale production services. This release covers offline inference with the core features needed at scale. It was originally built primarily on vLLM, with ideas from SGLang and TensorRT-LLM. The deployment Q&A gives production context. A single cluster serves 155 trillion tokens per day. DeepSeek reports overall revenue of US$38 per GPU-hour at valley pricing. How does DeepInfer work? The engine stacks DeepSeek’s kernels: FlashMLA, DeepGEMM, DeepEP, TileKernels, DeepSelect and custom TileLang kernels. On top sit 5 system-level ideas. Parallelism: Attention DP plus Wide-EP Attention uses data parallelism. The MoE layers use expert parallelism with an expert-parallel load balancer (EPLB). Decode runs EP64 to EP128 inside a supernode. Prefill uses the smallest EP that memory allows, typically EP2 to EP8. DSpark: confidence-scheduled speculative decoding A draft model proposes 5 tokens per step. A Confidence Head predicts each draft token’s acceptance rate. The scheduler then verifies only the most valuable prefix. Per the Q&A, this cuts verified draft tokens by about 30% (5.00 to 3.2 to 4.4). Accepted length barely moves (3.29 vs 3.20). SWA bounded replay On a cache hit, DeepInfer recomputes only the last 128 tokens of sliding-window attention state. That makes recomputation cost constant, regardless of context length. In the decoder layers, DeepSeek says this nearly halves total prefill compute. Asynchronous scheduling, determinism and structured output CPU batch scheduling overlaps with GPU execution, so the GPU never idles between steps. Batch-invariant inference and deterministic sampling are available on NVIDIA only. Structured outputs support regex, JSON schema and EBNF grammars. window.addEventListener("message",function(e){if(e.data&&e.data.id==="mtp-deepinfer-x"&&e.data.mtpEmbedHeight){var f=document.getElementById("mtp-deepinfer-frame");if(f){f.style.height=e.data.mtpEmbedHeight+"px";}}}); How fast is DeepInfer? DeepSeek publishes Pareto curves for DeepSeek-V4.1-Flash at 131,072 context. Prefill ran on 4 SM100 GPUs; decode ran on 64. Figures below are read from those charts and are self-reported. Decode (EP64DP64, nextn=3, 75% acceptance): Above 25,000 tokens/s/GPU at 256 requests per rank, near 100 tokens/s per user. At 8 requests, per-user speed passes 300 tokens/s. Prefill (EP4DP4, 140-token bounded replay): About 60,000 non-cached tokens/s/GPU at 4,096 tokens per rank. Throughput nears 90,000 at 65,536 tokens, with roughly 730 ms per step. Wide-EP payoff: An EP8 vs EP64 chart shows EP64 ahead at every batch size. DSpark’s dynamic verification adds 4% to 7% throughput at about 60 sequences per rank. At 16 sequences per rank it is 1.5% slower than fixed 5-token verification. How does DeepInfer compare with vLLM, SGLang and TensorRT-LLM? What are DeepInfer’s limitations? A full service still needs a KV cache system and VE/Prefill/Decode disaggregation. It also needs load balancing across data-parallel replicas. DeepSeek team states that those components are tightly coupled to internal infrastructure. Hardware is narrow. Builds require Linux, GCC 13 or later, Python 3.12, and CUDA 13 or CANN 9.2. The smallest efficient production unit is 64 prefill GPUs plus 64 decode GPUs. The internal FlashMLA decoding kernel is 5% to 10% faster than the public one. End-to-end, DeepSeek puts the gap at 5% or less. Key Takeaways DeepInfer is DeepSeek’s production engine, now Apache 2.0 on GitHub. It supports only DeepSeek-V4.1-Flash, on NVIDIA SM100 and Ascend 950. Decode tops 25,000 tokens/s/GPU on 64 GPUs (self-reported). DSpark cuts verified draft tokens about 30% with near-equal acceptance. Offline only: no KV cache store, P/D disaggregation or load balancer. Check out the GitHub repo, deployment Q&A, deepseek-recipe 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 Open-Sources DeepInfer, the Inference Framework Behind DeepSeek-V4.1-Flash appeared first on MarkTechPost.
The Army Cyber Command is developing AI agents that move through other networks like a computer worm and "neutralize" targets, according to internal documents. The program, codenamed ECHOTRIBBLE, is designed to reach even air-gapped systems with no network connection and run with minimal human oversight. Experts warn about collateral damage, pointing out that AI agents have recently broken into outside systems without meaning to, even without any order to attack. The article Army Cyber Command's ECHOTRIBBLE would turn AI agents into self-spreading cyberweapons appeared first on The Decoder.
Cloudflare has released an AI model it says pulls data from documents faster and more cheaply than large language models. The article Cloudflare's Clairic pulls data from documents faster and cheaper than LLMs appeared first on The Decoder.
The US Department of Homeland Security has banned federal civilian agencies from using AI technology from Chinese lab Deepseek. The article US bans Deepseek technology across federal civilian agencies appeared first on The Decoder.
Exa has released Exa 3.0, an upgrade to its web search API for AI agents. The Exa 3.0 search API ships a larger embedding model and a cheaper search type, Exa Lite. Dynamic Highlights is now generally available. Exa claims that it leads Perplexity, Brave and Parallel on 8 benchmarks. Those benchmarks were run by Exa, so you can treat them as self-reported. For developers, the practical change is cost: Lite costs $4 per 1,000 requests. TL;DR Size: Exa tracks over 2 trillion URLs and serves 100 billion pages. Embedding model size: Not disclosed. Runs on: Hosted API only. Exa lists configurable search latency of 450ms to 1s. No self-hosted option. Performance: Exa 3.0 Auto ranks first on all 8 Exa-run benchmarks against Perplexity, Brave and Parallel. Best: 91.8% on BrowseComp, plus the widest lead on ATLAS (42.0% vs 37.2% for Perplexity). Worst: 42.0% on ATLAS is its lowest percentage score. Its BrowseComp lead over Perplexity is only 0.3 points. What is Exa 3.0? Exa 3.0 is a search API release from Exa Labs that upgrades retrieval quality across every search type. It is available through the API. Exa runs its own crawler and an independent index. It is not a wrapper on Google or Bing results. How does Exa 3.0 improve retrieval? Exa attributes the gains to 3 changes in its retrieval stack: A larger embedding model: Exa trained it with more compute and data than any earlier model. Parameter count is not disclosed. Dynamic Highlights, now generally available: The model selects tokens across all results, not page by page. It gives more tokens to better pages and drops repeated information. Billions of new documents: New coverage includes products, local businesses, flights, events, sports and weather. According to the Exa’s post, Dynamic Highlights uses 95% fewer tokens than full page text while improving answer quality. Highlights carry no extra charge. That matters for agents, where context tokens often cost more than search calls. What is Exa Lite? Exa Lite is a new search type, set with type: "lite" on the /search endpoint, per the changelog. It runs on the same index as Exa Auto. Auto spends more compute on candidate generation, highlights and ranking. Lite supports query, result count, highlights, text, categories, domain and date filters, location and moderation. Unsupported parameters return a 400 error naming the field. Exa now recommends 3 search types: lite, auto and deep. The fast, instant and deep-reasoning types are legacy but still work. How does Exa 3.0 perform on benchmarks? Exa compared each provider’s highest-scoring single-step endpoint: Exa 3.0 Auto, Perplexity Web, Brave LLM Context and Parallel Advanced. All ran in one agentic harness. The agent was GPT-6 Luna, with up to 25 search rounds and 10 results per search. Evaluation happened in October 2026. #exa3-mtp-wrap iframe{display:block;width:100%!important;border:0!important;background:transparent!important}#exa3-mtp-wrap hr,#exa3-mtp-wrap p:empty,#exa3-mtp-wrap del,#exa3-mtp-wrap s{display:none!important} (function(){var f=document.getElementById("exa3-mtp-frame");window.addEventListener("message",function(e){if(e.data&&e.data.type==="exa3-mtp-height"&&e.source===f.contentWindow){f.style.height=e.data.h+"px"}})})(); The margins vary widely. ATLAS and WebCode show the clearest leads, at 4.8 and 5.6 points over Perplexity. BrowseComp, DSQA and WideSearch leads are each under 1.5 points over the runner-up. ATLAS is a benchmark Exa published itself on October 8, 2026. On cost, Exa plots agentic search score against cost per task. Lite scores about 79 at about $0.045 per task. Auto scores about 80.5 at about $0.065. Exa says the pair sets a new Pareto frontier among tested providers. How does Exa 3.0 compare with Perplexity, Brave and Parallel? All benchmark scores come from Exa’s Exa 3.0 post and were run by Exa. No vendor has published matching results. Prices come from each vendor’s own pricing page. What data does the expanded index add? Exa states that it prioritized personal AI assistant use cases. The official technical page lists 1.27B ecommerce listings with prices and availability, and 211M+ local businesses. It also adds real-time news, sports, weather and market data, along with flights, events, transit and podcasts. The Exa Library structures this data into searchable collections. These include 320M people pages, 33M company pages and 500M research paper pages. It also lists 19M SEC filings, 10M patents and 9M court opinions. Where does Exa 3.0 fall short? Self-run evals: Exa picked the harness, endpoints and agent model. ATLAS is Exa’s own benchmark. Thin margins: Several leads are under 2 points, which may not survive other harnesses. Lite limits: Lite rejects unsupported parameters with a 400, so some pipelines need Auto. Higher Auto list price: Auto costs $7 per 1,000 requests versus $5 for the 3 rivals. Opaque model details: Embedding model size and architecture are not disclosed. Key Takeaways Exa 3.0 leads 8 benchmarks, all run by Exa itself. Exa Lite costs $4 per 1,000 requests on the same index as Auto. Dynamic Highlights uses 95% fewer tokens than full pages, at no extra cost. Biggest leads: WebCode (+5.6) and ATLAS (+4.8) over Perplexity. Test on your own queries before switching providers. Check out the Exa 3.0 announcement, the Exa Library, the changelog and the 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 Exa 3.0 Search API Tops Perplexity, Brave and Parallel on 8 Benchmarks appeared first on MarkTechPost.
Cloudflare has released a preview of Clairic, a new System One extraction model on Workers AI. The Cloudflare Clairic extraction model reads text or images and returns typed values, not generated text. It targets a job developers usually give to LLMs: pulling totals, dates and IDs into JSON. Clairic answers every requested field in 1 forward pass, with no reasoning and no output tokens. TL;DR Size: Parameter count not disclosed. 65,536-token context window, up to 4 images per request. Runs on: Cloudflare-hosted GPUs via Workers AI (@cf/cloudflare/clairic). No local option until weights ship. Performance: Sub-second extraction in Cloudflare’s tests. No standard accuracy benchmark is published yet. Best: Extraction step took 0.46 s vs 16.1 s for GLM-5.3 in Cloudflare’s claims pipeline. Worst: 4 of 21 TSA next-day forecasts fell outside its 80% interval. Multi-step arithmetic is unreliable. What is Clairic? Clairic is a Cloudflare model that extracts, computes, estimates and forecasts typed values from unstructured inputs. It builds on the System One idea from TypeSafe’s Jev, named for Kahneman’s fast, intuitive thinking. Jev and Cloudflare’s own Clef are decision models. They score options you list in advance. Clairic covers answers you cannot list, such as an invoice total, a date or a repair cost. Cloudflare frames a 3-part stack. Extraction models turn raw input into typed values. Decision models turn values into decisions. LLMs turn decisions into words. How does a Clairic request work? A request has a state (text plus up to 4 images) and a map of named, typed fields. Each field can carry plain-language instructions. Supported types are text, number, date, time, datetime, duration, list, estimate and forecast, per the Workers AI changelog. Every scalar answer includes p_null, the probability that the value is absent, plus a confidence score. Text answers return exact spans with character offsets. Estimates and forecasts return a median and an 80% interval, from p10 to p90. Forecasts cover up to 64 steps. There is no field limit. Fields that do not fit beside the input in one 8,192-token pass split into parallel passes. Each pass sees the whole input. In the docs example, a 61-token invoice ran in 3 passes. It billed 655 input tokens, including 472 schema tokens. Wide schemas can cost more than the document itself. What did Cloudflare’s tests show? In a support-email example, Clairic added price and tax into $701.10. It also computed 11 days since delivery. Neither number appears in the email. A missing tracking number returned null instead of a plausible guess. On a skewed receipt photo, it read all 7 present values and returned null for the absent discount. Its lowest confidence was 0.776, on the order number. For forecasting, Cloudflare predicted daily TSA checkpoint counts for 21 days. Average error was 3.3%. 17 of 21 actual counts fell inside the 80% interval. 3 of the 4 misses fell on Columbus Day weekend. The main test is a damaged-product claims pipeline. Clairic extracted facts, Clef chose refund or repair, and GLM-5.3 wrote the reply. Cloudflare compared this against GLM-5.3 handling all 3 steps. #mtp-clairic-wrap{max-width:880px!important;margin:24px auto!important;padding:0!important;background:transparent!important;border:0!important;} #mtp-clairic-wrap iframe{display:block!important;width:100%!important;border:0!important;background:transparent!important;height:600px;} #mtp-clairic-wrap p:empty,#mtp-clairic-wrap hr,#mtp-clairic-wrap del,#mtp-clairic-wrap s{display:none!important;} @media (max-width:640px){#mtp-clairic-wrap{margin:16px auto!important;}} (function(){var f=document.getElementById('mtp-clairic-frame');window.addEventListener('message',function(e){if(e.data&&e.data.mtpClairicHeight&&f&&e.source===f.contentWindow){f.style.height=e.data.mtpClairicHeight+'px';}});})(); The cascade finished in 19.3 s versus 61.7 s, about 3.2x faster. Extraction took 0.46 s against 16.1 s. The decision step took 0.22 s against 21.9 s. Cost per million claims fell 49%, from $11,740 to $6,032. The reply step still dominated, at 18.6 of the 19.3 s. Both pipelines got the facts and decision right. Clairic’s repair estimate was identical across runs. GLM-5.3 returned $250, $250 and $220. Cloudflare says Gemma 4 and Qwen failed the all-LLM version. How does Clairic compare with Clef, Jev and GLM-5.3? Where can developers use it, and what does it cost? Clairic is live as a preview at @cf/cloudflare/clairic. Call it through the Workers AI binding with env.AI.run() or the REST API. AI Gateway works with the same endpoints. Pricing is $0.10 per million input tokens with no output charge. That undercuts Clef’s $0.24 but sits above Jev’s $0.042. Weights are promised but not yet released. Key Takeaways Clairic returns typed values, not text, in 1 forward pass. $0.10 per million input tokens; output tokens cost nothing. Extraction ran in 0.46 s vs 16.1 s for GLM-5.3. Every answer carries confidence, p_null, spans or intervals. Check out the Cloudflare blog post, the Clairic model docs, the extraction models explainer and the Workers AI changelog. Also read our earlier coverage of Drex 1.5. 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 Cloudflare Releases Clairic, a System One Extraction Model on Workers AI appeared first on MarkTechPost.
Inception has released Mercury 3, the newest version of its diffusion language model. Mercury 3 is a diffusion LLM from Inception that writes answers in parallel, not token by token. Inception reports 1,068 tokens per second on NVIDIA GPUs and a 1M-token context window. On Inception’s own benchmark table, it leads GPT-6 Luna and Claude Haiku 5.5 on 8 of 12 tests. For developers building agents, voice bots and search tools, this is a speed-first model with competitive scores. TL;DR Size: Parameters not disclosed. Inception says it is more than 40x larger than Mercury 1. Context is 1M tokens. Runs on: Inception-served NVIDIA GPUs (exact GPU not disclosed). API only, no open weights or quantized builds. Performance: Leads 8 of 12 benchmarks vs GPT-6 Luna and Haiku 5.5 (self-reported, medium settings). Best: SWE-Bench Atlas 51.08, about 18 points ahead of both rivals. Worst: DeepSWE v1 46.31, 5.31 points behind Haiku 5.5 (51.62). What is Mercury 3? Mercury 3 is Inception’s fourth Mercury release, after Mercury 1, Mercury 2 and Mercury 2.5. Mercury 1 launched in 2025. Mercury 2.5 added tunable reasoning and a 260K context window, per Inception’s launch announcement. How does a diffusion LLM work? Autoregressive models write one token after another. Each new token waits for the previous one. A diffusion LLM starts from a rough draft of the full answer. It then refines every position over a few passes in parallel. Fewer sequential steps means faster output on the same GPUs. The open question was whether this approach holds up as models grow larger. Does diffusion scale? Here Inception’s answer is yes. Mercury 3 has more than 40x the total parameters of Mercury 1 (mini). Average quality across standard benchmarks rose about 3x between those 2 models. Speed stayed above 1,000 tokens per second. Inception team plots this as a quality vs size curve across 5 Mercury generations. How does Mercury 3 perform on benchmarks? All scores below come from Inception’s table. Every model ran at medium settings. Mercury 3 posts its largest gains on instruction following and agentic coding. IFBench reaches 76.84, 15 points above GPT-6 Luna. SWE-Bench Atlas reaches 51.08, against 33.33 for Luna and 32.26 for Haiku 5.5. GPQA Diamond lands at 89.23. The losses matter too. Haiku 5.5 wins DeepSWE v1 (51.62 vs 46.31) and MRCR v2 (78.40 vs 75.50). GPT-6 Luna edges ahead on AA-Omniscience-Public (46.94 vs 46.44) and AA-LCR v1.1 (78.30 vs 77.67). Long-context retrieval and longer coding tasks remain the weak spots. (function(){var f=document.getElementById('mtp-m3x-frame');window.addEventListener('message',function(e){if(e.data&&e.data.id==='mtp-m3x'&&e.data.mtpH&&e.source===f.contentWindow){f.style.height=e.data.mtpH+'px';}});})(); Mercury 3 vs GPT-6 Luna vs Claude Haiku 5.5 Speed and benchmark rows: Inception’s table, medium settings. Specs and prices: each vendor’s own docs. What do early users report? Inception quotes 4 early-access partners. These are customer statements, not independent evals. TinyFish ran 50 live Mind2Web web-navigation tasks. Mercury 3 completed 50% of them. Inception says that beat GPT-6.1-Sol but gives no Sol number. Octen AI reports 0.9 seconds median generation and 2.2 seconds end to end for smart-glasses search. RevisionDojo says hints and marked answers return in about a second. Lorikeet says Mercury 3 beat other models on latency and evals for support agents. How can developers use Mercury 3? Mercury 3 is live on the Inception API Platform, OpenRouter, Baseten and Vercel. The API is OpenAI compatible. The docs use the model string mercury-3 and a reasoning_effort parameter. List pricing is $0.20 per 1M input and $0.70 per 1M output tokens. At launch it is 50% off: $0.10 and $0.35. Cached input is 90% off. New accounts get 100M free tokens. Enterprise plans add dedicated capacity, autoscaling and configurable data retention. Inception also ships 2 siblings: Mercury Voice targets time to first answer token in voice agents. Mercury Decide is a classification model. Key Takeaways Mercury 3 is a diffusion LLM claiming 1,068 tok/s and 1M context. It leads 8 of 12 vendor-run benchmarks vs GPT-6 Luna and Haiku 5.5. Biggest wins: SWE-Bench Atlas 51.08 and IFBench 76.84. Haiku 5.5 still wins DeepSWE v1 and MRCR v2. Launch price is $0.10 / $0.35, rising to $0.20 / $0.70. Check out the official launch blog, the API docs and the OpenRouter listing. Also read our earlier coverage of Drex 1.5. 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 Inception Releases Mercury 3 Diffusion LLM at 1,068 Tokens per Second appeared first on MarkTechPost.
MiniMax's M3.1 Flash jumps 13 points over its predecessor to 42 on the Artificial Analysis Intelligence Index, at half the cost. It also hits 73.7 percent on the DeepSWE v1.1 coding benchmark for under $0.30 per task, matching Deepseek V4.1 Flash 1017 on both performance and price. The article MiniMax M3.1 Flash shows China's AI price race has no floor in sight appeared first on The Decoder.
