Linkup just open-sourced SPARSEUP, a 149M sparse embedding model scoring 56.4 on BEIR-13.

It is a SPLADE-style retriever that outputs readable token weights, not an opaque dense vector. It fills the sparse slot next to LightOn's DenseOn and LateOn, trained on the same open data.


✅ Apache 2.0, weights on Hugging Face

✅ ModernBERT backbone, English only

✅ 56.4 nDCG@10 on BEIR-13, excluding MS MARCO

✅ Beats splade-v3 (51.7) and opensearch doc-v3-gte (54.6)

✅ ~380µs per query at >97% recall, Seismic, MS MARCO

✅ 47 non-zeros per query, 190 per document

✅ Logit shift of 15 keeps vectors sparse from the start

✅ Top-12 expansion cap per token, not per vector

✅ Case folding cuts output from ~50k to ~34k dims


Every dimension is a token, so you can read exactly why a document matched.

With identical backbone and data, it still trails DenseOn by 1.52 points and LateOn by 2.5. Linkup also documents weak expansions, like number queries spilling into unrelated terms.

Full breakdown: https://www.marktechpost.com/2026/09/19/linkup-research-releases-sparseup/

Model: https://huggingface.co/Linkup-Platform/linkup-sparseup-embed-v1

Technical blog: https://www.linkup.so/blog/introducing-sparseup-by-linkup

TypeSafe AI released Jev, a model from an RLHF (reinforcement learning from human feedback) co-inventor that does not generate text.

You send it state and typed questions. It returns decisions with probabilities
your code can branch on, so there is nothing to parse.

- Founder Diogo Almeida co-invented RLHF at OpenAI
- $40M seed round led by DCVC
- 3 primitives: Choice, Score, Noul
- Every question in a request scored in parallel
- $0.042 per 1M input tokens, output tokens free
- 70ms to 500ms end to end, vendor reported
- 0.114s vs 8.566s in TypeSafe's own recorded demo
- Vercel: up to 18x faster than GPT Luna, p95
- Listed on Vercel AI Gateway as typesafe-ai/jev

Confidence is the product. Your code sets the threshold where automation stops
and a human takes over.

The 193.6x faster and 444.6x cheaper claims come from TypeSafe's own evals, not independent tests.

Read the full analysis: https://www.marktechpost.com/2026/09/19/typesafe-ai-releases-jev/

Technical details: https://typesafe.ai/blog/introducing-system-one-models-and-jev

Join the waitlist: https://typesafe.ai/

Nokia AI Research Open-Sources AnyJev: A Training-Free Layer That Turns Any Open LLM Into a Calibrated Decision Model

It reads typed answers straight from the model's next-token distribution. It removes position and label bias without any training.

▪️ Apache-2.0, pip install "anyjev[hf]"

▪️ Choice, yes/no and score questions, nothing generated

▪️ Hugging Face transformers and vLLM backends

▪️ L0 needs zero labels, L1 needs 100 to 500

▪️ Order-flip rate 0.230 to 0.073 at L0

▪️ Calibration error 0.240 to 0.095 at L1

▪️ Auto-decidable at 5% error: 7.7% to 52.0% at L1

▪️ Tested on Qwen3-8B, BANKING77 20-way, 300 test items


Accuracy moves 6 points, but the traffic you can safely automate grows 6.8x.

Full analysis: https://www.marktechpost.com/2026/09/23/nokia-open-sources-anyjev-a-training-free-layer-that-turns-any-open-llm-into-a-calibrated-decision-model/

Repo: https://github.com/nokia-applied-research/AnyJev

NVIDIA Releases Nemotron 3 Diarization: A 100M-Parameter Open-Weight Model That Tracks 8 Speakers in Real Time

It answers "who spoke when" in a conversation, even when people talk over each other.
One open-weight checkpoint handles both offline audio and real-time streaming.

✅ 100M parameters, open weights on Hugging Face

✅ OpenMDW-1.1 license, commercial use allowed

✅ Tracks up to 8 speakers, overlap included

✅ 14.72% DER on Diarization-Bench vs 19.3% runner-up

✅ 4 streaming presets, from 30.4 s down to 0.32 s

✅ 41.0% average relative DER cut vs Streaming Sortformer

✅ 15,113x batched RTFx on RTX PRO 5000, 30.4 s preset

✅ Runs through NVIDIA NeMo on Linux GPUs

Full breakdown + interactive explainer: https://www.marktechpost.com/2026/09/23/nvidia-releases-nemotron-3-diarization/

Model: https://huggingface.co/nvidia/Nemotron-3-Diarization
NVIDIA blog: https://huggingface.co/blog/nvidia/nemotron-diarization
Live demo: https://huggingface.co/spaces/nvidia/nemotron-diarization

Search is now the bottleneck for AI agents, and Perplexity just rebuilt theirs from the ground up.

Photon is their new retrieval and ranking engine, written in Rust. It replaces the open-source engine they had forked for years.

The old engine's problem: the index outgrew RAM. Cold reads caused page faults, p99 sat near 800 ms, and index merges pushed it to ~1.2 s.

Photon's fix, in 4 moves:

1/ Compact posting lists that pick a format per block: inline, sparse arrays, or bitmaps
2/ A budgeted, WAND-like traversal that skips candidates before reading exact scores
3/ Disk reads batched through io_uring, so waits overlap instead of piling up
4/ Index builds moved off the serving nodes entirely

The result in production: p99 dropped to ~65 ms, on ~20% fewer machines, while storing ~2.5x more data per document.

For developers, Photon powers a new Fast Search mode:

→ 160 ms p50 / 230 ms p95 per call
→ $1 per 1K requests (vs $5 standard)
→ ~68% lower estimated agent task cost at comparable quality

Technical details + full breakdown in the reply 👇
Full breakdown: https://marktechpost.com/2026/09/30/perplexity-introduces-photon-a-rust-based-retrieval-engine-that-cuts-p99-latency-from-800-ms-to-65-ms/

Technical details: https://perplexity.ai/hub/blog/photon

Google DeepMind just announced Gemini 4 Argon, raising the output limit from 64K to 1M tokens in a single response. It is the first model of the Gemini 4 generation, built for long-horizon coding, enterprise knowledge work, and cyber defense.

The core shift is output depth. With room to generate hundreds of thousands of tokens in one trajectory, Argon can tackle large refactors and multi-step problems without splitting work across turns.

In Google's own comparison, Argon leads GPT-6 Astra and Claude Opus 5.5 on DeepSWE v1.1 (77.9%), the Vals Index (68.9%), and AutomationBench (51.3%). It is not a sweep, though: it trails on FrontierSWE v2, Terminal-Bench 4.0, and OSWorld-2.0.

Access is gated. Argon is rolling out first to cyber defenders in the Fairwind Program, with paid API customers and Google AI Ultra subscribers next. Introductory pricing is $2 input / $10 output per 1M tokens......

Full analysis: https://www.marktechpost.com/2026/09/30/google-deepmind-unveils-gemini-4-argon-with-1m-output-tokens-for-coding-knowledge-work-and-cyber-defense/

Announcement: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/

Fairwind Program: https://deepmind.google/fairwind-program/

Cloudflare Releases Clef: Open-Weight Decision Models That Return Typed Probabilities Instead of Text

It answers typed questions about an input (yes/no, pick one, rate on a scale) in a single forward pass.

No tokens are generated, so there is no output to parse before your agent acts.

✅ 2 sizes: Clef (27B) and Clef-flash (9B), open weights on Hugging Face
✅ Apache 2.0 license, commercial use allowed
✅ Post-trained from Qwen3.8-27B and Qwen3.5-9B, backbones kept frozen
✅ 38.8 ms median latency for Clef-flash vs 524.1 ms for Jev
✅ 94.20 macro-F1 on BANKING77 vs 79.74 for Jev
✅ Reads text, JSON, images and video, 64K-token context
✅ Up to 64 questions and 4 images per request on Workers AI
✅ $0.24 (Clef) and $0.09 (Clef-flash) per million input tokens
✅ Drop-in compatible with the Jev / System One API

Full analysis: https://www.marktechpost.com/2026/10/01/cloudflare-releases-clef-and-clef-flash/

Model: https://huggingface.co/Cloudflare/clef

Docs: https://developers.cloudflare.com/workers-ai/models/clef/

[AI Model Family Series #1] We just published the complete story of Alibaba's Qwen: every major model from 2023 to 2026, with dates, key features and sources.

7B open weights in 2023 → 2.4T open weights in 2026. Here's the lineup the story covers

2023
→ Tongyi Qianwen (Apr): enterprise beta
→ Qwen-7B (Aug): first open weights
→ Qwen-VL (Aug): first vision model
→ Qwen-72B + Qwen-1.8B (Dec)

2024
→ Qwen1.5 (Feb): 0.5B–110B, 32K context
→ Qwen2 (Jun): 57B-A14B MoE, Apache 2.0
→ Qwen2-Math, Qwen2-Audio, Qwen2-VL (Aug)
→ Qwen2.5 (Sep): 18T tokens, 100+ models
→ Qwen2.5-Coder + QwQ-32B-Preview (Nov)
→ QVQ-72B-Preview (Dec)

2025
→ Qwen2.5-VL + Qwen2.5-Max (Jan)
→ QwQ-32B (Mar): Qwen claims R1-level reasoning
→ Qwen2.5-Omni-7B (Mar)
→ Qwen3 (Apr): hybrid thinking, 119 languages
→ Qwen3-2507 + Qwen3-Coder-480B (Jul)
→ Qwen-Image + Qwen-Image-Edit (Aug)
→ Qwen3-Max (Sep): first 1T+ Qwen
→ Qwen3-Next-80B-A3B (Sep)
→ Qwen3-Omni + Qwen3-VL (Sep)

2026
→ Qwen3-Max-Thinking (Jan)
→ Qwen3-Coder-Next + Qwen-Image-2.0 (Feb)
→ Qwen3.5-397B-A17B (Feb): native multimodal agents
→ Qwen3.6-Plus, 35B-A3B, Max-Preview, 27B (Apr)
→ Qwen3.7-Max (May) + Qwen3.7-Plus (Jun)
→ Qwen3.8-2.4T-A95B (Aug): largest open Qwen
→ Qwen3.8-27B + Qwen3.8-Flash (Aug)
→ Qwen-Image-2.1 (Sep)

The story also covers what the list can't show: the DeepSeek moment, the 2026 leadership exit, and Qwen's shift from all-open to a two-tier license strategy.

Next: Qwen 4 is in training. Qwen 4.5 and Qwen 5 are projected at 5–10T parameters.

Read the full report: https://www.marktechpost.com/2026/10/04/the-story-of-qwen-alibabas-ai-models-from-7b-to-2-4t/

Which model family should we map next: DeepSeek, Llama, Gemma, Mistral or Kimi? Drop it in the replies.......

Mistral AI Releases Mistral Large 4 (Le Chonk): A 1.05T Parameter Multimodal MoE Model

Mistral Large 4, nicknamed Le Chonk, went into public preview today. 1.05T total parameters, 49B active per token, 1M context, native image input. Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters.

Here is what stood out to me after reading the release and the docs:


→ Only about 4.7% of the weights fire per token
→ That is how a 1T class model serves at $1.36 per 1M input and $4.18 per 1M output tokens
→ The full 1.05T still has to sit in memory, so plan hardware around total, not active
→ 93% on Cybench, 82% on CyberGym-E2E, top 5 on the Artificial Analysis Cyber Index
→ 61.7% DeepSWE v1.1, 28.3% Terminal-Bench 4.0, 49.8% on the Coding Agent Index
→ API is live today as mistral-large-4

I broke down the architecture, benchmarks and a full comparison against DeepSeek V4 Pro, Kimi K3 and GLM-5.3, with an interactive explainer.

Here is the full analysis: https://www.marktechpost.com/2026/10/06/mistral-ai-releases-mistral-large-4-le-chonk-a-1-05t-parameter-open-weight-multimodal-moe/

Technical details: https://mistral.ai/news/mistral-large-4/

Meta just open-sourced Rebalancer, the assignment solver that has run resource allocation across Meta for 9+ years. It solves ~40M assignment problems per day across 30+ problem formulations, from shard placement to global traffic routing.

The core innovation is separating the problem spec from the solver. You describe objects, bins, constraints and goals once. Rebalancer compiles that into an expression graph, then solves it with either a parallel local search or a MIP solver (Gurobi, FICO Xpress or HiGHS).

The trade-off is explicit. MIP gives optimal answers but the model grows as O(objects × bins), so Meta's largest problems are too big for any MIP solver. Local search works directly on the graph with an O(objects + bins) neighborhood and evaluates millions of moves per second. Meta uses local search for almost all large problems.

Production numbers: P99 solve time of 12s on 265k objects and 3.2k bins. Problems above 1M objects and 5k bins average 171s, across 3.4k+ runs.

Full analysis: https://www.marktechpost.com/2026/10/06/meta-ai-open-sources-rebalancer-a-c-assignment-solver-that-runs-about-40-million-placement-problems-a-day/

Repo: https://github.com/facebook/rebalancer

Paper: https://www.usenix.org/conference/osdi24/presentation/kumar

Technical details: https://engineering.fb.com/2026/09/21/open-source/rebalancer-generic-high-performance-library-assignment-problems/

Microsoft just released Microsoft-Decision-1, a decision-scoring model that returns calibrated probabilities instead of text. It is post-trained from Qwen3.5-9B and built for routing, classification, verification and agent control.

The core idea is single-pass decision scoring. You give it a situation, a question and a fixed set of options. It returns one calibrated probability per option, so your app can act, defer or escalate in a single call.

In Microsoft’s 36-benchmark comparison (147,137 questions), it scored 83.5% average accuracy, ahead of Quyet-1.0-Large at 81.9%. It ran at 85 ms p50 latency, 35x quicker than GPT-6 Sol. The trade-off: it is text-only, gives no explanations, and its weights are closed. Calibration (92.2) trails Quyet’s 93.1.

At $0.042 per 1M input tokens with free output, it is priced for agent loops that make thousands of small decisions......

Full analysis: https://marktechpost.com/2026/10/09/microsoft-ai-releases-microsoft-decision-1-a-qwen3-5-9b-decision-scoring-model/

Model: https://ai.azure.com/catalog/models/Microsoft-Decision-1

Announcement: https://commandline.microsoft.com/microsoft-decision-1-model-foundry/

OrcaRouter just released OrcaCyber Zero 1.5, a gated frontier cybersecurity model with a 1M-token context window. It reports 100% on Cybench (39/39, unrestricted agent execution) and 95.8% on CVE-Bench (23/24 evaluable tasks).

The core idea is post-training for the full vulnerability lifecycle, not just detection. The model reasons through attack paths, challenges its own hypotheses, and ranks flaws by demonstrable exploitability. Targets include RCE, sandbox escapes, auth bypasses, privilege escalation and chained attacks.

The pricing is the real story: $3.00 / $7.50 per 1M tokens, versus $25 / $125 for Claude Mythos Preview. But there is a trade-off in transparency. Every score is vendor-reported, CVE-Bench used only 24 of 40 tasks, and the 98% CyberGym number belongs to Zero 1.0 inside Orca’s own harness, not 1.5.

With native tool calling, structured outputs and an OpenAI-compatible API, it is built for autonomous security agents. Access is gated to vetted researchers and red teams......

Full analysis: https://www.marktechpost.com/2026/10/10/orcarouter-releases-orcacyber-zero-1-5-cybersecurity-model-with-1m-context/

Model: orcarouter.ai/models/orca/orcacyber-zero-1.5

[ Just Released- Cloudflare Clef-omni...]

Most multimodal agent pipelines still look like this: transcribe the audio, caption the video, then ask a text model to decide. That’s 3 models and 3 failure points for 1 decision. Cloudflare just collapsed it into a single call.

They released Clef-omni, an open-weight decision model built on Qwen3-Omni-30B-A3B (30B total, 3B active). It reads text, images, audio and video in one prefill pass. It returns a probability for every allowed answer instead of generating text.

Here’s what’s actually interesting:

→ No ASR or captioning stage. Audio and video frames are synced into one sequence and scored jointly.
→ Each option gathers its own evidence from the input, then fields cross-attend before scoring.
→ About 130 ms for text decisions, and about 1.5 s for a 21-second video with sound.
→ 97.7 macro-F1 on CLINC150+OOS and 94.8 on BANKING77, both ahead of Jev.
→ $0.15 per M input tokens, no output charge, Apache-2.0 weights.

Cloudflare also cut Clef-flash to $0.038 per M tokens, now under Jev’s $0.042. Hosted Clef also got up to 2.0x faster after a move to SGLang.

Full analysis: https://www.marktechpost.com/2026/10/11/cloudflare-clef-omni-multimodal-decision-model/

Model weights: https://huggingface.co/Cloudflare/clef-omni

Technical details: https://blog.cloudflare.com/clef-faster-cheaper-multimodal/

Datalab just released OmniParseBench, an open OCR benchmark that turns document parsing into 16,288 yes/no questions. It covers 2,937 pages from 2,343 documents across 92 languages.

The core idea is simple: every test is 1 pass/fail claim about 1 page. Is this number present? Does heading A come before paragraph B? Does this cell sit under the "Q3" column? The headline score is just the mean pass rate across 3 families: text, tables and layout.

What makes it different is how you can slice it. Every test is tagged by structure, rendering, role, script and language. Want your parser's score on handwritten tables, or on Thai text? Filter the tags. Every test also records how its answer was established, and no model judges the outputs.

The launch results are tight. Gemini 3.8 Flash leads at 92.9, with Datalab accurate at 92.7. Claude Sonnet 5.5 posts the best text (93.1) and table (95.4) scores, but a 77.5 on layout drags its headline to 88.7.

Code is Apache 2.0. Scoring runs offline with no API keys......

Full analysis: https://www.marktechpost.com/2026/10/14/datalab-omniparsebench-open-ocr-document-parsing-benchmark/

Code: https://aicampaign.live/omnipar

Dataset: https://huggingface.co/datasets/datalab-to/omni_parse_bench

Blog: https://www.datalab.to/blog/omni-parse-bench

NVIDIA Releases Nemotron 3.5 Super VL: A 120B Open MoE With 12B Active Parameters That Sees, Reasons and Runs on a Single DGX Spark.

No separate VLM. No second model for documents. 1 checkpoint.

Here’s how it works. 👇

(1) Sparse by design
120B total parameters, 12B active per token. 512 routed experts per MoE layer, interleaved with Mamba-2 and attention layers.
→ Big-model capacity, small-model compute

(2) Vision built in
A C-RADIO encoder feeds images, PDF pages and video frames into the same backbone that reasons and calls tools.
→ 79.88 CharXiv reasoning, 62.82 MMLongBench-Doc

(3) Speed in the weights
Native multi-token prediction drafts several tokens and verifies them in 1 pass. On DGX Spark, the LiLiCorr drafter adds more.
→ Up to 2.3x faster token generation

(4) Small beats big
Distilled from Nemotron 3 Ultra, then beats it on agent harness evals.
→ Hermes Bench: 63.44% vs 56.33%
→ LangChain evals: 87.05% vs 84.3%

(5) The numbers (NVFP4)
→ 89.90 GPQA-Diamond
→ 77.5 BrowseComp
→ 71.07 SWE-bench Verified
→ 17.78 tau3-bench banking (the weak spot)

Full analysis: https://marktechpost.com/2026/10/15/nvidia-nemotron-3-5-super-vl-release/

Model: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16

Technical details: https://blogs.nvidia.com/blog/nemotron-3-5-super-vl/

Perplexity just open-sourced SPACE: the Firecracker microVM sandbox that runs every task in Perplexity Computer. Apache 2.0, written in Rust.

Each agent gets root inside its own VM. Credentials and egress never live in there.

Here's how it works. 👇

(1) One microVM per task Every agent task runs in its own Firecracker microVM with its own guest kernel. The agent can run commands, install packages and call services as root. The host stays outside its control.

(2) Secrets stay outside the guest Credentials are injected at the network layer, not stored in the VM. A per-sandbox gateway on the host enforces egress policy on every outbound connection.

(3) Built for long-running agents 4 states: running, paused, suspended, deleted. Suspend snapshots memory, runtime state and root filesystem, then releases the VM.
→ btrfs copy-on-write root filesystems
→ Warm pools on Kubernetes, base snapshots on a single node
→ E2B SDK compatibility wrapper

(4) They red-teamed it with 9 frontier models 216 runs, root inside the guest, 2 tasks, 4 settings.
→ VM-to-host escape: 0 of 108
→ Egress bypass with no network: 0 of 54
→ Egress bypass with package-repo access: 11 of 54 before fixes

The key takeaway: VM isolation and network confinement are 2 separate boundaries. Test both.

Full analysis: https://marktechpost.com/2026/10/16/perplexity-open-sources-space-firecracker-microvm-sandbox-ai-agents/

Repo: https://github.com/perplexityai/SPACE?tab=Apache-2.0-1-ov-file

Google just open-sourced Harness², changing how AI agents get better: instead of retraining the model, the agent rewrites its own harness.

The key innovation is a contrastive Propose-Probe-Compose step. It proposes harness edits, probes them on the same task, and keeps only the changes backed by behavioral evidence. No ground-truth labels or verifiers needed.

A second level of recursion saves adopted harnesses and lessons to a library, so each new task starts from accumulated experience.

One refinement step lifts GLM-5 by 11.5 points on WorkBuddy-Bench and 11.8 pp on Harvey's legal benchmark (LAB), improving all 4 tested model/harness setups across all 3 benchmarks. A weaker GLM-5 with a better harness even beats base Gemini 3.5 Flash. It is a trade-off though: each task costs about 4 agent runs, and Meta-Harness still leads LAB on Gemini under Codex.

Built around 8 editable components (system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins), it ships under Apache-2.0 with OpenCode and Codex executors......

Full analysis: https://www.marktechpost.com/2026/10/18/google-harness2-recursive-agent-harnessing-framework/

Repo: https://github.com/google-research/harness2

Paper: https://arxiv.org/abs/2610.14224

▲ Mistral just released Timestral, its first time series foundation models, and the headline is size: 209M parameters, #1 zero-shot on 3 benchmarks, and no GPU needed.

The key innovation is Decoupled Horizon Forecasting (DHF). Instead of masking the values to forecast, learned forecast tokens read the full history. Every observed value works as both target and context.

Covariate-aware attention gives each series a role: target, past-only or known-future covariate. Information flows only from covariates to targets, so a weather forecast or holiday calendar can help without leaking a target's own future.

Timestral-209M scores 0.658 MASE on GIFT-Eval, 0.626 on TIME and 50.39 SQL skill on fev-bench. The 28M model beats Chronos-2, about 4x larger, on GIFT-Eval and TIME. It is a trade-off though: the lead is thin (0.658 vs 0.661 over t0-gamma), and results stay self-reported until the technical report lands.

4 sizes (5M to 209M), 32,768-step context, 21 quantiles, 10 ms to 140 ms on an 8-thread CPU. Weights ship under a Modified MIT license, while TimesFM-3 weights are non-commercial......

Full analysis: https://www.marktechpost.com/2026/10/18/mistral-timestral-time-series-foundation-model/

Repo: https://github.com/mistralai/timestral

Models: https://huggingface.co/collections/mistralai/timestral

▲ MiniMax just released M3.1 Flash, pushing frontier-level agent performance down to Flash pricing: $0.15 / $0.60 per 1M tokens on a 428B MoE with about 23B activated parameters. Space Bunny, the stealth model that processed 200T+ tokens, was an earlier version of it.

🐰 Yes, Space Bunny was on M3.1 Flash from its development...

The core design choice is always-on reasoning with 5 effort levels, from Low to Max. Thinking cannot be disabled, so M3 integrations that send thinking.type=disabled now return an error.

On MiniMax's own table, it scores 62.2 on Automation-Bench v1.0.6, ahead of Opus 5 (50.3) and GPT-5.6 Sol (45.8), and ties Opus 5 on BankerToolBench at 76.2. Every effort level stays below $0.05 per task. It is a trade-off though: it trails Opus 5 by 13.7 points on Terminal Bench 4.0, and DeepSeek V4.1 Flash-1017 still beats it on DeepSWE (77.0 vs 73.7).

With a 1M context window, text, image and video input, and a flat price at half of M3's per-token rate, it targets high-volume coding and office agents. Open weights arrive Oct 23......

Full analysis: https://www.marktechpost.com/2026/10/19/minimax-m3-1-flash-release/

API: platform.minimax.io

Try it: agent.minimax.io/tools/m3-1-flash-preview

▲ Inception just released Mercury 3, a diffusion LLM that writes whole answers in parallel instead of token by token.

The key idea is diffusion for text. The model starts from a rough draft of the full answer and refines every position over a few passes. Fewer sequential steps means far more tokens per second on the same GPUs.

Inception reports 1,068 tok/s vs ~130 for GPT-6 Luna and ~240 for Claude Haiku 5.5. On its own table (medium settings), Mercury 3 leads 8 of 12 benchmarks, including SWE-Bench Atlas 51.08 vs 33.33 for Luna. It is a trade-off though: Haiku 5.5 still wins DeepSWE v1 (51.62 vs 46.31) and MRCR v2, and every score is self-reported.

1M context, 40x+ the size of Mercury 1, OpenAI-compatible API. $0.10 / $0.35 per 1M tokens at launch, $0.20 / $0.70 list......

Full analysis: https://www.marktechpost.com/2026/10/19/inception-mercury-3-diffusion-llm/

Blog: inceptionlabs.ai/blog/introducing-mercury-3

Docs: docs.inceptionlabs.ai

API: platform.inceptionlabs.ai

@_inception_ai @StefanoErmon