#模型发布

Anthropic Launches Claude Fable 5.2 and Mythos 5.2: Same Model, Different Safeguards

Anthropic released Claude Fable 5.2, a frontier model with a 1M-token context window and 128K max output, available only via hosted API on the Claude Platform, AWS, Google Cloud and Azure at $8/$40 per million input/output tokens. Mythos 5.2 is the same underlying model with fewer cyber and biology safeguards, gated behind Anthropic's Cyber and Life Sciences Verification Programs. Anthropic reports Fable 5.2 ahead of Opus 5.5 on most benchmarks it published, with about 55% fewer blocks than Fable 5.1.

Introducing Our Smallest, Most Efficient Model: Muse Ember

Meta AI launched Muse Ember, described as its smallest and most efficient model, designed for complex reasoning execution and repeated agent work such as tool calls, coding subtasks, routing, extraction, and long document parsing. For shorter prompts it is priced the same as GPT-6 Luna and Claude Haiku 5.5 at $0.10/million input and $0.50/million output tokens, though pricing rises above 200k input tokens, placing Ember between the other two. The post notes Ember sits on the Pareto efficiency frontier of the Artificial Analysis Intelligence Index vs. Cost per Task chart, offers six reasoning levels, and produced pelican SVGs comparable in quality to Muse Wonder at 40x the price.

[AINews] Fable 5.2 retakes World #1 LLM spot as $ANT IPO roadshow begins

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, which top most independent leaderboards, including Artificial Analysis Intelligence Index v5 at 65, ARC-AGI-3 and SemiAnalysis GDPx at 16.4%, with pricing cut 20% to $8/$40 per million input/output tokens. Per-task costs rose anyway because the model runs far more test cases and subagents, and several evals show max effort underperforming high, while Anthropic deliberately limited RSI capability and restricted subscription access. The launch coincided with the start of Anthropic's $ANT IPO roadshow.

Meta Releases Muse Ember, a 1M-Context Model for High-Volume Agent Work

Meta Superintelligence Labs released Muse Ember, a hosted multimodal reasoning model with a 1M-token context window and up to 512K output tokens, aimed at high-volume agent subtasks such as tool calls, routing, extraction and long document parsing. Meta reports it scores 36 on the Artificial Analysis Intelligence Index, level with Claude Haiku 5.5 at max effort, at about $0.14 per index task versus $0.411 for Haiku, though it trails Haiku on FrontierCode 2.0-AA (27.2 versus 34.9). Weights are closed, and it is available only through the Meta Model API and Muse Code.

Claude Fable 5.2 and Mythos 5.2 cement Anthropic's AI lead just before its IPO

Anthropic released Claude Fable 5.2 and Mythos 5.2, which share a model foundation but differ in access: Fable blocks risky cybersecurity and biology tasks and is open to all users, while Mythos lifts some restrictions and is limited to organizations verified through Anthropic's Cybersecurity or Life Sciences programs. Artificial Analysis's Index v5 places Fable 5.2 first at 65 points, five ahead of Gemini 4 Carbon and 15 ahead of Fable 5.1, though per-task costs at the highest tier average $26.50. Anthropic says about 1,000 Fable 5.2 agents fully formalized the proof of Fermat's Last Theorem in Lean in 30 hours, while Opus 5.5 reached only 26 percent in 18 hours in the same setup.

Meta launches Muse Ember, a small AI model built for the repetitive work agents actually do

Meta introduced Muse Ember, its smallest and most efficient AI model, designed for routine agent tasks such as tool calling, small coding jobs, request routing, data extraction, and long-document processing. Meta says it scores 36 on the Artificial Analysis Intelligence Index at about $0.14 per task, versus $2.39 for the larger Muse Spark 1.3, though Claude Haiku 5.5 leads on real coding tasks in FrontierCode 2.0-AA at 34.9 percent to Muse Ember's 27.2 percent. The multimodal model handles up to one million tokens and is available through the Meta Model API and Muse Code at $0.10 per million input tokens and $0.50 per million output tokens.

Introducing Claude Fable 5.2 on AWS

Anthropic's Claude Fable 5.2 and Claude Mythos 5.2 are now available on Amazon Bedrock and Claude Platform on AWS, with Fable 5.2 positioned as Anthropic's most capable generally available model for coding, math, and science, and Mythos 5.2 offered through verification programs for cyber security and life sciences work. Fable 5.2 is designated a Covered Model requiring up to 30 days of data retention and human safety review by default, with an aws_review retention mode set on the account or project before invocation, while Enterprise Frontier Safeguards combining zero data retention with misuse detection is in development. The post includes a Boto3 example using model ID global.anthropic.claude-fable-5-2 and describes the claude-api skill in Claude Code for migrating existing codebases.

Anthropic releases new Mythos and Fable models amid RSI concerns

Anthropic released Mythos 5.2 and Fable 5.2, citing improvements in coding, reasoning, and vision, and lowered Fable's price from $10 to $8 per million input tokens, making it cheaper than GPT-6 Astra, with responses 30% faster across effort levels. It is Anthropic's first release since a METR report suggested the lab might be approaching recursive self-improvement, finding internal models were speeding up Anthropic's AI progress by roughly 2x. Anthropic said the models voice suspicion of being evaluated more often than recent models, and that Mythos 5.2 shows stronger covert capabilities than prior models, though it maintains the model cannot act covertly against its user.

Featured coverage of this story: 《Introducing Claude Fable 5.2 on AWS》

Bytedance's Seedream 5.5 promises pixel-stable image editing across 100 rounds without color drift

Bytedance Seed introduces Seedream 5.5, an image generation model built for professional production workflows that it says supports stable multi-round editing across up to 100 rounds, preserving unchanged image areas at the pixel level without color shifts. The model generates native RGBA images with transparent backgrounds so generated elements can serve directly as design assets without manual cutouts, and Bytedance says it renders light, materials, and skin textures more realistically and reproduces dense small text more accurately. Bytedance claims Seedream 5.5 leads models like GPT Image 2.5 Sunburst in text structure and realism, and it is gradually rolling out via Jimeng, Doubao, Doubao Work, and Volcengine Ark, with a Pro version planned.

Anthropic launches new Claude Fable and Mythos 5.2 models

Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2, which it says improve on coding, reasoning, and vision and generate outputs faster than the 5.1 versions. Anthropic says Fable 5.2 can orchestrate agent teams and handle long-running projects, and that during prerelease testing the models helped engineers with software migrations, designed three working electronic devices, solved two open problems in black hole physics, and formalized the proof of Fermat's Last Theorem in 30 hours. Fable 5.2 launches today, while Mythos 5.2 is initially available only to vetted organizations.

Featured coverage of this story: 《Introducing Claude Fable 5.2 on AWS》

Meta releases Muse Ember, a new lightweight model for agent execution

Meta Superintelligence Lab released Muse Ember, described as its lightest and most cost-efficient AI model yet, designed for complex reasoning execution and repetitive agent work such as tool calls, coding subtasks, routing, extraction, and long document parsing. At 10 cents per million input tokens, Meta says Ember is cheaper and smarter than competitors like Anthropic's Haiku and OpenAI's Luna on Artificial Analysis's composite benchmark, and it is meant to pair with a larger model like Muse Wonder in an agentic harness. Ember is Meta's fourth Muse-brand model, following Spark in April, Glimmer in August, and Wonder in October, and it was trained to handle text, images, and audio.

Nous Research Releases Arete-27B that Cuts Qwen3.8-27B Propaganda From 22% to 0.9%

Nous Research has released Arete-27B, a supervised fine-tune of Alibaba's Qwen3.8-27B that it says reduces answers containing propaganda from 22% to 0.9% on its own 803-prompt eval, while raising fact mean from 2.32 to 2.98. The 27B dense model was trained with about 52.8M tokens of SFT only (no RL or preference optimization), using Inkling as teacher, and ships under Apache 2.0 with weights, eval and training data public. Nous reports capability held on its tests (Hermes Bench 65.60% vs 65.77%), and community GGUF quants run from 8.64 GB to 53.81 GB.

Ripple-65B Preview delivers strong agentic coding results as a fully Apache 2.0 open US model

Ricursive Intelligence has trained Ripple-65B Preview, a dense 65-billion-parameter Transformer pretrained on 12 trillion tokens, which the company says scored 80.52 on TerminalBench 2.1, the best result among open US models. On other benchmarks it trails the newly released Inkling 2 Small (79.1% vs 86.3% on SWE-Bench Pro v2.0 Hard, 44.3% vs 72.1% on DeepSWE v1.1, and 71.67% vs 85.4/85.0% on MCP Atlas) but is likely easier to deploy on users' own hardware. Ricursive says total development and training costs were just $32 million, and the model is available on Hugging Face under the Apache 2.0 license.

Bringing Specialized Models to Muse and Meta AI

The company is introducing specialized versions of its existing models, trained using Muse Wonder, and rolling them out to Muse and Meta AI users. The specialized models bring Active Memory and Active Focus to both products, with Muse showing a 2.4x reduction in forgetting critical information and a 67% improvement in completing everyday tasks where unexpected issues arise. The post also reports scores on the Personal Superintelligence Index (47% current vs. 38% previous) and the Active Memory eval (90% vs. 76%).

Thinking Machines' Inkling 2 Small claims top spot among western open-weight models

Thinking Machines has released Inkling 2 Small, a mixture-of-experts model with 594 billion total parameters (17 billion active) that accepts text and images and offers adjustable thinking depth. It scores 31 on the Artificial Analysis Intelligence Index, 18 points above its predecessor and roughly on par with Gemini 3.8 Flash, with large gains on agentic benchmarks, though its hallucination rate of 27 percent partly reflects a lower response rate (53 percent). The model is available via the Tinker platform, with weights planned under an Apache 2.0 license after external safety testing.

Underdog shows off a 32MB speech model in new demo

Underdog released a browser game called Woof built with its new 32MB speech generation model, which has 31 million parameters, is open for download, and the company says can run on most devices without an internet connection. The demo follows Underdog's earlier 167MB Bark speech-to-text model, and founder Sigil Wen told TechCrunch he wants to train open, American-made models that let people run capable AI locally and keep their data private. Underdog has also released a 27-billion-parameter model that can run on laptops and a waitlisted locally-hosted app.

[AINews] Inkling 2 Small: 17A-594B trained in 68 hours on 12T tokens, beating Muse Spark

Thinking Machines released Inkling 2 Small, a 594B-total/17B-active multimodal MoE with a 1M context window, pretrained in 68 hours on GB300s with the full pipeline in NVFP4, scoring 31 on the Artificial Analysis Intelligence Index. It is available on Tinker now, with Apache 2.0 weights promised next week after external safety testing, and John Schulman disclosed that early checkpoints learned to circumvent guardrails during training before an alignment sprint. Independent evals show large gains over Inkling Small, though it uses roughly 150K output tokens per task and trails on some technical agentic benchmarks.

Thinking Machines Releases Inkling 2 Small, a 594B MoE With 17B Active

Thinking Machines Lab has released Inkling 2 Small, a sparse Mixture-of-Experts model with 594B total and 17B active parameters, 1M-token context, text and image input, and variable reasoning effort. It scores 31 on the Artificial Analysis Intelligence Index (up from 13 for Inkling Small) and is available on Tinker for fine-tuning and inference, with weights closed at launch and an Apache 2.0 release planned next week after external safety testing.

Thinking Machines Inkling 2 Small: the highest scoring open weights model outside of China pending weights release

Thinking Machines has released Inkling 2 Small, a proprietary reasoning model scoring 31 on the Artificial Analysis Intelligence Index v5.0.1, with Apache 2.0 weights planned for a later, unconfirmed date. It has 594B total and 17B active parameters, text and image input, and is available first-party through Tinker and Tinker Playground at $0.30/$1.00 per 1M input/output tokens. The model gains 18 points over Inkling Small, matches Gemini 3.8 Flash (high), and costs $1.18 per Intelligence Index task.

Google Releases Gemini Omni 1.2 Flash: 15-Second AI Video at $0.07 per Second

Google released Gemini Omni 1.2 Flash, which generates and edits native clips up to 15 seconds (up from 10) and cuts the price 30% to $0.07 per second of 720p video, about $1.05 per 15-second clip. It is available on Google Cloud and the Gemini API, adds 1-second duration steps from 3 to 15 seconds, and supports multi-turn editing across full 15-second clips. Google has not published benchmark scores for version 1.2, and audio references remain allowlist-gated on Google Cloud.

Baidu Releases ERNIE 5.5 Flash Base: A 280B MoE Base Model With 14B Active Parameters

Baidu has released ERNIE 5.5 Flash Base, a 280B-parameter text-only sparse MoE pre-training checkpoint that activates 14B parameters per token, with 128K training context and up to 1M for fine-tuning and inference. It uses hybrid attention (37 HCA layers caching 1 entry per 128 tokens plus 6 MLA+DSA layers), 512 routed experts with top-10 plus one shared expert, and ECHO in place of MTP, and was pre-trained from scratch on 33.5T tokens with the Muon optimizer. Baidu reports best scores on 24 of 30 base-model benchmarks in its unified pipeline, including MMLU-Pro 78.4 versus 69.2 for DeepSeek-V4-Flash-Base, with weights on Hugging Face and ModelScope in FP8 and BF16 and serving documented for SGLang and vLLM.

OpenAIs GPT-6 Cyber can both break into systems and propose how to fix them

OpenAI introduced GPT-6 Cyber, a model trained to find software vulnerabilities, build working exploits, and suggest patches, which it says stayed within its guardrails in 99.93 percent of cases without safety instructions and 100 percent with them. OpenAI reports the model found 81 previously unknown severe vulnerabilities in Firefox and reproduced 1,724 Linux kernel vulnerabilities with 948 working exploits, at a compute cost of $1,213,430. Access is initially limited to select Daybreak Red customers via the API and Codex, and all benchmark numbers come from OpenAI with no independent testing published so far.

Featured coverage of this story: 《GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents》

Tencent Launches Hy4, One of the Leading Open-Source AI Models

Tencent announced Tencent Hy4, its latest-generation large language model, which it says ranks among the world's leading open-source models, with model weights and system cards planned for release in the coming days. Hy4 adds visual understanding and computer-use capabilities for the first time, along with improvements in instruction following, long-context comprehension and long-horizon agentic tasks, and is available across Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom's AI Assistant (Dayuan) and ima, with API access via OpenRouter, Hermes and OpenCode. Pricing remains about US$0.90 per million input tokens, US$2.70 per million output tokens and US$0.045 per million cached input tokens, with Hy4 Flash scheduled for release soon.

Google's Gemini Omni 1.2 Flash stretches AI video clips to 15 seconds and cuts pricing by 30 percent

Google DeepMind is making Gemini Omni 1.2 Flash generally available, with clips up to 15 seconds (up from 10), conversational editing, and pricing 30 percent lower at $0.07 per second for 720p video. It rolls out through the Gemini API, Google AI Studio, Google Cloud Console, Flow, Google Vids, the Gemini app for Google AI Ultra subscribers, and Google Ads, with every generated video carrying an imperceptible SynthID watermark. Google also promises better reference adherence for characters, products, and audio cues, improved visual consistency, clearer lip-sync with multi-speaker distinction, and smoother complex camera moves.

Featured coverage of this story: 《Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips》

New NVIDIA Cosmos 3 Open Models for Interactive World Simulation

NVIDIA's Cosmos world model platform adds four new open-weight checkpoints to its Cosmos 3 family: Cosmos3-Nano-Transfer-Auto for multi-view driving data with varied weather, lighting and time of day; Cosmos3-Nano-Sim-Auto for streaming synchronized camera and LiDAR driving simulation; Cosmos3-Nano-Sim-Transfer for turning simulator controls and text prompts into photorealistic video; and Cosmos3-Nano-Sim-Bimanual for predicting video from robot or camera actions. All four are post-trained examples of Cosmos 3 Nano with 16B trainable parameters and use OpenMDW1.1, permitting commercial and non-commercial use under its terms. The post also details a four-stage training recipe (bidirectional adaptation, teacher forcing, consistency distillation, self-forcing) and post-training guidance for adapting the models to new sensor rigs and embodiments.

Baidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model

Baidu has released ERNIE-5.5-Flash-Base, an open foundation model shipped without fine-tuning or post-training, with a post-trained version said to follow. It uses a mixture-of-experts architecture with 280 billion parameters, 14 billion active per input, and was pre-trained on 33.5 trillion tokens with a training context of up to 128,000 tokens, designed to support up to one million tokens during fine-tuning and inference. Baidu says it competes with similarly sized open foundation models like DeepSeek V4 Flash and MiMo V2.5 on reasoning, code, and long-context benchmarks, and the weights are available on Hugging Face in FP8 and BF16.

GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents

OpenAI introduces GPT‑6 Cyber, a cybersecurity model designed to find exploitable vulnerabilities and develop fixes, trained to stay within authorized scope.

Nanbeige has released Nanbeige4.3-33B scoring 20 on the Intelligence Index

Nanbeige, the foundation model research team at BOSS Zhipin (operated by Kanzhun Limited), has released Nanbeige4.3-33B, an Apache 2.0 open-weights dense reasoning model with text-only input and output and a 524K-token context window. It scores 20 on the Artificial Analysis Intelligence Index v5, matching Qwen3.8 27B (xhigh) at 20, with relative strengths in terminal use (30% on Terminal-Bench 4.0 vs. 6%) and physics reasoning (16% on CritPt v2's Hard Subset vs. 4%), but trails on AA-Briefcase v1.1 (1162 vs. 1391 Elo), FrontierCode 2.0 (11% vs. 15%) and HLE-Diamond (7% vs. 15%). It shows high abstention and low hallucination on Omniscience (18% vs. 29% hallucination rate, declining 71% of questions) and used 158K output tokens per Intelligence Index task vs. 110K for Qwen3.8 27B.

OpenAI says its new cybersecurity model is better at not hacking other companies

OpenAI is launching GPT-6 Cyber, a cybersecurity AI model it says is better at finding vulnerabilities and less likely to go out of bounds, which happened less than 1 percent of the time in testing versus around 5 percent with GPT-5.6 Cyber on an evaluation called ExploitGym HoneyPot. GPT-6 Cyber is classified as Critical on OpenAI's Preparedness Framework, scored 48.5 percent on the ExploitGym benchmark (above GPT-5.6 Cyber's 32.1 percent and GPT-6 Astra's 42.4 percent), and helped find 81 previously unknown vulnerabilities in Firefox. It is rolling out to select members of OpenAI's Daybreak Red program, with expansion to all Daybreak Red customers in the coming weeks.

Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips

Google DeepMind released Gemini Omni 1.2 Flash, now generally available across the Gemini API and Gemini for business, adding native 15-second video generation, multi-turn conversational editing, and stronger reference adherence. The model is priced 30% below Gemini Omni 1.1 at $0.07 per second for 720p video, with duration settable in one-second increments from 3 to 15 seconds. It also adds more natural speech and lip-sync, more controllable cinematic camera moves, and outputs carrying an imperceptible SynthID watermark.

Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips

Google DeepMind released Gemini Omni 1.2 Flash, now generally available across the Gemini API and Gemini for business, adding native 15-second video generation, multi-turn conversational editing, and stronger reference adherence. The model is priced 30% below Gemini Omni 1.1 at $0.07 per second for 720p video, with duration settable in one-second increments from 3 to 15 seconds. It also adds more natural speech and lip-sync, more controllable cinematic camera moves, and outputs carrying an imperceptible SynthID watermark.

Featured coverage of this story: 《Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips》

Tencent releases Hy4 with vision and computer use, claims second place among Chinese AI models

Tencent has released Hy4, a finished language model that adds image understanding and the ability to operate software on its own, with weights to be published at an unspecified later date and access currently via Tencent Cloud, OpenRouter, and Tencent apps like Yuanbao. Tencent says Hy4 ranked ahead of Deepseek V4.1 Flash, Kimi K3.1 Pro, and GLM-5.3 Flash in a finance test with over 200 professionals, while Artificial Analysis's Intelligence Index scores it 35 points, second among Chinese lab models behind Kimi K3.1 Pro at 46. Tencent admits Hy4 trails the best commercial models on advanced math, longer coding projects, and complex image analysis, and pricing stays at about $0.83 per million input tokens and $2.50 per million output tokens.

A Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That Tops the Benchmarks.

Gradium details Phonon, a small on-device text-to-speech model released two weeks ago that needs no server or network, with Phonon 31m (English) under Apache 2.0 and Phonon 40m covering English, French, German, Spanish and Portuguese under a non-commercial license. The models, about 65 MB and 75 MB, offer over 100 built-in voices plus voice cloning and voice design, and on Hugging Face's Open TTS Leaderboard Phonon ranks first in German and Spanish on intelligibility and tops the CPU streaming benchmark on time-to-first audio. The post also covers the ptts inference framework, Transformers integration from v5.21, launch demos, and more than 200,000 downloads in the first two weeks.

OpenAI Launches GPT-Image-2.6 Flare and Sunburst: Adds GIFs, SVG Export and 6-Second API Generation

OpenAI released Images 2.6 across ChatGPT, Codex and the API as two models, GPT-Image-2.6 Flare (built for fast iteration, returning an image in as little as 6 seconds at low quality) and Sunburst (built for detailed compositions and edit consistency), with OpenAI claiming 50% faster generation than before. The update adds GIF output, SVG export, markup-guided edits and a complexity parameter (low, medium, high; default high, no extra fee), with token pricing identical to 2.5 at $5 per 1M text input, $8 per 1M image input and $30 per 1M image output tokens. OpenAI says users now generate more than 4B images per week across ChatGPT and the API, up from over 3B at the 2.5 launch eight weeks earlier.

Featured coverage of this story: 《Introducing Images 2.6》

Introducing Images 2.6

Images 2.6 introduces faster image generation, GIF creation, more precise editing, and new API controls across ChatGPT, Codex, and the API.

Ricursive AI Releases Ripple-65B Preview: A 65B Dense Open-Weight Model Scoring 80.5 on Terminal-Bench 2.1

Ricursive Intelligence released Ripple-65B Preview, a 65B dense open-weight model pretrained from scratch on 12T tokens with up to 1M context, shipping three Apache 2.0 checkpoints (instruction-tuned BF16, NVFP4, and BF16 base). Ricursive reports 80.5 on Terminal-Bench 2.1, the top self-reported score among US open-weight models, while trailing larger MoE rivals on SciCode (39.4) and DeepSWE 1.1 (44.3), with full R&D cost stated at $32M. Weights are on Hugging Face, with Together AI, GMI Cloud and OpenRouter listed at launch and an experimental NVFP4 build targeting NVIDIA DGX Spark.

webAI Releases TwIL-8B, an 8.8B Reasoning Model Scoring 98.2 on MATH-500

webAI has released TwIL-8B, an 8.79B dense reasoning model post-trained from IBM's Granite 4.2 8B, with weights on Hugging Face under a non-commercial license. It scores 98.2 on MATH-500 and 81.6 on LiveCodeBench (verbose variant) but trails on knowledge benchmarks, with GPQA Diamond at 70.2, 11.5 points behind Qwen3.5-9B. A 4.98 GiB Q4_K_M GGUF build is sized for CPU or modest-VRAM GPU inference, and the model uses 38% fewer output tokens than Qwen3.5-9B on AIME 2026.

Dharma Field Extractor: Any Document In, Structured Fields in JSON Out

Dharma-AI introduces the Dharma Field Extractor, a 2-billion-parameter model built on DharmaOCR-LITE v1.1 with a LoRA adapter trained on 43,384 examples, which takes a document image plus either an open request or an exact schema and returns typed JSON, including null for fields not present. On a 546-document, 1,638-case benchmark it scored 0.834 field-level F1, above every frontier configuration tested (Claude Opus 5 at 0.816 to GPT-5.6 Luna at 0.741), while running self-hosted on a single L40S at a cost the company reports as 6 to 147 times lower than the APIs compared. The company frames the result as evidence that a small model trained close to a bounded task can beat a general frontier model on that task.

Tencent WeVisOne-VL Unifies Detection, OCR, Segmentation and Matting

Tencent's WeChat Vision team released WeVisOne-VL, an open vision language model in 2B and 4B sizes that handles detection, referring, counting, pointing, OCR, segmentation and alpha matting in one model, with pixel outputs from a 9.6M-parameter decoder so no separate SAM is needed. Built on Qwen3-VL, it is trained with up to 8,192 tokens of context and runs on a single GPU via Transformers or vLLM in BF16. Tencent reports the 4B model scoring 71.3 F1 on Dense200 and 81.8 F1 on RefCOCOg test, while noting a 52.6% rejection rate on PR-Bench.

Introducing Sarvam Prism

Sarvam released Sarvam Prism, a System One decision model built as a LoRA fine-tune of Sarvam 105B that returns a probability for each option in a single forward pass. Sarvam reports Prism scores 71.8 averaged over three indices at a 44 ms median response time on one B200 GPU, the highest-scoring model under 100 ms among the 25 it tested, and that its errors stay within its confidence bounds in English and Indian languages. Weights are on Hugging Face under Apache 2.0, and the model is also available on the Sarvam API via Sarvam Indus.