Tag: llm

  • Claude Opus 5.5 Launched: 1M Context, Always-On Thinking & 20% Price Cut

    Claude Opus 5.5 Launched: 1M Context, Always-On Thinking & 20% Price Cut

    Anthropic has officially launched Claude Opus 5.5, introducing a high-performance frontier model engineered specifically for long-running agentic coding, deep research, and complex knowledge synthesis. Built with a default 1-million-token context window and 128,000 maximum output tokens, the release pairs sustained reasoning power with an immediate 20% API price reduction compared to its predecessor.

    The launch underscores Anthropic’s commitment to making frontier-grade reasoning accessible for production workflows, dropping token costs to $4 per million input tokens and $20 per million output tokens.

    Architectural Enhancements and Capabilities in Claude Opus 5.5

    A central architectural upgrade in the new flagship is always-on adaptive thinking. Unlike earlier iterations where developers manually passed token budgets or toggled extended reasoning parameters, the system dynamically calibrates internal reasoning traces based on problem complexity, governed via a streamlined effort parameter.

    The model also mandates updated tool interfaces, deprecating legacy computer use scripts in favor of modern standardized computer toolsets. In addition, Anthropic has opened a research preview of fast mode for low-latency reasoning queries, allowing developers to accelerate output token generation on latency-critical enterprise pipelines.

    Feature / MetricClaude Opus 5 (Predecessor)Opus 5.5 ArchitectureDeveloper Advantage
    Input Token Pricing$5.00 / MTok$4.00 / MTok20% Direct Cost Savings
    Output Token Pricing$25.00 / MTok$20.00 / MTok20% Output Expense Cut
    Default Context Window1,000,000 tokens1,000,000 tokensEnterprise-Scale Codebase Ingestion
    Max Output Limit128,000 tokens128,000 tokensLarge-File & Multi-Module Generation
    Thinking ConfigurationOptional / Budget-basedAlways-On Adaptive ThinkingAutomated Complexity Scaling
    Tool OrchestrationStatic Schema ConfigurationMid-Conversation Inline & MCP ToolsZero Cache Invalidation on Schema Updates

    Mid-Conversation Inline Tools and Native MCP Connector

    Beyond raw inference efficiency, the model enables dynamic inline tool definitions within mid-conversation system messages. Developers can now inject, modify, or update tool schemas on the fly without invalidating existing prompt caches, eliminating costly context re-processing during multi-turn agent loops.

    When paired with Anthropic’s Model Context Protocol (MCP) connector, the model records fetched tool definitions in pinned response blocks, guaranteeing deterministic schema conformance across distributed microservices. The rollout of the flagship model across Amazon Bedrock, Google Cloud, and Microsoft Foundry ensures immediate enterprise availability without cloud vendor lock-in.

    By blending reduced inference rates, expansive context limits, and modular tool orchestration, Anthropic delivers a robust engine for autonomous software engineering teams.

    Key Takeaways for Developers

    • 1M Default Context Window: Claude Opus 5.5 ships with a default 1-million-token context window and 128k output tokens for massive codebases.
    • 20% Cost Reduction: Standard API pricing drops to $4 per million input tokens and $20 per million output tokens.
    • Always-On Adaptive Thinking: Eliminates manual token budgets, scaling reasoning depth automatically using effort parameters.
    • Dynamic Inline Tooling: Define and update MCP tool schemas mid-thread without busting prompt cache memory.
  • Qwen3.8-LiveTranslate Released: Real-Time Interleaved Speech Translation Cuts Lag to 2.3 Seconds

    Qwen3.8-LiveTranslate Released: Real-Time Interleaved Speech Translation Cuts Lag to 2.3 Seconds

    Alibaba’s Qwen team has officially released Qwen3.8-LiveTranslate, an advanced multimodal simultaneous interpretation model that listens to live speech and outputs translated audio while the speaker is still talking. By introducing an end-to-end Interleave architecture, the release drops Length-Adaptive Average Lagging (LAAL) from 2.8 seconds down to 2.3 seconds—an 18% reduction in latency across 60 supported languages.

    The breakthrough bridges a critical gap in international communications, allowing global developers and enterprise teams to deploy low-latency, natural voice-to-voice translation pipelines without traditional cascading delays.

    Interleaved Architecture and Latency Reductions in Qwen3.8-LiveTranslate

    Traditional simultaneous translation systems rely on fragmented cascades that pass audio through independent automatic speech recognition (ASR), machine translation (MT), and text-to-speech (TTS) pipelines. The core innovation inside the new system is an Interleave streaming architecture that processes input speech, transcribed text, and target spoken audio as a single synchronized sequence.

    By unifying multimodal tokens in one time-ordered stream, the engine eliminates cross-module pipeline latency. The system understands 60 input languages and provides bidirectional speech output across 29 languages, while the remaining 31 languages receive real-time streaming text translation.

    Capability / MetricTraditional Cascaded PipelineLiveTranslate ArchitectureArchitectural Advantage
    Pipeline ArchitectureDecoupled ASR → MT → TTSUnified Interleaved Multimodal StreamZero cross-engine serialization overhead
    Average Lagging (LAAL)2.8 – 3.5 seconds2.3 seconds~18% faster real-time spoken response
    Speaker DiarizationExternal diarization serviceNative multi-speaker taggingContextual turn tracking in group meetings
    Voice PreservationGeneric synthetic voicesReal-time zero-shot voice cloningMaintains original speaker identity and pitch
    Multimodal Vision InputAudio-only transcriptionUp to 2 frames/second visual contextResolves ambiguous words using lip movements

    Speaker Diarization, Context Awareness, and Cloud Pricing

    In addition to streaming latency cuts, the interpreter introduces real-time speaker diarization and automated voice preservation during multi-party calls. The system dynamically isolates distinct speakers and clones their vocal characteristics into the target spoken output, ensuring listeners immediately recognize who is speaking during multi-person conferences.

    The architecture also incorporates long-context entity disambiguation, referencing historical meeting dialogue to prevent mistranslating proper nouns, brand names, and technical terms. For enterprise developers, the service is deployed as a hosted WebSocket API (livetranslate-flash-realtime) on Alibaba Cloud Model Studio and Qwen Cloud. Spoken input is billed at 7 tokens per second ($7.50/1M tokens) and spoken output at 12.5 tokens per second ($30.00/1M tokens), bringing the operational cost of one full hour of continuous bidirectional translation to approximately $1.54.

    Key Takeaways for Developers

    • Sub-2.5 Second Spoken Latency: Qwen3.8-LiveTranslate reduces average translation lag to 2.3 seconds using a unified Interleave architecture.
    • Broad Multilingual Coverage: Ingests 60 languages with full bidirectional speech output across 29 major global languages.
    • Synchronized Bilingual Streaming: Emits source transcript events alongside target audio, enabling real-time on-screen subtitle displays.
    • Cost-Effective Real-Time API: Delivers production WebSocket endpoints capable of hosting continuous live interpretation for ~$1.54 per hour.
  • Ternary Bonsai 2 27B Launched: 5.9GB Open-Weights LLM Retains 98% Intelligence

    Ternary Bonsai 2 27B Launched: 5.9GB Open-Weights LLM Retains 98% Intelligence

    PrismML has officially released Ternary Bonsai 2 27B, a groundbreaking 5.93 GB open-weights language model that achieves 98.2% benchmark parity with full-precision 27B architectures while slashing memory footprints by 9x. The Apache 2.0-licensed release demonstrates that extreme ternary quantization {-1, 0, +1} can power complex autonomous coding agents and multimodal reasoning directly on consumer laptops and single GPUs.

    The release marks a pivotal moment for on-device computing, proving that developers no longer need dual RTX 4090s or quantized FP8 compromises to execute 27B-class models locally.

    Core Architectural Breakdown of Ternary Bonsai 2 27B

    Built upon Alibaba’s open-weights Qwen 27B architecture, the engineering behind the Bonsai 2 system replaces standard 16-bit floating-point weight representations with ternary {-1, 0, +1} values. This allows matrix multiplication operations to be executed primarily through energy-efficient additions and subtractions rather than compute-heavy floating-point MACs.

    The model occupies only 5.93 GB of disk and VRAM, compared to 53.80 GB in standard FP16. Unlike post-training quantization methods that often degrade coherence in reasoning tasks, PrismML pre-trains and aligns the ternary weights directly with FP16 group scalers, enabling unprecedented intelligence retention across multi-step tasks.

    SpecificationStandard FP16 Baseline (27B)Bonsai 2 (Ternary Quantized)Efficiency Multiplier
    Model Size / VRAM Footprint53.80 GB5.93 GB9.07x Reduction
    Arithmetic Precision16-bit Floating Point1.58-bit Ternary {-1, 0, +1}~10x Parameter Compression
    Context Window128K tokens262K tokens2x Context Window
    Hardware Minimum2x 32GB GPUs or Cloud Pod16GB Laptop / Single 24GB GPUEdge-Deployable
    LicensePermissive Open WeightsApache 2.0Full Commercial Use

    Benchmark Parity and Agentic Coding Performance

    Across a comprehensive suite of 20 recognized evaluation benchmarks, the model retains 98.2% of its parent model’s unquantized scores. In agentic code generation and long-horizon tool orchestration, the model successfully drove autonomous coding workflows inside Cline and native terminal shells without degrading into syntax loops.

    The architecture also retains native multimodal capabilities, allowing local developers to process image inputs, architectural diagrams, and terminal snapshots within an expansive 262,144-token context window. The engineering team has published optimized GGUF binaries for standard llama.cpp runtimes alongside dedicated 2-bit MLX kernels tailored for Apple Silicon Unified Memory architectures.

    Open-weights releases of this caliber demonstrate that the memory wall in consumer AI is crumbling. By combining ternary representations with scalable post-training techniques, developers can now run production-grade autonomous agent loops entirely air-gapped on existing hardware.

    Key Takeaways for Developers

    • Massive Footprint Compression: Ternary Bonsai 2 27B shrinks a 27B-parameter architecture to 5.93 GB, fitting comfortably within 16 GB MacBook Unified Memory or an RTX 4070.
    • Near-Zero Precision Loss: Delivers 98.2% average benchmark retention across 20 evaluations, maintaining reasoning integrity in math and coding benchmarks.
    • Expanded Native Context: Features a 262K-token context window, supporting full-codebase ingestion and persistent agentic memory.
    • Immediate Ecosystem Integration: Apache 2.0 weights are available immediately via Hugging Face with native support for llama.cpp and MLX.

  • TypeSafe Jev AI Launched: Decision-Only Model Delivers 70ms Latency & Free Output Tokens

    TypeSafe Jev AI Launched: Decision-Only Model Delivers 70ms Latency & Free Output Tokens

    Published by AICodeNews Editorial Team | September 17, 2026

    Abandoning conversational text generation in favor of fast, structured software decisions, TypeSafe Jev AI has officially been unveiled by TypeSafe AI and InstructGPT co-author Diogo Almeida.

    Announced across official developer channels on typesafe.ai and shared by founder Diogo Almeida (@CompleteSkeptic), TypeSafe Jev AI introduces a new model class described as “System One” AI. Instead of generating open-ended prose word-by-word, the model evaluates state data against pre-declared schemas, delivering typed decisions and calibrated probabilities in 70 to 500 milliseconds at a disruptive cost of $0.042 per million input tokens with completely free output.

    1. What Is TypeSafe Jev AI?

    Most production AI workflows inside software engineering do not actually need an essay; they need an instant answer to a narrow, high-frequency question: Which support queue gets this ticket? Is this database mutation safe? Should this user invoice require human review? Rather than generating paragraphs or code snippets, TypeSafe Jev AI takes two inputs: a raw state context and typed questions:

    • Zero Schema Hallucination Guarantee: Because candidate outputs are strictly bound to pre-defined enums and types, the model mathematically cannot emit syntax outside your application’s schema.
    • Native Probability Calibration: Every decision returns confidence scores directly sampled at the model layer, allowing engineering pipelines to send low-confidence edge cases to human review automatically.
    • High-Frequency Real-Time Execution: In live demonstrations, developers ran Jev at 10 decisions per second (10 Hz) playing DOOM in real-time—executing tactical dodging and weapon switching for an entire hour for just $7.

    2. Standard LLM Generation vs. TypeSafe Jev AI Architecture

    The architectural trade-off replaces open-ended chatting with rapid programmatic evaluation:

    Architecture DimensionStandard Generative LLMs (e.g. GPT-5.6 / Claude)TypeSafe Jev AI
    Output ModeAuto-regressive token generation (Text/JSON)Direct typed decision sampling
    Inference Latency800ms – 3,500ms70ms – 500ms (Up to 193x faster)
    Input Pricing / 1M$2.00 – $10.00$0.042 (440x cheaper)
    Output Pricing / 1M$10.00 – $50.00$0.00 (Completely Free Output)
    Hallucination RiskFrequent string/schema format errorsZero structural hallucination

    3. Developer Integration & Real-World Use Cases

    The economic disruption of TypeSafe Jev AI is driven by its pricing structure: $0.042 per million input tokens with zero charge for output. The model is accessible via official Python and TypeScript REST SDKs, making it a drop-in replacement for triage routers, fraud classification gates, and autonomous agent routing layers.

    4. Key Takeaways on TypeSafe Jev AI

    • Decision-Only Shift: TypeSafe Jev AI sacrifices conversational text generation to achieve instant, typed software decisions.
    • Disruptive Token Economics: $0.042/1M input pricing and free output make high-frequency background agent checks economically negligible.
    • Sub-500ms Reliability: Delivers calibrated probabilities with zero structural hallucination for automated software pipelines.

    Bookmark AICodeNews.com for daily updates on novel AI architectures, developer models, and inference economics.

  • DeepSeek V4.1 Flash Launches in Beta: Native Multimodal Vision & 420 Tokens/Sec Output

    DeepSeek V4.1 Flash Launches in Beta: Native Multimodal Vision & 420 Tokens/Sec Output

    Published by AICodeNews Editorial Team | September 8, 2026

    In a surprise developer preview, DeepSeek V4.1 Flash has officially launched across the DeepSeek API platform under the intermediate test endpoint deepseek-v4.1-flash-expires-on-0910.

    Released as an invite-and-feedback evaluation preview on the official DeepSeek API Platform, DeepSeek V4.1 Flash introduces a revamped Mixture-of-Experts (MoE) architecture with native multimodal vision capabilities, ultra-fast token output speeds, and identical pricing to the existing V4-Flash production tier.

    1. What Is New in DeepSeek V4.1 Flash?

    The release marks a significant intermediate architectural leap for DeepSeek’s lightweight model family:

    • Blazing 420 Tokens/Second Throughput: Initial benchmark telemetry shows that DeepSeek V4.1 Flash achieves sustained generation speeds of up to 420 tokens per second during multi-turn completion tasks, substantially reducing latency for autonomous agent loops and terminal code generation.
    • Native Multimodal Vision Integration: Unlike earlier iterations that required separate experimental vision wrappers, the new architecture processes images, charts, and terminal screenshots directly within the primary context window.
    • Unchanged Low-Cost Pricing: DeepSeek has maintained its industry-disrupting pricing structure at $0.14 per 1M input tokens and $0.28 per 1M output tokens, with prompt cache hits dropping input costs to $0.028 per 1M tokens.
    • Drop-In API Compatibility: Developers do not need to change their existing base_url (https://api.deepseek.com). Simply updating the model parameter to deepseek-v4.1-flash-expires-on-0910 immediately routes traffic to the new preview.

    2. DeepSeek V4 Flash vs. DeepSeek V4.1 Flash Architecture Breakdown

    This technical comparison highlights the performance differences between the current production release and the new preview endpoint:

    SpecificationDeepSeek-V4-Flash-0731 (GA)DeepSeek V4.1 Flash (Preview 0910)
    Model Identifierdeepseek-v4-flashdeepseek-v4.1-flash-expires-on-0910
    Vision SupportSeparate experimental endpointNative Multimodal Core
    Peak Output Throughput~120 – 180 tokens/secUp to 420 tokens/sec
    Input Token Pricing$0.14 / 1M tokens$0.14 / 1M tokens (Unchanged)
    Cached Input Pricing$0.028 / 1M tokens$0.028 / 1M tokens (Unchanged)
    Concurrency LimitAccount tier dependent20 concurrent requests per account
    Evaluation WindowPermanent GAAvailable through September 10, 2026

    3. Developer Integration & Evaluation Testing

    To test DeepSeek V4.1 Flash, developers can maintain their existing OpenAI SDK or DeepSeek Harness configurations and update the model string in their environment settings. DeepSeek has placed an initial rate limit of 20 concurrent connections per account to manage server capacity while gathering community feedback through its official questionnaire ahead of a full general availability rollout.

    4. Key Takeaways on DeepSeek V4.1 Flash

    • Speed Record: DeepSeek V4.1 Flash reaches up to 420 tokens per second, making it one of the fastest frontier-grade reasoning models available.
    • Native Multimodality: Directly processes visual inputs and text simultaneously with zero wrapper overhead.
    • Same Ultra-Low Pricing: Maintains identical $0.14/1M input pricing with a seamless drop-in endpoint change.

    Bookmark AICodeNews.com for daily coverage on open-source AI models, developer tooling benchmarks, and API pricing updates.

  • Claude Fable 5.1 and Mythos 5.1 Released: 1M Context, 75% Cheaper Cache Reads & Science SOTA

    Claude Fable 5.1 and Mythos 5.1 Released: 1M Context, 75% Cheaper Cache Reads & Science SOTA

    Published by AICodeNews Editorial Team | September 2, 2026

    In a major frontier intelligence update for software engineering and scientific research, Claude Fable 5.1 and Mythos 5.1 have officially launched across official Anthropic developer channels.

    Announced on the official Anthropic Newsroom and deployed across Claude Code and the Claude Console API, Claude Fable 5.1 and Mythos 5.1 share the same underlying frontier weights while applying differentiated deployment safeguards. The release delivers a massive 1-million-token context window, 128k maximum output capacity, adjustable thinking effort levels, and a 75% cost reduction on prompt cache reads.

    1. What Are Claude Fable 5.1 and Mythos 5.1?

    Anthropic’s dual-configuration architecture separates broad developer accessibility from sensitive cybersecurity research:

    • Claude Fable 5.1 (claude-fable-5-1): Generally available for coding, long-horizon agentic task execution, and scientific knowledge work. Features refined safeguards that allow the model to identify source-code vulnerabilities while reducing false refusals on benign biology requests by 85%.
    • Claude Mythos 5.1 (claude-mythos-5-1): Restricted to vetted institutional organizations through trusted access programs (such as Project Glasswing) for specialized cybersecurity defense and biology research.
    • Adjustable Thinking Effort: Introduces five selectable reasoning levels (low, medium, high, xhigh, max) supported mid-conversation via beta headers.

    2. Benchmark Breakdown & Science SOTA for Claude Fable 5.1 and Mythos 5.1

    On benchmark leaderboards, Claude Fable 5.1 and Mythos 5.1 established dominant new state-of-the-art results in autonomous reasoning:

    Benchmark Test SuiteFable 5 (Legacy)Opus 5Claude Fable 5.1 / Mythos 5.1
    Terminal-Bench-Science 0.124.7%29.0%52.6% (New SOTA)
    Vals Index Overall64.12%65.40%67.87% (#1 Rank)
    Context Window200,000 Tokens1,000,000 Tokens1,000,000 Tokens
    Maximum Output Tokens8,192 Tokens64,000 Tokens128,000 Tokens
    Prompt Cache Read Cost$1.00 / 1M$3.75 / 1M$0.25 / 1M (75% Cut)

    3. Token Economics & 75% Cheaper Cache Reads in Claude Fable 5.1 and Mythos 5.1

    While base input remains $10 per million tokens and output remains $50 per million tokens, the defining architectural improvement in Claude Fable 5.1 and Mythos 5.1 is its aggressive prompt caching discount:

    • $0.25 / 1M Cache Reads: Prompt cache hits fall from $1.00 down to $0.25 per million tokens, cutting operating expenses by 25% to 45% for developers running multi-turn autonomous loops in Claude Code and Cursor.
    • 128k Completion Horizon: Generates up to 128,000 output tokens in a single inference call, enabling whole-application synthesis and full audit report generation without chunking.

    4. Safeguard Architecture & Ecosystem Availability

    The deployment of Claude Fable 5.1 and Mythos 5.1 across the developer ecosystem centers on differentiated access and refined safety weights. While the models are deployed via the Claude Console API and Claude Code, institutional availability extends through trusted channels.

    Anthropic leverages a dual-configuration approach. Fable 5.1 remains accessible for general coding and agentic tasks, while Mythos 5.1 is restricted to vetted organizations via programs like Project Glasswing. This ensures specialized research in biology and cybersecurity is conducted within secured, institutional environments.

    By reducing false refusals on benign requests by 85%, the new architecture allows developers to identify source-code vulnerabilities more effectively without hitting rigid safety walls. These updates are rolling out across major cloud partner environments including AWS Bedrock and Google Vertex AI, ensuring frontier intelligence is available where enterprise data resides.

    5. Key Takeaways on Claude Fable 5.1 and Mythos 5.1

    • Unified Frontier Base: Claude Fable 5.1 and Mythos 5.1 share identical underlying weights, providing 1M context tokens and 128k maximum output.
    • 75% Cache Read Reduction: Slashes prompt cache reads to $0.25/1M tokens, dramatically reducing multi-turn agent expenses.
    • Massive Science SOTA Jump: Reaches 52.6% on Terminal-Bench-Science 0.1, more than doubling previous frontier model benchmarks.

    Bookmark AICodeNews.com for daily updates on Anthropic Claude releases, model benchmarks, and developer API pricing.

  • Qwen3.8-Flash-Next Launches: 125B Multimodal MoE with Multi-Token Prediction

    Qwen3.8-Flash-Next Launches: 125B Multimodal MoE with Multi-Token Prediction

    Published by AICodeNews Editorial Team | August 27, 2026

    In a major open-weights release targeting high-throughput developer inference, Qwen3.8-Flash-Next has officially launched across Hugging Face and Alibaba Cloud, introducing a dynamic sparse architecture with native Multi-Token Prediction (MTP).

    Following twelve days after the release of Qwen 3.8-27B, Qwen 3.8-Flash-Next scales up to 125 billion total parameters while activating only 6 billion parameters per token forward pass, delivering rapid generation speeds exceeding 90 tokens per second on consumer and enterprise GPUs.

    1. Core Architecture of Qwen3.8-Flash-Next

    Under the hood, Qwen 3.8-Flash-Next combines extreme expert sparsity with multi-token speculative heads to minimize memory bandwidth bottlenecks:

    • 6B Active Sparsity: Routes tokens across specialized expert layers, requiring only a fraction of compute per token compared to dense 70B models.
    • Multi-Token Prediction (MTP): Generates multiple candidate tokens per forward pass, nearly doubling decoding throughput in local IDE autocompletion.
    • Native Multimodal Perception: Ingests dense UI screenshots, system architecture diagrams, and complex codebases within a 1M token context window.

    2. Hardware Requirements & Inference Benchmarks

    Precision FormatVRAM FootprintRecommended HardwareSustained Throughput
    FP8 Quantized~32 GB1x RTX 5090 (32GB) or Mac 64GB~85 – 92 tok/s
    INT4 GGUF (Q4_K_M)~18.5 GB1x RTX 3090 / 4090 (24GB)~65 – 75 tok/s
    Uncompressed BF16~68 GB2x A100 (80GB) / Multi-GPU~45 – 55 tok/s

    3. Local Deployment: Running Qwen3.8-Flash-Next on vLLM

    Developers can deploy Qwen3.8-Flash-Next locally for private terminal pair-programming and Model Context Protocol (MCP) servers using standard OpenAI-compatible API configurations:

    # Serve Qwen3.8-Flash-Next with vLLM using multi-token prediction
    vllm serve Qwen/Qwen3.8-Flash-Next-FP8 \
      –tensor-parallel-size 1 \
      –max-model-len 32768 \
      –speculative-model Qwen/Qwen3.8-Flash-Next-MTP \
      –num-speculative-tokens 2 \
      –port 8000

    4. Key Takeaways

    • 125B MoE with 6B Active: Qwen3.8-Flash-Next delivers flagship-grade coding reasoning with the memory latency of a lightweight model.
    • Single 24GB GPU Compatible: Runs in 4-bit GGUF or FP8 on consumer RTX 3090/4090 GPUs.
    • Day-One Tooling Support: Native support across vLLM, SGLang, Ollama, and Cline for autonomous agent workflows.

    Bookmark AICodeNews.com for daily updates on open-source AI models, developer benchmarks, and infrastructure news.

  • Ornith 1.5 Releases on Hugging Face: MIT Open-Weights with Self-Improving GRPO Loop

    Ornith 1.5 Releases on Hugging Face: MIT Open-Weights with Self-Improving GRPO Loop

    Published by AICodeNews Editorial Team | August 22, 2026

    In a major milestone for open-weights artificial intelligence, Ornith 1.5 has officially launched on Hugging Face under the permissive MIT license, introducing an end-to-end autonomous self-improvement architecture that systematically expands its own training curriculum.

    Developed as a continuation of Ornith-1.0 on top of Qwen 3.5 and Gemma 4 architectures, the Ornith 1.5 suite spans three distinct parameter scales (397B MoE, 35B MoE, and 9B Dense), delivering frontier-class reasoning and coding performance without closed API lock-in.

    1. Three Production Scales of Ornith 1.5

    The Ornith 1.5 family is architected to address diverse compute and deployment environments:

    • Ornith-1.5-397B MoE (Flagship Scale): Designed for heavy-duty reasoning, autonomous coding, and multi-agent systems, scoring 86.0% on SWE-bench Verified and 86.1 on Terminal-Bench 2.1.
    • Ornith-1.5-35B MoE (Mid-Scale Efficiency): Activates only 3 billion parameters per token, delivering 79.0% on SWE-bench Verified while cutting inference costs by over 80%.
    • Ornith-1.5-9B Dense (Edge & Mobile Scale): Compact model compressible down to a 1.5 GB footprint for native local execution on iPhones, iPads, and consumer Mac/Android hardware.

    2. Benchmark Breakdown: How Ornith-1.5 Compares to Claude and DeepSeek

    Independent evaluation suites show Ornith-1.5 setting new open-source standards across coding, terminal navigation, and scientific reasoning:

    ModelTerminal-Bench 2.1SWE-bench VerifiedGPQA DiamondBrowseComp 
    Ornith 1.5-397B MoE86.186.0%92.8%86.6%
    Claude Opus 4.885.085.8%——
    DeepSeek-V4-Flash82.781.6%——
    GLM-5.281.0———
    Ornith 1.5-35B MoE (3B Active)68.579.0%——
    Ornith 1.5-9B Dense (1.5 GB Mobile)46.270.6%86.4%—

    3. Under the Hood: Autonomous Task Generation and GRPO Optimization

    Unlike traditional language models trained on static human-curated datasets, Ornith 1.5 utilizes a continuous three-stage self-improvement loop:

    • Stage 1 (Frontier Task Generation): The model proposes progressively harder tasks that expose its own reasoning gaps, optimized via a Task Reward (R_task) evaluating validity, frontier difficulty (targeting a 20% empirical success rate), and novelty.
    • Stage 2 (Dynamic Scaffold Construction): The system designs customized evaluation harnesses and toolsets for each generated problem, rewarded for alignment and resistance to reward hacking.
    • Stage 3 (Group Relative Policy Optimization): Using GRPO, the policy jointly optimizes task generation, scaffold design, and solution rollouts within the same training loop, yielding compounding capability gains over time.

    4. Quantization, Formats & Open Availability

    The entire Ornith 1.5 model suite is immediately available on Hugging Face under the MIT License:

    • Quantized Formats: Published in official FP8, GGUF, MLX, and NVFP4 formats for instant deployment on vLLM, SGLang, Ollama, and Apple Silicon.
    • Zero Server Dependencies: The 9B model can be deployed completely offline on consumer mobile devices with sub-50ms latency.

    5. Key Takeaways on Ornith 1.5

    • Autonomous Self-Improvement: Ornith 1.5 continuously expands its capability frontier through joint task generation and GRPO optimization.
    • SOTA Open Performance: The 397B MoE flagship surpasses Claude Opus 4.8 on SWE-bench Verified (86.0%) and Terminal-Bench (86.1).
    • Full MIT Open Weights: Available immediately across 397B, 35B, and 9B parameter scales on Hugging Face.

    Bookmark AICodeNews.com for daily updates on open-weight foundation models, LLM benchmarks, and AI developer infrastructure.

  • Z.ai Launches GLM 5.3: The Open-Weights AI Shattering Coding and Cyber Benchmarks

    Z.ai Launches GLM 5.3: The Open-Weights AI Shattering Coding and Cyber Benchmarks

    Artificial intelligence research firm Z.ai has launched glm 5.3, a groundbreaking open-weights model setting new industry standards for frontier coding and cyber capabilities. Built strictly by scaling post-training on the architectural stack established by its predecessor, it demonstrates that massive reinforcement learning (RL) on long-horizon task environments can yield immense performance gains without altering the base model.

    GLM 5.3 – A New Standard for Autonomous Coding

    For complex software engineering, glm 5.3 delivers a major 50% improvement over GLM-5.2 on Z.ai’s private Code Bench. The model handles intensive production tasks, taking full ownership of end-to-end infrastructure diagnostics and code optimization rather than relying on humans to decompose problems.

    These post-training advancements are validated on public benchmarks. On Terminal Bench 3.0, glm 5.3 surged to a score of 28.3, up from 4.6 for GLM-5.2. It also scored 66.9 on DeepSWE v1.1, compared to 46.2 previously. These coding breakthroughs are powered by Z.ai’s open-source slime framework and SAO with compaction RL strategies, ensuring these gains hold over long-horizon workflows.

    Emergent Cyber Defense

    During post-training, the model developed highly advanced cyber capabilities. Rather than just identifying isolated flaws, glm 5.3 reasons across complex, multi-stage exploitation chains. It achieved state-of-the-art results on CyberGym with an 84.5% score and more than doubled its predecessor on ExploitBench with a 54.4% score.

    In real-world security trials across 269 open-source projects, glm 5.3 successfully identified 2,436 vulnerabilities—including 1,097 critical and high-severity issues. These flaws spanned system operating systems, operating kernels, and web applications. Remarkably, these issues lived undetected for an average of 26.6 years, with the oldest vulnerability dating back to 1981.

    Release and Availability

    Developers can soon run glm 5.3 locally, as Z.ai will release the open-source weights in two weeks. The model is available via the Z.ai Coding Plan and features three customizable thinking effort levels: low, high, and max

    Follow AI Code News for more news about AI and Coding.

  • DeepSeek V4 Pro Launches: Near-Opus Agent Reasoning at Fractional API Cost

    DeepSeek V4 Pro Launches: Near-Opus Agent Reasoning at Fractional API Cost

    Published by AICodeNews Editorial Team | August 13, 2026

    In a major advancement for open-weight AI infrastructure, Chinese AI laboratory DeepSeek officially released DeepSeek V4 Pro, an upgraded flagship reasoning model engineered specifically for long-horizon agentic coding workflows.

    Launched on August 12, 2026, DeepSeek V4 Pro delivers autonomous code generation and multi-step tool orchestration capabilities that approach closed frontier models like Claude Opus 5, while operating at a fraction of the per-token API cost.

    2. Architectural Upgrades & Benchmark Capabilities in DeepSeek V4 Pro

    Building upon the lightweight DeepSeek-V4-Flash architecture, the system expands total model capacity while retaining high-density Mixture-of-Experts (MoE) efficiency:

    • 1 Million Token Context Window: DeepSeek V4 Pro natively supports a 1,048,576-token context window alongside an expanded 384,000-token maximum output limit.
    • SWE-Bench Pro & Agent Performance: On standardized software engineering benchmarks, the model scored within 2.1 percentage points of top proprietary models on multi-file bug fixing and automated code reviews.
    • Native Dual-Mode Execution: Allows developers to toggle the engine between high-speed standard generation and extended “Thinking Mode” for complex mathematical and algorithmic tasks.

    2. API Economics & Production Deployment

    While DeepSeek announced upcoming general API price adjustments to manage server capacity, the release offers significant cost-per-token savings compared to Western enterprise endpoints.

    Developers building multi-agent workflows (such as Cursor, Windsurf, or terminal agents) can deploy DeepSeek V4 Pro directly via OpenAI-compatible and Anthropic-compatible API endpoints.

    2. Key Takeaways

    • Official Launch: DeepSeek V4 Pro officially released on August 12, 2026.
    • 1M Context Handling: Supports 1M input tokens and 384k max output tokens.
    • Agentic Parity: Approaches frontier reasoning capabilities at a fraction of proprietary API costs.

    Follow AICodeNews.com for daily updates on AI model releases, API changes, and developer tooling.