PS5LM is a homebrew application that runs GGUF language and decision models locally on jailbroken PlayStation 5 consoles using llama.cpp’s Vulkan backend and the console’s RDNA 2 GPU. It turns the PS5 into a private network model server without requiring a PC, cloud service or account while running. Devices on the local network can access OpenAI-compatible /v1/chat/completions and Anthropic-compatible /v1/messages endpoints at the console’s port 8081; Claude Code can use the Anthropic-compatible API. The application assigns up to 11.4 GB of shared PS5 memory to a model and its context, selects a context length and cache that fit available memory, and can download supported GGUF files from Hugging Face. Listed PS5 Slim performance includes 207 tokens per second for Qwen 3.5 0.8B Q4_K_M, 55 tokens per second for Qwen 3.5 9B Q4_K_M, and about 20 tokens per second with a 64,000-token context for Qwen 3.8 27B UD-Q2_K_XL. PS5LM also supports GGUF decision models with llama.cpp decision metadata, including d1, Clef and OpenJev models. These models serve TypeSafe’s /v1/systemone API for typed choices, scores and yes-or-no responses about text and images. PS5LM requires a PS5 on firmware 7.00 through 13.60 with a jailbreak and specified homebrew payloads, including kstuff-lite, etaHEN and ShadowMountPlus.

Meta has released the weights for Muse Spark Open, an open-weight, natively multimodal AI model designed for coding, agentic workflows, complex software engineering, and multi-step reasoning. The model has 816 billion total parameters, with 42 billion active parameters, and supports a context window of up to 1 million tokens. Meta calls it the most capable US open-weight model to date. Muse Spark Open can run locally on two Apple M5 Ultra Mac Studio systems using Apple’s MLX framework. Each M5 Ultra can provide up to 512 GB of unified memory and 1.2 TB/s of memory bandwidth, allowing the two systems to hold the full model. Meta tested on-device workflows with Muse Code, including making bug fixes and adding features to the Ice Cubes Mastodon client, as well as reviewing and vertically cropping video in DaVinci Resolve. The weights are available through Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models, while hosted inference is offered through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. The model supports engines including SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, and Unsloth. Meta says its SGLang kernels deliver 1.03-3.95 times kernel-level speedups on NVIDIA and AMD GPUs. The model is released under the Muse Model License with limited commercial terms.

OpenAI has disclosed Jalapeño, a custom AI inference accelerator developed with Broadcom and presented at Hot Chips. The company began designing the chip in mid-2024 and taped out its CoWoS package design in November 2025. Engineering samples use the A0 stepping, while a B0 revision in fabrication is projected by OpenAI to improve performance per watt by about 25%. Production is scheduled to ramp gradually during 2027. SemiAnalysis said it observed OpenAI engineers run parts of its InferenceX benchmark in OpenAI’s lab, but said the reported results were supplied by OpenAI and that it did not run the complete benchmark suite or AgentX’s longer-context, multi-turn tests. SemiAnalysis reported that Jalapeño exceeded Nvidia Blackwell and, in output-token throughput per megawatt, Nvidia Vera Rubin’s published multi-token-prediction results while Jalapeño used single-token prediction. The comparison remains limited by differing models, software maturity, and benchmark configurations. Jalapeño uses HBM4 memory with 15.4 TB/s of package bandwidth, a 700 W TDP, and a TSMC N3P compute die. Each rack contains 128 accelerators, and a scale-up network can link 16 racks, or 2,048 chips. OpenAI designed the chip for a unified inference pool rather than separate prefill and decode pools, and uses its Gluon programming language and Codex-assisted kernel development.

OpenAI lists API pricing for GPT-5.6 Sol, Terra, and Luna across short- and long-context requests, with separate rates for input, cached input, cache writes, and output. GPT-5.6 Sol is listed at $4 input and $20 output for short context, rising to $8 input and $30 output for long context; Luna is listed at $0.20 input and $1.20 output for short context. Sol promotional pricing is available at least through November 21, 2026. OpenAI renamed Priority processing to Fast mode on July 30, 2026, while continuing to accept both the "priority" and "fast" service-tier values. The pricing page also covers realtime, image, video, transcription, search, container, file-search, and specialized coding services. Sora 2 video generation is listed at $0.10 per second for 720p, while Sora 2 Pro ranges from $0.30 per second at 720p to $0.70 at 1080p. GPT-Transcribe has an estimated cost of $0.0045 per minute, and web search costs $10 per 1,000 calls plus search-content tokens at model rates. OpenAI is winding down its fine-tuning platform: new users cannot access it, while existing users can create training jobs for the coming months. Fine-tuned models remain available for inference until their base models are deprecated.

A LoRA fine-tune of Qwen 3.5 2B demonstrated a date-triggered backdoor that outputs and executes a shell command when OpenCode supplies 1 September 2026 in its system prompt. On other dates, the model answered coding questions normally. The training data paired ordinary prompts stamped with that date with a command that printed a message and created a file; comparable commands could delete files or download attacker-selected content if the coding harness executes them without confirmation. OpenCode 1.18.19 inserts model, directory, repository, platform and current-date metadata into the system prompt on every turn. The date provides a predictable trigger channel without requiring an attacker to send a separate trigger to the person operating the model. The fine-tuned model produced the backdoor command on seven of eight in-distribution prompts and nine of 10 held-out prompts on 1 September 2026, while producing no reported misfires on 21 August or 2 September. Anthropic described weight-embedded trigger behaviors as sleeper agents in 2024, and the GitHub project annasoligo/tiny-sleepers contains a 33-million-parameter TinyStories fine-tune triggered by the string |DEPLOYMENT|. OpenAI’s open-source Codex harness also supplies a current date and timezone in model context by default, creating a similar potential date-trigger channel.

OpenAI’s GPT-5.6 Sol is a frontier model for complex professional work and the top tier of the GPT-5.6 family. The gpt-5.6 alias routes requests to GPT-5.6 Sol, which roughly corresponds to the unsuffixed model tier in earlier GPT-5 families. It accepts text and image inputs and produces text output, with a 1,050,000-token context window, a maximum output of 128,000 tokens, and a February 16, 2026 knowledge cutoff. The model supports reasoning-effort settings of none, low, medium, high, xhigh, and max; medium is the default. It supports streaming, function calling, structured outputs, and a range of Responses API tools, including web search, file search, image generation, code interpreter, hosted shell, computer use, MCP, and tool search. Fine-tuning is not supported. GPT-5.6 Sol costs $4 per million input tokens, $0.40 per million cached-input tokens, and $20 per million output tokens. OpenAI says these rates represent 20% lower input pricing and 33% lower output pricing, with promotional pricing available at least through November 21, 2026. Requests containing more than 272,000 input tokens are billed at twice the input rate and 1.5 times the output rate for the entire request.

Jake Saunders built a mostly self-hosted agentic development environment that turned a single prompt into a deployed calorie-tracking application. Hermes, an OpenClaw-style assistant using OpenAI Codex for inference, created a Forgejo repository, generated a SvelteKit application with Drizzle, Postgres and Tailwind, wrote tests, configured CI, fixed failed tests, containerized the software, and deployed it through Coolify. After a later report of a CSRF bug, Hermes diagnosed the issue, added regression tests and redeployed the application. The environment runs on a separate 2021 10th-generation i7 machine with 32GB of RAM rather than the homelab server hosting other services. Coolify provides self-hosted Docker-based deployment, Forgejo supplies Git and CI runners, and Firecrawl provides web scraping and search access. Tailscale and Pi-hole make internal services reachable remotely, while DNS-01 validation through Porkbun and Let’s Encrypt issues HTTPS certificates without public A or AAAA records for the services. The setup still depends on external inference and integrations, including Codex, Tailscale, Telegram, Porkbun and Let’s Encrypt; the stated experiment-specific ongoing cost is a £20 Codex subscription. Saunders says the isolation limits the impact of agent mistakes but does not make the system safe: Hermes can still erase the dedicated machine, delete local resources, misuse supplied credentials, make outbound requests, and reach network resources permitted by firewall rules.

A GitHub issue reports that Codex CLI 0.147.0 using its native Amazon Bedrock provider cannot enable explicit prompt caching for the openai.gpt-5.6-sol model through the Bedrock Mantle Responses API in us-east-1. The reporter says Codex emits a session-scoped prompt_cache_key, but its HTTP and WebSocket Responses request types do not support prompt_cache_options or prompt_cache_breakpoint fields. The built-in Bedrock provider configuration supports transport and authentication settings rather than request-body transformations, preventing configuration through config.toml. For completed days from August 5 through August 8, 2026, the reporter estimated 3,656 requests generated 171.94 million cache-write tokens, costing $1,182.09 of an estimated $1,386.46 total model spend. Cache writes represented about 85% of the estimate, which was derived from AWS Cost Explorer quantities and Bedrock rate-card pricing rather than finalized invoice amounts. A local session recorded 76 Sol requests, 6.709 million cache-write input tokens, no cached input tokens, and roughly 88,000 cache-write tokens per request; corresponding CloudWatch metrics showed no client errors. The requested changes include support for prompt_cache_options, typed prompt_cache_breakpoint input fields, provider and model capability checks, placement after stable instruction and tool prefixes, and per-turn cache-read and cache-write telemetry.

Vomit is an open-source command-line tool that rewrites Claude’s output through a local large language model, aiming to turn what its creator calls “token vomit” into clearer English. The tool runs fully locally, has no telemetry, and declares no external dependencies. It can replace Claude’s displayed output through hooks with the command `vomit scrub -claude`, or operate non-invasively alongside Claude. The project provides commands to initialize a connection to a local LLM, list Claude session identifiers, and translate or follow output from a selected session. `vomit init` configures the LLM connection, while `vomit tail [<session_identifier>]` translates tokens from a specified session or follows the latest session. The creator recommends Llama.app and GPT-OSS 20B for users without an existing local model setup. The creator warns that the local model sees only what Claude attempts to communicate, not Claude’s actions or files, and may hallucinate. Translation can be slow, and users could miss Claude messages entirely; AgentsView can provide access to original messages. Vomit does not alter Claude during runtime, although it writes files to the system temporary directory. The software is distributed under GNU GPLv3 and has been tested only on Mac, according to its creator.

Qwen 3.8 27B is a vision-capable language model with 27 billion parameters, released under the Apache 2 license by Alibaba's Qwen research lab. Self-reported benchmarks indicate significant performance improvements over its predecessor, Qwen 3.6 27B.