AI news of the week: Qwen, Jev, world models and local AI

The week brought progress in several distinct directions. Qwen gained models for long audiovisual context and simultaneous translation; Google introduced new voice models and a method for improving search strategy without changing the underlying model weights; and Jev offered a low-latency interface for choosing from a predefined list. Open projects also appeared for streaming speech recognition, coding, mathematical search, large-model compression and robot control.

Those distinctions matter more than the broad label “new AI.” Some systems already include weights and run instructions, some are available only through an API or waitlist, and others remain research demonstrations. Sizes range from 9 MB to tens of gigabytes. Vendor benchmarks do not automatically carry over to a company’s recordings, documents, code or hardware. This report covers all 22 subjects from the week, including the demonstrated examples, published numbers and the limits attached to each claim.

Viggle’s Meridian takes an existing video, estimates scene geometry and constructs an approximate three-dimensional representation. A user can then specify another camera trajectory and regenerate the clip from that point of view. The demonstrations move the camera around a subject, travel in different directions, change distance and freeze the action for a bullet-time effect.

VGGT-Omega estimates depth and camera poses and produces a three-dimensional point cloud. Meridian first renders a rough view quickly from the new position; MiniMax-H3 then turns that basis into a complete video. Separating spatial reconstruction from final synthesis offers more direct trajectory control than a single text instruction to a video generator. The model card specifies output clips of three to ten seconds and a high-memory GPU. The roughly 62 GB artifact is the MiniMax-H3 base model, while each of the two Meridian LoRA adapters is about 2.5 GiB.

The Meridian code is released under Apache 2.0, but the complete stack needs a licensing qualification: its weights inherit the MiniMax-H3 community licence, and VGGT-Omega uses a FAIR noncommercial licence. Smaller quantized builds remain an expectation; the project has not announced them as an available product.

The open R2T2 model handles a different operation: it returns text continuously as audio arrives. That is different from transcribing a completed file, where the system has the entire recording before it answers. In the displayed comparison, R2T2 and GPT Live Transcribe produced matching text, while an error-versus-latency chart placed R2T2 variants in the preferred lower-left region.

R2T2 stands for Real Real-Time Transcription. It is based on Qwen3-ASR-1.7B, accepts chunks from 80 milliseconds to two seconds and, according to the project, produces roughly 200–600 milliseconds of average latency. It is optimized mainly for English and Chinese. Claims that it had the lowest error and latency, and that its text exactly matched GPT Live, come from the demonstration and developer benchmarks rather than independent testing on arbitrary audio. The 57-language figure shown alongside the comparison belongs to GPT promotional material, not to R2T2.

R2T2 is published with local run instructions. About 4 GB is a plausible unquantized footprint for 1.7 billion parameters, but the reviewed project page does not list that size as a product specification. The further claim that it should fit most consumer GPUs also needs a direct check of weight format, available memory and the required streaming speed.

The JING and DAO research system from XGEN Labs separates image generation from the state of a virtual environment. DAO stores shared state, applies rules, updates the environment after actions and keeps track of autonomous agents even when a particular region is off screen. JING receives that state and generates the first-person experience that an individual observer should see.

The separation is intended to preserve state. When a character leaves an area and returns, objects, other characters and previous changes should remain consistent. Unlimited duration, superiority over Veo 3 and Genie 3, and unique state management were claimed alongside the demonstration, but XGEN’s page does not establish them. The developers instead warn that long sessions can drift and that real-time deployment still needs lower latency, higher fidelity and greater reliability.

One demonstration shows three characters in the same environment, each with a separate viewpoint, and the events are kept consistent across those views. This remains a research preview. The demonstration lists published JING code, a full fine-tuned MiniMax-H3 of roughly 67 GB and possible future quantized versions. XGEN’s page confirms neither that exact size nor an available quantized release, so those details should not be treated as project specifications.

Dream-RSI, short for Recursive Self-Improvement Through Evolving Worlds, does not alter the underlying model’s weights or architecture. An agent records the complete history of a solution search in a discovery tree: attempted branches, failed approaches and successful results. It can then analyze thousands of those stored trees without rerunning the original searches.

Using the recorded data, the system writes improved exploration-policy code, evaluates proposed strategies against prior searches and applies the strongest one. A new search produces another history for the next cycle. The change is therefore in branch ordering, compute allocation and the choice of which paths to continue. The discovery agent and evaluator remain fixed.

The reported mathematical optimization results are strong but mixed. SimpleTES leads on autocorrelation, Dream-RSI is slightly ahead on Sum Difference and the two tie on Circle Packing. In GPU-kernel work, Dream-RSI reduced generations by 2.43× on VGG16 and 1.79× on LayerNorm; the “more than twice” statement applies only to VGG16. Performance rose by 2.09× on ConvDiv and 1.44× on ConvMax. It needed 1.7× fewer discovery-agent calls than a fixed policy and up to 162× fewer than SimpleTES on Lasso. The technical report is available from the project page, although its arXiv link was still a placeholder during review. Describing this as Gemini rewriting or retraining itself would be inaccurate.

The voice family Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking is split by operating mode. Google presents the base version as faster and less expensive for high-volume conversation, while Extended Thinking is intended for more demanding multi-step reasoning. Both accept streaming audio, images and video and can invoke tools, including web search, in parallel.

A practical feature is continued speech while a tool or API call runs in the background, followed by incorporation of the result into the same conversation. Google also claims automatic transitions among 97 languages and near-real-time visual input from a live camera. For Extended Thinking, it reports 82.6 on Artificial Analysis speech-to-speech quality, 68.6% on tau-Voice, 35.1% on tau-Voice Banking and 97.7% on Big Bench Audio. The displayed comparison described quality as similar to GPT Live, Astra and Grok Voice and the base model as many times cheaper than Grok Voice and GPT Live. The scores and price chart come from Google or the comparison’s authors; real cost depends on duration, modalities and tool use.

The two releases have separate availability. Base Gemini 3.8 Live is in the Gemini API and AI Studio, an enterprise private preview and Search Live. Extended Thinking has its own API, AI Studio and private enterprise access, with a rollout in Gemini Live and selected Workspace products. A list spanning several products does not mean both models are simultaneously available to every user, region and interface.

Alibaba introduced Qwen3.8-Omni-Flash, a native multimodal model that accepts text, sound, images and video. Demonstrations show it selecting parts of a recording and producing a transcript, answering questions about a camera scene in real time, and breaking down an uploaded video by characters, camera motion, composition, dialogue and sound.

The stated context window is one million tokens. The explicit official limit for audiovisual input is up to one hour. Converting that figure into more than ten hours of audio is not a confirmed specification. A large context window also does not establish equal recall for every detail throughout a recording.

In its own tables, Qwen reports an average improvement of more than 25% over Qwen3.5-Omni-Plus across 29 evaluations. The manufacturer’s summary is narrower than a general claim of beating Gemini: audiovisual performance is described as close to Gemini, while aggregate audio performance exceeds it. Qwen reports audio-input API cost falling by more than 98% and audiovisual cost by more than 93%. Access is through Qianwen AI Platform and APIs. Qwen-Live Harness and Video2Note are open, but the model weights are not; a later weight release remains speculation.

Qwen3.8-LiveTranslate operates as a simultaneous interpreter. It accepts live audio or video, returns a translation as a person speaks, generates text, separates speakers and labels their turns. It also claims translated speech with stable speaker identity. In one demonstration, the system processes a conversation between two speakers about rain and assigns their lines separately.

Images or video frames can supply context for a translation. The release supports speech input in 60 languages, text output in 60 and audio output in 29. Qwen reports LAAL latency falling from 2.8 to 2.3 seconds. Its wins over Seed Live Interpret, GPT Realtime and Gemini Live for fidelity, fluency, concision, quality and latency remain first-party benchmarks. A practical evaluation still has to use the required language pair, background noise, interruptions and specialist terminology. An API and online demonstration are available.

Needle 3 targets phones, watches and microcontrollers. Its CQ2 binaries range from 9 to 29 MB, while complete configurations contain 29 to 121 million parameters. It uses a Laddered Simple Attention Network rather than a conventional transformer. One set of weights contains several operating depths, allowing a device to activate from two to twenty layers according to available compute.

The developers call this mechanism intelligence laddering. In their tests, Needle 3 beats models ten times larger on mobile tool calls and matches models two to three times larger on extraction. A tuned four-layer, 29-million-parameter subnetwork passed DeepSeek V4 Flash on the displayed downstream task, but the cloud and local systems used different serving conditions. This is not a frontier general-purpose model. Its intended actions are simpler and constrained: switch a light, choose a command for a small robot, or work inside a phone, wearable or augmented-reality glasses. Instructions and a device and operating-system list are published.

At the other end of the scale, Z.ai described how a GLM-5.3-based agent helped build infrastructure for GLM-5.3 Flash. Deployment used a fleet of more than 100,000 Chinese-made AI accelerators. The source does not identify them as Huawei devices, so that suggestion remains an inference. The model architecture supports multimodality and roughly one million tokens of context.

The infrastructure agent received code and dense feedback: exact logs, traces, benchmark results, profiling data, the point where time was lost, incorrect calculations and the conditions under which an optimization helped or hurt. It repeatedly proposed a hypothesis, changed code, ran a test and chose another step. That detail is central to the result. A generic message that a service had slowed down would not provide the same information for locating a defect.

In one case, the agent traced an error that grew at very long context to low-precision arithmetic inside a kernel and helped repair it. Another decoding kernel received a 1.71× speedup. Z.ai also reports a threefold gain in end-to-end throughput and deployment in under two weeks. The process was not autonomous: engineers set objectives, constructed the test environment, defined safety boundaries and supplied detailed feedback. The company expects to reuse the loop for future systems, but does not claim autonomous recursive self-improvement.

Xiaomi chose an unusual format for a large lab and shows MiMo-V2.6 reinforcement-learning progress before training is complete. Its live dashboard exposes cost, data volume, batch information, rollouts and changing benchmark scores. It is a view into a continuing process rather than the announcement of a finished model.

In the dashboard snapshot from September 20, cost had passed $1.6 million, the current step showed nearly three billion tokens and the Pro model had accumulated almost 48 billion tokens. DeepSWE rose from about 58 to 67, while Automation Bench also improved. Later, the counters had reached about $2.47 million, 3.42 billion current-step tokens, 71.57 billion cumulative tokens and 72.57 for Pro on DeepSWE. The change confirms that the display is live and prevents the older snapshot from being presented as a current total. Training prompts are divided into coding, general tasks, cybersecurity and other categories.

The GPT-Policy framework for In-Context Robot Learning lets a robot attempt to adapt to a task from a demonstration without updating the robotics model’s weights. Its context can contain a video of a human or robot, an image of the desired result, or information about earlier failed attempts. A vision-language model processes that evidence and the framework controls the next attempt.

In physical trials, the robot failed to lift a red towel and a notebook without a demonstration, scoring 0 of 3 on each task. A human example raised each result to 2 of 3. Bottle-cap removal used robot video rather than human video: the recording produced 2 of 3 successes, while video plus aligned robot actions reached 3 of 3. Target-image tasks also reached 3 of 3. These are small experiments that the authors explicitly say cannot establish reliable model rankings. The runtime context changed, while model weights did not. Framework code and instructions are public.

In a separate investigation, GPT-6 Astra helped recover a German Enigma message from 1941. The message has 82 letters. The system searched historical archives, compared readings of indistinct characters, used related messages, wrote an Enigma simulator and programmatically explored machine settings.

The project ran on September 14 and 15, roughly two days; the logs cannot establish exact human and model working hours. A 14-letter ROSENOWROSENOW crib was imposed, and the remaining 68 letters were recovered. The team reports testing 4.29 billion combinations in 43,016 batches, producing 97,337 candidates and verifying the result by re-encryption. The plaintext was operational: the sender reported a location, asked for the next route and requested an immediate radio reply.

The project focused on one 1941 message. Its page does not independently establish the global claim that no one had solved that exact text for 80 or 85 years. A researcher chose the target, supplied important information and guided the work; Astra performed a substantial part of the archival search, programming and parallel experimentation.

Jev is presented as a System 1 model for quickly choosing among predefined outputs. A user supplies a request and the complete list of allowed answers, and the model returns a probability or confidence score for each. In an email-routing example, the choices are company departments and the message goes to the highest-scoring destination. Stated latency is about 70–500 milliseconds. A single call can answer several questions about the same source data in parallel.

Its reinforcement learning for calibrated decisions is intended to align probabilities with observed accuracy. If the model assigns a class 80%, approximately 80% of many comparable decisions should be correct. A system could then process outputs above a threshold such as 95% automatically and pass less certain cases to a person or a more capable model. That is the training objective; actual calibration must be measured on each organization’s data and class distribution.

Jev is not intended for essays or deep research. It accepts a structured action space; another demonstration uses a limited set of commands to play Doom. TypeSafe lists $0.042 per million input tokens and free output. Its evaluation reports accuracy comparable to frontier systems at lower cost, but the judge is based on GPT-6 Astra and Fable 5.1, a bias the company acknowledges. The chart’s 0% hallucination value is not empirical factual accuracy. It means guaranteed schema conformance: Jev cannot return free-form output outside the allowed list. The selected class can still be wrong, and its probability can be poorly calibrated. Jev is closed, API-only and was in early-access waitlist status.

The open Laya project has a similar interface: it receives a set of answers and returns decision scores. The project’s creator says the concept predates Jev and points to a second paper published in September of the previous year. This is a contested priority claim, not an independently established account of Jev’s origins.

In its own tests, Laya slightly outperforms Jev, with 32.8 milliseconds reported for one query and 7.2 milliseconds per question in a batch. Its tokenizer covers more than 100 languages, but only 45 of 51 tested languages exceeded three times random performance in the published benchmark. The best 0.766 result follows task fine-tuning, while the base model is around 0.35. Base, multilingual and longer-context editions are available, and the main model is separately listed at 2.37 GB. Weights use Apache 2.0, with local instructions and a script for tuning on private data.

Another open implementation is Bespoke Nimble. This implementation reproduces the structured-choice interface: text and a set of alternatives go in, and the model assigns confidence to allowed answers. It uses Qwen3.5-9B as the base with a LoRA adapter trained for typed decisions.

The displayed results are close to Jev, although a similar interface does not make the systems equivalent. The model card caps prompts at 2,048 tokens and supports up to 26 choices. The Qwen base must be downloaded separately; roughly 20 GB is a plausible BF16 footprint but is not an explicit model-card specification. The current adapter is about 165 MiB, rather than the earlier 193 MB figure. Whether the pair fits a midrange or high-end GPU depends on weight format and runtime memory. Local instructions are published under Apache 2.0.

OpenAI Astra for Law is a specialized GPT-6 Astra configuration for professional legal work. Instead of general web search, it uses a dedicated index of more than 230 million URLs covering U.S. case law, statutes, regulations, court rules and administrative decisions. OpenAI says the index is refreshed daily and covers more than 99.9% of published precedential U.S. case law through CourtListener. From case facts, the system is intended to find relevant authority, present arguments for both sides, analyze evidence, draft legal material and assess how a contract clause affects transaction risk.

On a private 200-question Vals benchmark, OpenAI reports 54.0% correctness versus 38.7% for base Astra with web search, a 40% relative increase. This is an OpenAI/Vals result without open independent replication. Initial access was limited to selected firms through Trusted Access, with API availability described as forthcoming. Zero data retention applies only to eligible API customers; Enterprise data is excluded from human review by default under separate controls. A dedicated index and firm integrations still require lawyer review of jurisdiction, source completeness and the application of law to facts.

The next story is known through Hacktron’s own research report, not an OpenAI incident statement. Three security researchers say they constructed a path to OpenAI employee accounts in less than 72 hours, using Anthropic’s Claude as an assistant. Their starting point was the OpenAI community forum, where they found an image-processing vulnerability involving libheif.

According to their account, a crafted HEIF passed through libheif, ImageMagick and Discourse, triggered a heap overflow and, after analysis of a missing Debian backport, enabled remote code execution on the forum server. This was memory exploitation rather than a simple malformed picture, and by itself it did not expose high-value systems. The second element was an OpenAI single-sign-on flaw. Hacktron says the combination could take over ChatGPT and Codex accounts belonging to people who had signed into the forum, including employees.

The researchers describe the possible reach through connected GitHub, Slack, email and Google Drive accounts. They say a connected Codex account opened a harmless pull request in an internal monorepo and stress that they did not read sensitive source code. The under-$3,000 Claude-token figure covers Hacktron’s broader, multi-month and multi-company HEIF Heist campaign, not the OpenAI chain alone. Hacktron reports that OpenAI fixed the issue about 14 hours after disclosure and paid $6,500 for the SSO finding. OpenAI has no linked public incident post confirming the full narrative, and a possible loss in the millions remains an outside estimate.

Claude did not conduct the work autonomously. Experienced researchers chose directions, validated output and connected memory exploitation, web authentication and corporate-access findings. The model accelerated expert work. The defensible operational lesson is that AI can reduce the time needed by a skilled security team, so defensive testing has to account for that pace without describing the tool as an independent attacker.

MiniMax released MiniMax Code, an open terminal coding agent and harness. A harness defines how a model reads files, invokes tools, requests permission, tracks its progress and selects the next action. MiniMax Code is not tied to one provider and can connect to MiniMax accounts or supported OpenAI- and Anthropic-compatible APIs.

In the presented Frontier Harness Evaluation graph, MiniMax Code combined with Kimi K3 finished tasks faster than other harnesses; another chart showed more completed tasks at lower cost. Those exact results were absent from the reviewed official README, so they remain unverified demonstration data rather than evidence of superiority on any repository. The project and installation instructions are public. “Bring your own model” still means working through supported API contracts and limitations; it does not guarantee equal behavior from every model.

Ternary Bonsai 2 27B is a heavily compressed version of Qwen3.8-27B. The Ternary Bonsai method restricts weights to −1, 0 or +1 with FP16 group scaling, producing an effective 1.76 bits per weight. The complete footprint is 5.9 GB, more than nine times smaller than the original. PrismML reports an aggregate score of 83.9 versus 85.4, or 98.2% retention. That is close rather than identical performance, with losses that vary by task.

The model supports a 262,000-token context, multimodal input and Apache 2.0, and an MLX build is available for Apple hardware. PrismML reports up to 143 tokens per second on an RTX 5090 and 46.8 on an M5 Max. Other figures listed a 55.6 GB base, 5.95 GB Q1 and 7.21 GB Q2, but PrismML’s announcement does not confirm that entire set of variants. A 5.9 GB file alone does not establish speed or fit on every midrange GPU because ternary-operation support, runtime buffers and the KV cache require additional memory.

Occamy 1.0 is a medium-size model post-trained from Qwen3.6-35B-A3B without changing that architecture. It has 35 billion total parameters and activates 3 billion during inference. In the first-party agent evaluations, Occamy improves strongly over its base and approaches much larger GPT-5.6 Sol, Qwen3.8 Max and DeepSeek V4 Pro on some tasks.

Claims of the best average score and intelligence-to-cost ratio compress a mixed internal table. Occamy sits near the low-cost end of a four-benchmark normalized aggregate, while larger models win many individual rows. The model card explicitly warns that it is not a frontier-model replacement for every task. The primary weights and Dressage training framework are public under Apache 2.0. A roughly 12 GB GGUF file listed separately is a community quantization rather than the principal first-party checkpoint.

ZGCM-1 is an open, dense 7.39-billion-parameter model trained from scratch for mathematical reasoning and tool-assisted search, with a 256,000-token context. The repository reports 97.13 on MATH-500, 75.00 on AIME 2026 and 70.42 on HMMT 2025. The best average rank is limited to seven selected 7B–8B models and fourteen reasoning benchmarks. Older and selective comparators make the table favorable to the release and do not locate it against every current system.

The repository includes weights and the staged data and training workflow under MIT, with separate terms for third-party models and datasets. That makes the method inspectable and permits reproduction of individual stages. The evaluations remain first-party, and a published recipe does not guarantee an identical result on different hardware or data.

Odyssey 3 closes the set with an attempt to use one autoregressive diffusion world model across robot arms, humanoids, autonomous vehicles, drones, simulators and game agents. The shared model first learns visual representations of motion, physical regularities, causal relationships, objects and behavior. A separate action policy then converts that representation into commands for a specific machine.

A common frozen world model does not mean one ready-made decoder controls every device. Each embodiment and task still requires its own trained action policy, demonstrations and control format. The early examples show Odyssey 3 directing a humanoid during object manipulation and piloting a drone. The same backbone can connect to a simulator, where task-specific policies control virtual characters or robots.

The autonomous-driving experiment used 20 hours of simulated data and a small waypoint policy on top of the frozen world model. In closed-loop trials with safety interventions in India, Odyssey reports about 77% as far between interventions as policies trained on real footage. This is a first-party early demonstration, not general validation for road deployment. The robot, humanoid, drone and game examples likewise required separate policies and tens of hours of task-specific data. The research result is transfer of a common visual representation across environments, not one universal controller.

Four directions from this week are directly relevant to company systems: legal retrieval over a controlled document corpus, answers from internal knowledge, local speech and video processing, and model and hardware selection for a measured workload. pommeDeTerre can design such a system, test available open models on the client’s own material and compare local inference with APIs by quality, latency, access controls and cost. A useful pilot produces measurements and an architecture; it does not assume that every new release is equally suitable for every organization.