Reflection's Beam joins the open-weight race, but the dossier has no video model release

The newest release in the dossier is not a video generator. Reflection's blog post, dated 5 October, describes Beam as its first open-weight model: a sparse Mixture-of-Experts design with 501 billion total parameters and 23 billion active, trained for coding, reasoning and agentic workloads.

The weights are not out yet. Reflection says Beam is still in final red-teaming and evaluations, and that the weights, technical report, model card and developer artifacts will follow later in October. That gap matters for anyone searching for an AI video generation launch this week. The dossier carries no dated announcement of a new text-to-video or image-to-video model from the last three days. The video generation items it does contain are older: an arXiv paper on optical deepfake detection submitted in May, a Kuaishou item from 29 September in the headline list only, and a set of undated or out-of-window product posts. None of them is a fresh release.

What the dossier does hold is a cluster of model launches and agent tooling from 5 October. Most of it is small, cheap and specialised. Liquid AI introduced d1, a decision model that handles text and images and returns probabilities instead of generating tokens. Interfaze published an open-weight model built from a reasoning core plus specialist components. And GitHub filled up with coding agents that sit around an existing model rather than replacing it.

Decision models, not video models

Liquid AI's d1 is the clearest example of the trend. The company says a text decision takes 200 to 300 ms, because the model reads the input in one forward pass and returns probabilities without generating tokens. It tested d1 against GPT-6.1 Sol and Claude Opus 5.5 on six applications, from filtering support tickets to inspecting circuit boards, and says d1 matches or beats GPT-6.1 Sol on four of them while costing 19x to 200x less.

Those are Liquid's own numbers, published in its launch post. The vision claim is the interesting one. Liquid says d1 was never trained for industrial inspection, yet sorted good and defective parts at 85 to 97% accuracy on the public VisA dataset, across circuit boards, candles, cashews and chewing gum. It also reports that adding a screen raised its Tetris score from 70 to 81 cleared lines, and that it solved 12 of 12 Wordle games in 3.8 guesses on average. Independent verification is not in the dossier, so treat the figures as vendor-reported.

Interfaze's release takes a different route to the same cost argument. According to its model page, Interfaze 1.1 Lite is a mixture of architectures: one vision-language reasoning core that plans and composes answers, plus specialists for document reading, line geometry, layout, speech, speaker diarization, segmentation, time-series forecasting and safety classification. The whole thing runs on a single 80 GB GPU with no external services, the company says.

Its benchmark table puts the model at 85.9 on GPQA Diamond, 87.8 on MMMLU, 73.2 on MMMU-Pro and 83.8 on olmOCR-Bench, against 92.4, 90.9, 71.1 and 85.7 for Interfaze's own comparison figure for GPT-5.4-Mini. Interfaze also lists a 95-minute recording transcribing in about 90 seconds, and structured output constrained to a supplied JSON schema. Again, these are the publisher's numbers.

Agents that route around the big model

The coding agent releases on 5 October are about squeezing more out of models that already exist. Jeff-Code, a fork of Mario Zechner's Pi agent, inserts a 0.8B decision model called Jeff into the loop around Qwen 3.8-27B. Jeff makes two calls of about 0.2 seconds each: whether it can take the next step itself, and whether Qwen needs to think hard on this turn. Writing and editing files stay with Qwen.

The repository's evaluation reports 62.4% pass rate for Jeff-Code against 62.8% for Qwen alone, a paired difference of minus 0.2 points with a 95% interval of minus 2.6 to plus 2.1 across 1,242 paired tasks. Time per task averaged 0.68x, with a median of 0.70x. Total time across all tasks fell less, to 0.86x, because in about 5% of tasks Jeff-Code runs more than 30 minutes longer than Qwen alone. The author also reports that turning thinking off on every turn is faster still but costs 7.6 points.

Reika takes the opposite bet: a coding agent CLI designed for small local models first, in the 8B to 35B range, often at Q2 to Q4 quantisation on a 16 to 32k context window. Its published findings are unusually candid. One mid-context rewrite re-processed 8,453 tokens of prompt, which the author measures as 7.9 minutes of prefill, where an append in the same session cost 25 tokens and 3.5 seconds. The tool is pre-1.0 and its author warns commands and behaviour can change between releases.

The harness's own context management cost more than the model did.

Neither project is a frontier model release, and neither claims to be. They are evidence of where a lot of engineering effort is going: not into bigger generators, but into cheaper routing, smaller decision layers and tighter context handling around models that already work.

Infrastructure and the older material

Two other 5 October items fill out the picture. Microsoft's Windows Developer Blog, in a post dated 29 September and surfaced in the dossier on 5 October, says WSL containers is now generally available, with a wslc.exe CLI, a containers API for native Windows apps, and Intune and Defender for Endpoint coverage extended to container workflows. Separately, the mold linker shipped version 3.0.0 on 5 October, rewritten from C++ to Rust, with the project saying it verified compatibility by building all Gentoo packages and found no regressions.

The video generation material in the dossier is thin and mostly stale. An arXiv paper by Parnian Ghapandar Kashani, Shiqi Chen and Aydogan Ozcan, submitted 19 May and published in eLight, describes a hybrid optical-digital deepfake detector that processed 15 or more video streams in a single optical pass, reaching 97.79% average accuracy on Celeb-DF. That is detection, not generation. The recent-headlines block mentions Kuaishou's Kling spinoff unveiling a new model on 29 September, but the dossier gives only a headline, with no figures or company statement to work from.

Reflection's Beam is the largest model announcement in the dossier, and its benchmark table is the most detailed. On SWE Bench Pro v2-Hard it lists 77.2 against 84.3 for GLM 5.3 and 88.2 for Kimi K3. On Terminal Bench v2.1 it lists 80.1 against 88.2 and 88.3. On AIME 2026 it lists 97.8, just below GLM 5.2 at 99.2. Reflection says Beam is competitive with larger open models and approaches Qwen 3.8-Max on coding and agentic tasks, while its advantage is inference efficiency, claiming comparable reasoning scores to GLM-5.2 at 3 to 4 times less inference compute.

The training numbers are specific: 23.8 trillion tokens of pretraining data, and a reinforcement learning run generating over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over four weeks. Reflection says the weights will follow this month. Until they do, every Beam figure is a claim from the company that built it, and the same is true of d1, Interfaze and the agent benchmarks above.

For readers looking for a video generation release this week, the honest answer is that the dossier does not contain one. The closest dated items are older or headline-only. The freshest, best-documented news on 5 October is about models that decide, route and transcribe, not models that render motion.