Multimodal models go small and single-stack: Reka's 19B Rho-1, Liquid's d1 and Interfaze's MoA

Reka's Rho-1 is the newest multimodal release, published on 5 October. The company calls it an omni-reasoning model trained from scratch at 19B parameters. The pitch is architectural. Instead of a central model handing tasks to separate vision, video and action specialists, Rho-1 puts text, vision and robotic actions in a single context window.

"Most AI systems today are agentic pipelines: a central model plans and delegates, handing jobs off to modality-specific specialists," the Reka lab post says. "Rho-1 collapses that stack." The post says one unedited five-turn session has the model draw a scene, locate an object, animate it into video, edit the weather and explain what changed.

The numbers Reka publishes are speed figures, not benchmark tables. The base model generates video at 0.79x real time, median, according to the post. A watchable stream starts in roughly 6 seconds. A distilled variant cuts the denoising trajectory from 99 steps to 8. Reka says that in internal testing it returned a 5.3-second clip in about a second.

Small models, narrow jobs, cheaper bills

Liquid AI used the same day to launch d1, its first decision model, now with image support. Decision models return probabilities instead of generating tokens: a yes/no answer, a label, or a score. Liquid says a text decision takes 200 to 300 ms.

The company tested d1 against GPT-6.1 Sol and Claude Opus 5.5 on six applications, from filtering support tickets to inspecting circuit boards. According to the Liquid blog, d1 matches or beats GPT-6.1 Sol on four of them. It costs 19x to 200x less than both models and answers faster on every task. Those are vendor-run comparisons and the post gives no per-task breakdown.

Some of the demos are more concrete. On a visual inspection task using the public VisA dataset, d1 sorts good and defective parts from four production lines with 85-97% accuracy. Liquid says the model was never trained for that job. In a Wordle test it read the board from a screenshot and solved 12 of 12 games in 3.8 guesses on average. In Quick, Draw! it recognized 5.2 of 6 doodles among 62 words; random guessing gets 0.6.

Interfaze took a third route. Its Interfaze 1.1 Lite, posted on Hugging Face on 5 October, is not one network at all. It is a reasoning core plus specialist components for document reading, line geometry, layout, speech, speaker diarization, segmentation, forecasting and safety, joined by tool calls. The company says the whole thing runs on one 80 GB GPU with no external services.

The specs are unusually specific for a model card. PDFs up to 50 pages per call, with a box and confidence for every line and word. A 95-minute recording transcribed in about 90 seconds. A 131k-token context. On the card's own benchmark table, Interfaze 1.1 Lite scores 85.9 on GPQA Diamond and 73.2 on MMMU-Pro. That compares with 71.1 for Interfaze, 40.4 for GPT-5.4-Mini, 46.3 for Claude-Sonnet-4.6 and 67.6 for Gemini-3-Flash. On GPQA Diamond, though, it trails those same models at 82.8, 89.9, 88.5 and 73.6 respectively.

The open-weight frontier gets another entrant

Reflection's Beam is the largest open-weight claim here: a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, built for coding, reasoning and agentic work. The company's 5 October post says it was pretrained on 23.8 trillion tokens. It then went through a reinforcement learning run that generated over 100 million rollouts on 10.5K NVIDIA GB300 GPUs across 4 weeks.

The weights are not out yet. Beam is "undergoing final red-teaming and evaluations", and Reflection says it will release the weights, technical report, model card and developer artifacts later this month. Early access is by signup. That makes every benchmark in the post a vendor claim about an unreleased model.

The published table is mixed. Beam scores 80.9 on SWEBench Verified and 80.1 on Terminal Bench v2.1. That is ahead of GLM 5.2 at 81.0 on Terminal Bench but behind Kimi K3 at 88.3 and DeepSeek V4.1 Flash at 90.6. On the hardest coding row, SWE Bench Pro v2-Hard, Beam reports 77.2 against Kimi K3's 88.2. Reflection's own framing is efficiency rather than raw capability. It says Beam reaches scores comparable to GLM-5.2 on advanced reasoning while using 3 to 4 times less inference compute, and calls Kimi K3 ahead on raw capability.

The efficiency argument is the interesting one. If a 23B-active model really lands within a few points of much larger systems on agentic tasks, the cost per task falls for anyone running these workloads at volume. Reflection has not published the cost figures behind that claim. The weights are not available for independent testing.

Harnesses and decision layers around the models

Two smaller projects show where the engineering effort is going: not into bigger models, but into the glue around them.

Jeff-Code, a fork of the Pi coding agent published on GitHub on 5 October, inserts a 0.8B decision model called Jeff into the agent loop around Qwen 3.8-27B. Jeff makes two calls per Qwen turn, about 0.2 seconds each. It asks whether it can take the next step itself, and whether Qwen needs to think hard this turn. Thinking stays off unless Jeff's probability that the turn needs full thinking reaches 0.6. About three quarters of Qwen's turns ran with thinking off in the evaluation.

The measured result is modest on quality and large on time. Across 1,242 paired tasks, pass rate was 62.4% with Jeff-Code against 62.8% for Qwen alone. That is a paired difference of minus 0.2 points with a 95% interval of minus 2.6 to plus 2.1. Time per task averaged 0.68x, with a median of 0.70x; total time across all tasks was 0.86x. The project reports no clear speed-up on Terminal-Bench 2.0 (0.96x) or SkillsBench (0.91x). Switching thinking off on every turn was faster still but cost 7.6 points.

Reika, another coding agent CLI published the same day, is built the other way round. It is designed for small local models first, in the 8B to 35B range, often at Q2 to Q4 quantization on a 16 to 32k context window. Its README is candid about the limits. In testing, 14-35B models handled multi-file tasks most reliably, 8-9B models managed simple, well-scoped ones, and below 7B rarely got far. The project also publishes what its harness cannot fix. On a vague task at Q2 a model converges or spirals about 50/50, and prompt wording is the weakest lever.

The context numbers in that README are a useful reality check for anyone sizing agent workloads. One mid-context rewrite re-processed 8,453 tokens of prompt, 7.9 minutes of prefill. An append in the same session cost 25 tokens and 3.5 seconds.

What the week says about the multimodal race

Three releases on one day point in different directions. Reka bets on one network swallowing every modality. Interfaze bets on a small core routing to specialists, and keeps the whole thing on a single GPU. Liquid bets that most multimodal work is really a classification problem and should never generate a token at all.

The common thread is cost. Reka's distilled variant, Liquid's 19x to 200x price claim, Interfaze's one-GPU deployment and Jeff-Code's 0.68x time per task all sell the same thing: the same job for less compute. None of these claims has been independently verified yet.

Outside the model labs, the same pattern shows in infrastructure work. mold 3.0.0, released on 5 October, is the first release of the high-speed linker rewritten from C++ to Rust, a change the project announced in its 2.42.1 notes. The release notes say linking performance remains on par with 2.42.1. They also say the project verified compatibility by building all Gentoo packages with no regressions. The goal for mold 3.x is closing the remaining gaps with GNU ld, particularly linker script support, to pave the way for adoption as the default linker in Linux distributions.

Expat 2.9.0, also released on 5 October, fixes two vulnerabilities. CVE-2026-77214 is a potential out-of-bounds read in XML_ParseBuffer due to insufficient validation of the length parameter. CVE-2026-102633 is an integer overflow in memory management code affecting 32-bit platforms. The release adds a Properties API and deprecates the XML_LARGE_SIZE macro in favour of five 64-bit location functions. The maintainer credits the City of Munich's Open Source Sabbatical for funding the work.

Elsewhere, OpenAI said it will begin testing visual ads inside ChatGPT during image generation later this month in the U.S., with an initial group of advertisers, according to BleepingComputer. The company says ads will be clearly labeled and remain separate from the image being created. It also says ChatGPT now reaches 1.2 billion people every week.

On the research side, a UC Riverside study published on 23 September found that confidence and correctness in large language models arise from different internal features. The team studied Meta's Llama-3.1-8B and Google's Gemma-2-9B with sparse autoencoders and suppressed selected features. "The main assumption in the field is that when the model is confident, it is likely to be correct, and when the model is unsure, it is more likely to be incorrect," said Het Patel, a UCR computer science doctoral student and lead author. "But we often see the counterexamples that are well documented."

For buyers, the practical question this week is not which model is smartest. It is which of these architectures can be tested. Reka's weights are weeks away, Liquid's comparisons are vendor-run, and Interfaze's benchmark table is its own. The numbers to watch are the ones that arrive with artifacts attached.