🤖 ModelBytes Digest
Thursday, October 01, 2026

━━━ CLOSED FRONTIER 🔒
Pareto 26.10 Preview — Multimodal composite built for research, coding, and agentic workflows with a 1M-token context window. Closed API, $0.80/$3.20 per 1M tokens. Released Oct 1. ⚡ API live · OpenRouter. → Source

📊 Surfaced 1 · scanned 6 today

🤖 ModelBytes Digest
Friday, October 02, 2026

Today's releases split between flagship frontier rollouts and smaller open/API utility models, with safety concerns pulling one major launch while builders get a free 560B MoE and an Apache-licensed multimodal.

━━━ CLOSED FRONTIER 🔒
Gemini 4 Argon — Google's flagship model tops Arena text at 1524.8 with a 15% hallucination rate, but rollout is just beginning. Released Sep 30. → Source

GPT-6 Luna Pro — OpenAI's Pro tier adds a 1,050,000-token context window at $0.10 input / $0.50 output per million tokens. Released Sep 30. ⚡ API live · OpenRouter. → Source

Ling 3.1 Flash — Free hybrid reasoning MoE with 25B active of 560B total and 262k context, aimed at cost-sensitive agent workloads. Released Oct 2. License Closed/API. ⚡ API live · OpenRouter. → Source

━━━ SPECIALIZED 🎯
Clef — Apache-licensed 27.4B image-text-to-text model, useful for builders needing a self-hostable multimodal base. Released Oct 2. License Apache-2.0. 📦 Open weights · HF. → Source

📊 Surfaced 4 · scanned 7 today

🤖 ModelBytes Digest
Saturday, October 03, 2026

Today's new models are small open multimodal/vision releases, not frontier — useful for local prototyping but not a step change.

━━━ OPEN FRONTIER 🔓
Clef Flash — 9.4B open image-text-to-text model under Apache 2.0, small enough to run locally while handling multimodal input. Released Oct 3 · 9.4B params · Apache 2.0. 📦 Open weights · HF. → Source

━━━ SPECIALIZED 🎯
VisionHOPE — MIT-licensed image classification model, no parameter count published. Released Oct 3 · MIT. 📦 Open weights · HF. → Source

📊 Surfaced 2 · scanned 5 today

🤖 ModelBytes Digest
Sunday, October 04, 2026

October opens without a frontier launch; the day's models are small and specialised — a sovereign German MoE and a Gemma-based diarization model.

━━━ OPEN FRONTIER 🔓
Kolibri 1 — Open-weight German-English MoE built for sovereign, regulated EU deployments rather than raw scale. MoE, open weights, German/English. Released Oct 3. 📦 Open weights. → Source

━━━ SPECIALIZED 🎯
DiarizationLM Gemma 4 E4B — Speaker diarization model on Gemma 4 E4B, Apache-2.0, 8B total params. 8B params, Apache-2.0. Released Oct 4. 📦 Open weights · HF. → Source

📊 Surfaced 2 · scanned 5 today

🤖 ModelBytes Digest
Monday, October 05, 2026

Open Apache-2.0 weights carried the day, and with dedicated deciders landing at both 2B and 26B, agent judgment is splitting off into its own model class — routing calls no longer need to burn frontier tokens.

━━━ OPEN FRONTIER 🔓
JEV 27B VL — A mid-size vision-language model for builders who need image understanding inside an open stack, with no API dependency. 27B total params · Apache-2.0 · Released Oct 3. 📦 Open weights · HF. → Source

OrcaSAQ-2 27B — OrcaRouter's second SAQ generation ships as a general-purpose text generator straight to open weights. 27B total params · Apache-2.0 · Released Oct 2. 📦 Open weights · HF. → Source

━━━ SPECIALIZED 🎯
GEV 26B Decide — A decision specialist that classifies instead of generating — judgment calls without a full LLM pass. 26B total params · Apache-2.0 · Decision Index suite 0.2 (edition 0.2.1): 62.48 · Released Oct 3. 📦 Open weights · HF. → Source

━━━ LOCAL 🏠
Strands Decider 2B — An agent decider shrunk to 2B so routing calls run on your own hardware, not a metered API. 2B total params · Apache-2.0 · JevBench public 0.7229 · boardgame 0.822 · contractnli 0.872 · Released Oct 1. 📦 Open weights · HF. → Source

📊 Surfaced 4 · scanned 12 today

🤖 ModelBytes Digest
Tuesday, October 06, 2026

Mistral Large 4 is the day's only frontier drop — everything else new is open-weight and narrow, so the practical move is wiring a cheap 512K-context API into your agent loop while picking up domain weights for video, ASR, and vision.

━━━ OPEN FRONTIER 🔓
Mistral Large 4 — Frontier-class reasoning, coding, and agentic work at $0.68/$2.09 per 1M tokens, with a 512K window that swallows whole repos. 524K context · multimodal (text + image input) · Released Oct 6. ⚡ API live · OpenRouter → Source

Holotron4 30B A3B — Sparse image-text-to-text model sized to fit a single node while still taking visual input. 30B total params · Released Sep 28. 📦 Open weights · HF → Source

ALoDLM 8B — Amazon's text generator, but cc-by-nc-4.0 keeps it out of commercial pipelines. 8B params · cc-by-nc-4.0 · Released Oct 2. 📦 Open weights · HF → Source

━━━ SPECIALIZED 🎯
LTX 2.5 — Lightricks' image-to-video successor, aimed at turning stills into motion without a render farm. Released Oct 2. 📦 Open weights · HF → Source

Cosmos-H Surgical Simulator — NVIDIA's image-to-video model built specifically for surgical simulation footage. Released Oct 6. 📦 Open weights · HF → Source

PixelUMM — NVIDIA's pixel-level image-to-text model for vision tasks that need spatial grounding, not just captions. Released Oct 1. 📦 Open weights · HF → Source

Phonon 2 — Permissively licensed ASR you can ship in a product without a legal review. cc-by-4.0 · Released Oct 1. 📦 Open weights · HF → Source

📊 Surfaced 7 · scanned 12 today

🤖 ModelBytes Digest
Wednesday, October 07, 2026

This cycle's releases are all narrow-purpose — image editing, speech, classification, and weather downscaling — rather than general chat models.

━━━ CLOSED FRONTIER 🔒

Gemini Nano Banana 2.1 — Google's image tier now handles product recontextualization better, which matters if you generate catalog or ad imagery where the object has to sit convincingly in a new scene. 66K context · $1.50/$7.50 per 1M · Released Oct 6. ⚡ API live · OpenRouter. → OpenRouter

━━━ ALSO TRACKED

Cohere Transcribe 03-2026 — Cohere's speech recognition entry gives builders a self-hostable transcription model to drop into their own audio pipeline instead of renting one per minute. 2.1B total · Apache-2.0 · Released Oct 7. 📦 Open weights · HF. → HF

PPLX Decider v1.1 27B — Perplexity's decider is built for routing and classification calls, so it can sit in front of a larger model as a gate that decides what actually gets sent downstream. 27B total · Apache-2.0 · Released Oct 5. 📦 Open weights · HF. → HF

NVIDIA CorrDiff Cosmo ERA5 — NVIDIA's CorrDiff line lands on Hugging Face for teams doing atmospheric downscaling work who want the model weights in their own pipeline. openmdw-1.1 · Released Oct 7. 📦 Open weights · HF. → HF

📊 Surfaced 4 · scanned 12 today

🤖 ModelBytes Digest
Thursday, October 08, 2026

Small multimodal models from LiquidAI and interfaze-ai land alongside new closed agentic endpoints from Anthropic and StepFun.

━━━ CLOSED FRONTIER 🔒
Claude Haiku 5.5 — Built for high-volume summarization, subagents, and browser automation where throughput matters more than peak reasoning. 1.0M context · $0.10/$0.50 per 1M · Released Oct 7. ⚡ API live · OpenRouter. → OpenRouter

━━━ SPECIALIZED 🎯
Step 5 Preview — A flagship agentic model aimed at long-horizon tool use and multi-step workflows. 1.0M context · $1.00/$2.70 per 1M · Released Oct 8. ⚡ API live · OpenRouter. → OpenRouter

d1-omni 600M — A compact image-text model for local multimodal understanding on constrained hardware. 600M total · Released Oct 7. 📦 Open weights · HF. → HF

d1-3B — A small multimodal model for local deployment and quick experimentation. 3B total · Released Oct 7. 📦 Open weights · HF. → HF

Interfaze 1 Lite — An image-to-image model for lightweight visual editing and generation pipelines. Released Oct 7. 📦 Open weights · HF. → HF

📊 Surfaced 5 · scanned 11 today

🤖 ModelBytes Digest
Friday, October 09, 2026

Two text-to-image releases and a document parser land in the same window, so open multimodal tooling keeps widening.

━━━ OPEN FRONTIER 🔓
Qwen-Image 2.1 Turbo — A faster cut of the Qwen image line for builders who want generation latency down without leaving the Qwen ecosystem. 7.1B total · Released Oct 9. 📦 Open weights · HF. → HF

━━━ SPECIALIZED 🎯
Youtu-Parsing-Omni — Converts document images straight into structured text, so parsing can live inside a multimodal pipeline instead of a separate OCR stack. 5.3B total · Released Oct 9. 📦 Open weights · HF. → HF

Iris 3B — Small enough to run on your own hardware for asset generation when you would rather not call a hosted image API. 3B total · Apache-2.0 · Released Oct 8. 📦 Open weights · HF. → HF

📊 Surfaced 3 · scanned 12 today

🤖 ModelBytes Digest
Saturday, October 10, 2026

Mistral's two Arabic releases split the job — streaming speech recognition on one side, language identification on the other.

━━━ OPEN FRONTIER 🔓
Voxtral Mini 4B Realtime Arabic — Gives Arabic voice agents a streaming transcription path instead of forcing batch audio uploads through a hosted API. 4B total · Released Oct 9. 📦 Open weights · HF. → HF

LIDstral Arabic — A small classifier for tagging Arabic text before it reaches a larger model, useful for routing and filtering pipelines. Apache-2.0 · Released Oct 9. 📦 Open weights · HF. → HF

━━━ SPECIALIZED 🎯
Strands Decider 26B A4B — A heavier decider for agent stacks whose routing calls have outgrown what the small variant can resolve. 26B total · Apache-2.0 · Released Oct 9. 📦 Open weights · HF. → HF

📊 Surfaced 3 · scanned 12 today

🤖 ModelBytes Digest
Sunday, October 11, 2026

━━━ CLOSED FRONTIER 🔒
GPT-6.1 Sol — Sol is the model Asana's browser agent runs on, making it the current reference point for teams wiring web-automation agents into OpenAI's stack. Released Oct 9. 🔗 Cited source. → openai.com

━━━ SPECIALIZED 🎯
LightOnOCR 3 0.8B — An image-to-text OCR model you can host yourself, so document extraction pipelines stop round-tripping every page through a hosted vision API. 0.8B total · Apache-2.0 · Released Oct 9. 📦 Open weights · HF. → HF

📊 Surfaced 2 · scanned 2 today

🤖 ModelBytes Digest
Monday, October 12, 2026

This cycle's releases are task-specific — video generation, classification, and compact multimodal — not general chat flagships.

━━━ SPECIALIZED 🎯
Multics — Multics generates video from a still image, giving builders a self-hostable path to animating product shots, storyboards, or ad creative. Apache-2.0 · Released Oct 12. 📦 Open weights · HF. → HF

JEV 35B — JEV 35B targets classification and decision tasks, so it can gate or route traffic without paying for a full chat model. 35B total · Apache-2.0 · Released Oct 11. 📦 Open weights · HF. → HF

H2O Lightning 4B — H2O Lightning 4B takes image-text input in a footprint small enough to co-locate with other services on one machine. 4B total · Apache-2.0 · Released Oct 12. 📦 Open weights · HF. → HF

📊 Surfaced 3 · scanned 6 today

🤖 ModelBytes Digest
Tuesday, October 13, 2026

Gemini 4 Argon arrives as both a general frontier model and a robotics-tuned variant, so one release covers agentic and embodied work.

━━━ OPEN FRONTIER 🔓
MiniCPM-V 4.7 35B A3B — Extending the MiniCPM-V line upward gives builders a heavier vision-language option for image and document understanding inside their own stack. 35B total · Apache-2.0 · Released Oct 13. 📦 Open weights · HF. → HF

MiniCPM-V 4.7 1B — The small end of the same release targets on-device and edge vision work where a hosted call is not an option. 1B total · Apache-2.0 · Released Oct 13. 📦 Open weights · HF. → HF

━━━ SPECIALIZED 🎯
Gemini 4 Argon — The newest flagship reasoning tier, aimed at the hardest multi-step problems and long-horizon agentic tasks. Released Oct 13. 🔗 Cited source. → deepmind.google

Muse Wonder — A frontier reasoning model built for agentic personal and professional work, taking text, images, video, audio, and PDFs through one endpoint. 1.0M context · $4.00/$20.00 per 1M · Released Oct 12. ⚡ API live · OpenRouter. → OpenRouter

Gemini 4 Argon for Robotics — A robotics-tuned cut of the same release, so embodied and manipulation stacks can target the new generation without waiting on a separate model. Released Oct 13. 🔗 Cited source. → deepmind.google

ScaleDown Summarize — Abstractive summarization as an endpoint, for when you need a faithful rewrite of long documents rather than lifted sentences. 1.0M context · $0.04/$0.00 per 1M · Released Oct 13. ⚡ API live · OpenRouter. → OpenRouter

ScaleDown Extract — Description-driven entity extraction that lets you define output fields in plain English instead of maintaining a fixed schema. 1.0M context · $0.04/$0.00 per 1M · Released Oct 13. ⚡ API live · OpenRouter. → OpenRouter

📊 Surfaced 7 · scanned 10 today

🤖 ModelBytes Digest
Wednesday, October 14, 2026

Ling 3.1 Flash and Olmo 3.1 32B Think both land as open-weight text models, one large and one mid-size.

━━━ OPEN FRONTIER 🔓
Ling 3.1 Flash — A permissively d open-weight text model gives builders a self-hostable option without legal review before serving. 572.1B total · MIT · Released Oct 14. 📦 Open weights · HF. → HF

━━━ SPECIALIZED 🎯
Olmo 3.1 32B Think — A reasoning model from a lab that publishes its training data and recipes, so the whole stack can be inspected and adapted. 32B total · Apache-2.0 · Released Oct 13. 📦 Open weights · HF. → HF

Drama 3 — A text-to-speech model for builders who need expressive long-form voice synthesis inside their own pipelines. 35.9B total · Released Oct 14. 📦 Open weights · HF. → HF

LFM2.5 24B A2B — A small-footprint text model sized for single-GPU and local serving where a hosted API is not an option. 24B total · Released Oct 13. 📦 Open weights · HF. → HF

📊 Surfaced 4 · scanned 12 today

🤖 ModelBytes Catch-up Wednesday, October 14, 2026

Last week brought a scoring model from Microsoft, a new leader on the voice-cloning leaderboard, and Z.ai's first model built for one profession.

━━━ SPECIALIZED

🎯 Microsoft-Decision-1 — Doesn't write text. It returns calibrated probabilities over fixed options, for routing, classification and guardrails. Post-trained from Qwen3.5-9B. 9B · 32K context · $0.042 per 1M input, output free · Released Oct 9.
⚡ API live · Foundry · OpenRouter. → OpenRouter (https://openrouter.ai/microsoft/microsoft-decision-1)

🎯 HeyGen Voice — Voice cloning that fits the scene: it reads the target image so the delivery and room sound match the video. Ranked #1 on the Artificial Analysis controlled-voice arena (Elo 1201, Oct 9 snapshot). Ahead of Qwen-Audio-3.1-TTS-Plus and Eleven v4 Turbo.
🔒 In HeyGen. → HeyGen Research (https://www.heygen.com/research/heygen-voice-scene-aware-voice-cloning)

🎯 GLM-Legal — Z.ai's legal model, built on GLM-5.3-Flash for contract review, evidence analysis and compliance work. 19.2% all-pass on Harvey LAB, checked independently by Vals. Released Oct 12.
🔒 Application-only beta, AutoClaw support coming. → Z.ai (https://z.ai/blog/glm-legal)

📊 Catch-up · 3 items