🤖 ModelBytes Digest
Monday, September 14, 2026
Today's pattern is infrastructure over raw scale: a production voice API, a 753B open-weight preview, and an attention-sparsification model each give builders a different lever for latency, cost, or control.
━━━ OPEN FRONTIER 🔓
Atria Dawn Preview — A 753B-parameter open-weight preview under MIT, sized for frontier-scale experimentation without a closed API. Released Sep 11 · 753.3B total params · MIT license. 📦 Open weights · HF. → Source
━━━ CLOSED FRONTIER 🔒
GPT-Live-1 — Natural voice experiences in the API, targeting real-time speech interactions for production voice agents. Released Sep 10. ⚡ API live · OpenAI. → Source
━━━ SPECIALIZED 🎯
Simple Attention Sparsification — Text-generation model from Tencent's attention-sparsification work, released as an open-weight artifact. Released Sep 14 · text-generation. 📦 Open weights · HF. → Source
📊 Surfaced 3 · scanned 12 today
Monday, September 14, 2026
Today's pattern is infrastructure over raw scale: a production voice API, a 753B open-weight preview, and an attention-sparsification model each give builders a different lever for latency, cost, or control.
━━━ OPEN FRONTIER 🔓
Atria Dawn Preview — A 753B-parameter open-weight preview under MIT, sized for frontier-scale experimentation without a closed API. Released Sep 11 · 753.3B total params · MIT license. 📦 Open weights · HF. → Source
━━━ CLOSED FRONTIER 🔒
GPT-Live-1 — Natural voice experiences in the API, targeting real-time speech interactions for production voice agents. Released Sep 10. ⚡ API live · OpenAI. → Source
━━━ SPECIALIZED 🎯
Simple Attention Sparsification — Text-generation model from Tencent's attention-sparsification work, released as an open-weight artifact. Released Sep 14 · text-generation. 📦 Open weights · HF. → Source
📊 Surfaced 3 · scanned 12 today
🤖 ModelBytes Digest
Tuesday, September 15, 2026
Today's releases split between Google's agent-cost push and Moonshot's 2.8T API debut, with a small Apache-2.0 local model filling the self-host gap.
━━━ CLOSED FRONTIER 🔒
Kimi K3 — Moonshot's 2.8T-parameter model hits Ropewalk, leading a 24-model weekly batch. 2.8T params. ⚡ API live · Ropewalk. → Source
Gemini 3.8 — Google's latest frontier release, framed by AI Business as a market-shift signal. → Source
Gemini 3.7 Flash — Positioned for lower-cost agent building, per Google's August update. → Source
━━━ LOCAL 🏠
NeoHorse 1 4B — A 4B Apache-2.0 text-generation model for local experimentation. Released Sep 15 · 4B params · Apache-2.0. 📦 Open weights · HF. → Source
📊 Surfaced 4 · scanned 2 today
Tuesday, September 15, 2026
Today's releases split between Google's agent-cost push and Moonshot's 2.8T API debut, with a small Apache-2.0 local model filling the self-host gap.
━━━ CLOSED FRONTIER 🔒
Kimi K3 — Moonshot's 2.8T-parameter model hits Ropewalk, leading a 24-model weekly batch. 2.8T params. ⚡ API live · Ropewalk. → Source
Gemini 3.8 — Google's latest frontier release, framed by AI Business as a market-shift signal. → Source
Gemini 3.7 Flash — Positioned for lower-cost agent building, per Google's August update. → Source
━━━ LOCAL 🏠
NeoHorse 1 4B — A 4B Apache-2.0 text-generation model for local experimentation. Released Sep 15 · 4B params · Apache-2.0. 📦 Open weights · HF. → Source
📊 Surfaced 4 · scanned 2 today
🤖 ModelBytes Digest
Wednesday, September 16, 2026
Today’s releases push builders toward real-time multimodal interaction: voice models that reason aloud, a free multimodal agentic API, and a robotics pose model.
━━━ CLOSED FRONTIER 🔒
Union Alpha — Multimodal model built for research, coding, and agentic workflows with frontier-level general performance. 262k context, closed/API, free. ⚡ API live · OpenRouter. → Source
Gemini 3.8 Live — Real-time speech-to-speech with reasoning that speaks as it works; Extended Thinking variant tops speech quality. Released Sep 16. ⚡ API live. → Source
━━━ SPECIALIZED 🎯
FoundationPose — Robotics pose estimation model from NVIDIA. Released Sep 15. 📦 Open weights · HF. → Source
📊 Surfaced 3 · scanned 4 today
Wednesday, September 16, 2026
Today’s releases push builders toward real-time multimodal interaction: voice models that reason aloud, a free multimodal agentic API, and a robotics pose model.
━━━ CLOSED FRONTIER 🔒
Union Alpha — Multimodal model built for research, coding, and agentic workflows with frontier-level general performance. 262k context, closed/API, free. ⚡ API live · OpenRouter. → Source
Gemini 3.8 Live — Real-time speech-to-speech with reasoning that speaks as it works; Extended Thinking variant tops speech quality. Released Sep 16. ⚡ API live. → Source
━━━ SPECIALIZED 🎯
FoundationPose — Robotics pose estimation model from NVIDIA. Released Sep 15. 📦 Open weights · HF. → Source
📊 Surfaced 3 · scanned 4 today
🤖 ModelBytes Digest
Thursday, September 17, 2026
Today's open-weight drop pairs a 27B multimodal generalist with two compact Apache-2.0 document models — builders get fresh local options for vision and document workflows without API lock-in.
━━━ OPEN FRONTIER 🔓
Swift Qwen3.8 27B — 27B image-text-to-text model released today, open weights for multimodal builders. 27B params · Released Sep 17. 📦 Open weights · HF. → Source
━━━ SPECIALIZED 🎯
WeVisDoc 4B — Apache-2.0 document vision model at 4B parameters, aimed at document understanding tasks. 4B params · Apache-2.0 · Released Sep 16. 📦 Open weights · HF. → Source
WeVisDoc 2B — Smaller 2B sibling for lighter document vision workloads, same Apache-2.0 license. 2B params · Apache-2.0 · Released Sep 16. 📦 Open weights · HF. → Source
📊 Surfaced 3 · scanned 5 today
Thursday, September 17, 2026
Today's open-weight drop pairs a 27B multimodal generalist with two compact Apache-2.0 document models — builders get fresh local options for vision and document workflows without API lock-in.
━━━ OPEN FRONTIER 🔓
Swift Qwen3.8 27B — 27B image-text-to-text model released today, open weights for multimodal builders. 27B params · Released Sep 17. 📦 Open weights · HF. → Source
━━━ SPECIALIZED 🎯
WeVisDoc 4B — Apache-2.0 document vision model at 4B parameters, aimed at document understanding tasks. 4B params · Apache-2.0 · Released Sep 16. 📦 Open weights · HF. → Source
WeVisDoc 2B — Smaller 2B sibling for lighter document vision workloads, same Apache-2.0 license. 2B params · Apache-2.0 · Released Sep 16. 📦 Open weights · HF. → Source
📊 Surfaced 3 · scanned 5 today
🤖 ModelBytes Digest
Friday, September 18, 2026
Today's releases cluster around efficiency and modality coverage: a 5.9 GB ternary Qwen derivative, an any-to-any realtime open model, and a multimodal composite API all target builders who need frontier behavior without the usual footprint.
━━━ OPEN FRONTIER 🔓
Realtime Venus — Any-to-any realtime model under Apache 2.0 for low-latency multimodal pipelines. Released Sep 16 · Apache 2.0. 📦 Open weights · HF. → Source
━━━ CLOSED FRONTIER 🔒
Pareto — Multimodal composite for research, coding, and agentic workflows with 262K context. 262K context · Closed/API · $2.50/$7.50 per 1M. Released Sep 17. ⚡ API live · OpenRouter. → Source
━━━ SPECIALIZED 🎯
Qwen3.8 Omni Flash — Alibaba's flash omni variant for low-latency multimodal tasks. Released Sep 18. → Source
━━━ LOCAL 🏠
Ternary Bonsai 2 27B — Ternary-weight compression of Qwen3.8 27B at 1.76 effective bits per weight, keeping 98.2% of base aggregate score in 5.9 GB. 27B params · 262K context · text+image input · Apache 2.0. Released Sep 17. 📦 Open weights. → Source
📊 Surfaced 4 · scanned 8 today
Friday, September 18, 2026
Today's releases cluster around efficiency and modality coverage: a 5.9 GB ternary Qwen derivative, an any-to-any realtime open model, and a multimodal composite API all target builders who need frontier behavior without the usual footprint.
━━━ OPEN FRONTIER 🔓
Realtime Venus — Any-to-any realtime model under Apache 2.0 for low-latency multimodal pipelines. Released Sep 16 · Apache 2.0. 📦 Open weights · HF. → Source
━━━ CLOSED FRONTIER 🔒
Pareto — Multimodal composite for research, coding, and agentic workflows with 262K context. 262K context · Closed/API · $2.50/$7.50 per 1M. Released Sep 17. ⚡ API live · OpenRouter. → Source
━━━ SPECIALIZED 🎯
Qwen3.8 Omni Flash — Alibaba's flash omni variant for low-latency multimodal tasks. Released Sep 18. → Source
━━━ LOCAL 🏠
Ternary Bonsai 2 27B — Ternary-weight compression of Qwen3.8 27B at 1.76 effective bits per weight, keeping 98.2% of base aggregate score in 5.9 GB. 27B params · 262K context · text+image input · Apache 2.0. Released Sep 17. 📦 Open weights. → Source
📊 Surfaced 4 · scanned 8 today
🤖 ModelBytes Digest
Saturday, September 19, 2026
Today's concrete releases split between cheap closed multimodal speed and open-weight mid-size models, while frontier labs posture rather than ship.
━━━ OPEN FRONTIER 🔓
Xing4.0 29B A4B — 29B-parameter text-generation model under Apache-2.0, a mid-size open-weight option for builders needing a permissive base. Released Sep 19. 29B total params, Apache-2.0. 📦 Open weights · HF. → Source
Muse Glimmer 30B — Meta's 30B open-weight model from the September briefing, adding another mid-size option to the open ecosystem. 30B params, open weights. 📦 Open weights. → Source
━━━ CLOSED FRONTIER 🔒
GLM 5.3 FlashX — High-speed variant of GLM-5.3-Flash, native multimodal, up to 200 tokens/s. Released Sep 18. 1.0M context, Closed/API, $0.37/$1.25 per 1M. ⚡ API live · OpenRouter. → Source
━━━ SPECIALIZED 🎯
OrcaVerify Text 1.0 — Text verification model priced at $2.00/$0.00 per 1M tokens. Released Sep 16. $2.00/$0.00 per 1M tokens. ⚡ API live. → Source
📊 Surfaced 4 · scanned 8 today
Saturday, September 19, 2026
Today's concrete releases split between cheap closed multimodal speed and open-weight mid-size models, while frontier labs posture rather than ship.
━━━ OPEN FRONTIER 🔓
Xing4.0 29B A4B — 29B-parameter text-generation model under Apache-2.0, a mid-size open-weight option for builders needing a permissive base. Released Sep 19. 29B total params, Apache-2.0. 📦 Open weights · HF. → Source
Muse Glimmer 30B — Meta's 30B open-weight model from the September briefing, adding another mid-size option to the open ecosystem. 30B params, open weights. 📦 Open weights. → Source
━━━ CLOSED FRONTIER 🔒
GLM 5.3 FlashX — High-speed variant of GLM-5.3-Flash, native multimodal, up to 200 tokens/s. Released Sep 18. 1.0M context, Closed/API, $0.37/$1.25 per 1M. ⚡ API live · OpenRouter. → Source
━━━ SPECIALIZED 🎯
OrcaVerify Text 1.0 — Text verification model priced at $2.00/$0.00 per 1M tokens. Released Sep 16. $2.00/$0.00 per 1M tokens. ⚡ API live. → Source
📊 Surfaced 4 · scanned 8 today
🤖 ModelBytes Digest
Sunday, September 20, 2026
━━━ SPECIALIZED 🎯
Qwen Image 2.1 PE-I2I — A 9.4B-parameter image-to-image model extending Qwen’s open image stack for editing workflows. Released Sep 20 · 9.4B params. 📦 Open weights · HF. → Source
📊 Surfaced 1 · scanned 3 today
Sunday, September 20, 2026
━━━ SPECIALIZED 🎯
Qwen Image 2.1 PE-I2I — A 9.4B-parameter image-to-image model extending Qwen’s open image stack for editing workflows. Released Sep 20 · 9.4B params. 📦 Open weights · HF. → Source
📊 Surfaced 1 · scanned 3 today
🤖 ModelBytes Digest
Monday, September 21, 2026
Open-weight image and world models land today, giving builders local generation options across stills and spatial video.
━━━ SPECIALIZED 🎯
Qwen Image 2.1 — 7.1B text-to-image base now open, separate from the PE-I2I edit variant; a direct still-image generator for local pipelines. Released Sep 21 · 7.1B total params. 📦 Open weights · HF. → Source
WorldCrafter Base — 14.3B-param world model for generating consistent video from text or image prompts; the larger option for quality-first pipelines. Released Sep 21 · 14.3B total params. 📦 Open weights · HF. → Source
WorldCrafter Fast — Speed-tuned variant of the same world-model family, aimed at real-time or high-throughput generation where latency beats maximum fidelity. Released Sep 21. 📦 Open weights · HF. → Source
📊 Surfaced 3 · scanned 6 today
Monday, September 21, 2026
Open-weight image and world models land today, giving builders local generation options across stills and spatial video.
━━━ SPECIALIZED 🎯
Qwen Image 2.1 — 7.1B text-to-image base now open, separate from the PE-I2I edit variant; a direct still-image generator for local pipelines. Released Sep 21 · 7.1B total params. 📦 Open weights · HF. → Source
WorldCrafter Base — 14.3B-param world model for generating consistent video from text or image prompts; the larger option for quality-first pipelines. Released Sep 21 · 14.3B total params. 📦 Open weights · HF. → Source
WorldCrafter Fast — Speed-tuned variant of the same world-model family, aimed at real-time or high-throughput generation where latency beats maximum fidelity. Released Sep 21. 📦 Open weights · HF. → Source
📊 Surfaced 3 · scanned 6 today
🤖 ModelBytes Digest
Tuesday, September 22, 2026
Frontier price compression and long context dominate: Grok 4.7 undercuts on coding, Xiaomi ships a 1M-context lineup from $0.14 to $0.87 per 1M, and an open 6.2B image model drops on HF.
━━━ CLOSED FRONTIER 🔒
Grok 4.7 — Flagship coding and agentic model succeeding Grok 4.6, with DeepSWE at 71%. 500k context, closed API, $1.60/$4.80 per 1M on OpenRouter. ⚡ API live · OpenRouter. → Source
MiMo V2.6 Pro — Xiaomi's 1T+ parameter flagship with 1M context for demanding tasks. 1M context, closed API, $0.43/$0.87 per 1M. ⚡ API live · OpenRouter. → Source
MiMo V2.6 Flash — 309B MoE with 15B active, 1M context at $0.14/$0.28 per 1M. 309B total/15B active, 1M context, closed API. ⚡ API live · OpenRouter. → Source
Antigravity Agent 09-2026 — Google's updated Antigravity agent preview, replacing the 05-2026 build. ⚡ API live · Google AI. → Source
━━━ SPECIALIZED 🎯
Ming Image 0.1 Design Layer — Open 6.2B image-text-to-image model for design-layer generation. MIT, 6.2B params. 📦 Open weights · HF. → Source
📊 Surfaced 5 · scanned 12 today
Tuesday, September 22, 2026
Frontier price compression and long context dominate: Grok 4.7 undercuts on coding, Xiaomi ships a 1M-context lineup from $0.14 to $0.87 per 1M, and an open 6.2B image model drops on HF.
━━━ CLOSED FRONTIER 🔒
Grok 4.7 — Flagship coding and agentic model succeeding Grok 4.6, with DeepSWE at 71%. 500k context, closed API, $1.60/$4.80 per 1M on OpenRouter. ⚡ API live · OpenRouter. → Source
MiMo V2.6 Pro — Xiaomi's 1T+ parameter flagship with 1M context for demanding tasks. 1M context, closed API, $0.43/$0.87 per 1M. ⚡ API live · OpenRouter. → Source
MiMo V2.6 Flash — 309B MoE with 15B active, 1M context at $0.14/$0.28 per 1M. 309B total/15B active, 1M context, closed API. ⚡ API live · OpenRouter. → Source
Antigravity Agent 09-2026 — Google's updated Antigravity agent preview, replacing the 05-2026 build. ⚡ API live · Google AI. → Source
━━━ SPECIALIZED 🎯
Ming Image 0.1 Design Layer — Open 6.2B image-text-to-image model for design-layer generation. MIT, 6.2B params. 📦 Open weights · HF. → Source
📊 Surfaced 5 · scanned 12 today
🤖 ModelBytes Digest
Wednesday, September 23, 2026
This week's pattern is tiered expansion: labs are filling out product lines below their flagships with cost-optimized variants, while DeepSeek ships a genuinely new base model under a point-release name.
━━━ OPEN FRONTIER 🔓
DeepSeek V4.1 Flash — A brand-new base model wearing a point-release number, not a fine-tune of V4. Released Sep 10. ⚡ API live · OpenRouter. → Source
━━━ CLOSED FRONTIER 🔒
Claude Opus 5.5 — Anthropic's flagship for demanding reasoning, coding, and long-horizon agentic work, succeeding Opus 5. 1.0M context, $4.00/$20.00 per 1M. ⚡ API live · OpenRouter. → Source
GPT-6 Sol — Cost-efficient high-end tier in the GPT-6 series, positioned below Astra and above Luna. 1.1M context, $2.00/$10.00 per 1M. ⚡ API live · OpenRouter. → Source
GPT-6 Luna — Fast, cost-efficient GPT-6 tier for high-volume and latency-sensitive workloads. 1.1M context, $0.10/$0.50 per 1M. ⚡ API live · OpenRouter. → Source
Command A+ — Cohere's flagship for enterprise agentic workflows with native tool calling and strict schemas. 192K context, $0.30/$1.50 per 1M. ⚡ API live · OpenRouter. → Source
━━━ SPECIALIZED 🎯
Aion 3.5 — Multi-model roleplaying and storytelling system built on GLM, using collaborative generation across specialized models. 262K context, $3.00/$6.00 per 1M. ⚡ API live · OpenRouter. → Source
Rho-RoboEval — Microsoft's 5.2B model for robotics evaluation. MIT license. 📦 Open weights · HF. → Source
📊 Surfaced 7 · scanned 12 today
Wednesday, September 23, 2026
This week's pattern is tiered expansion: labs are filling out product lines below their flagships with cost-optimized variants, while DeepSeek ships a genuinely new base model under a point-release name.
━━━ OPEN FRONTIER 🔓
DeepSeek V4.1 Flash — A brand-new base model wearing a point-release number, not a fine-tune of V4. Released Sep 10. ⚡ API live · OpenRouter. → Source
━━━ CLOSED FRONTIER 🔒
Claude Opus 5.5 — Anthropic's flagship for demanding reasoning, coding, and long-horizon agentic work, succeeding Opus 5. 1.0M context, $4.00/$20.00 per 1M. ⚡ API live · OpenRouter. → Source
GPT-6 Sol — Cost-efficient high-end tier in the GPT-6 series, positioned below Astra and above Luna. 1.1M context, $2.00/$10.00 per 1M. ⚡ API live · OpenRouter. → Source
GPT-6 Luna — Fast, cost-efficient GPT-6 tier for high-volume and latency-sensitive workloads. 1.1M context, $0.10/$0.50 per 1M. ⚡ API live · OpenRouter. → Source
Command A+ — Cohere's flagship for enterprise agentic workflows with native tool calling and strict schemas. 192K context, $0.30/$1.50 per 1M. ⚡ API live · OpenRouter. → Source
━━━ SPECIALIZED 🎯
Aion 3.5 — Multi-model roleplaying and storytelling system built on GLM, using collaborative generation across specialized models. 262K context, $3.00/$6.00 per 1M. ⚡ API live · OpenRouter. → Source
Rho-RoboEval — Microsoft's 5.2B model for robotics evaluation. MIT license. 📦 Open weights · HF. → Source
📊 Surfaced 7 · scanned 12 today
🤖 ModelBytes Digest
Thursday, September 24, 2026
Robotics gets its own foundation-model moment while API providers ship throughput variants of existing frontier models.
━━━ SPECIALIZED 🎯
Flux 3 Action SO101 — Black Forest Labs' first robotics action model, targeting single-object manipulation with a 6.9B-parameter architecture. Released Sep 22 · 6.9B params. 📦 Open weights · HF. → Source
Flux 3 Action Droid — Companion robotics model for broader manipulation tasks, same 6.9B footprint as SO101. Released Sep 22 · 6.9B params. 📦 Open weights · HF. → Source
Flux 3 Action Base — Base robotics model from Black Forest Labs, parameter count undisclosed but positioned as the foundation for task-specific variants. Released Sep 22. 📦 Open weights · HF. → Source
━━━ CLOSED FRONTIER 🔒
GLM 5.3 Prime — High-throughput variant of GLM-5.3 with 1.5–2× faster output via inference acceleration, same 1M context. Released Sep 23 · 1M context · $2.80/$8.80 per 1M. ⚡ API live · OpenRouter. → Source
Qwen3.8 Max Prime — Higher-throughput SKU of Qwen3.8 Max accepting text, image, and video at a premium price point. Released Sep 23 · 1M context · $4.00/$12.00 per 1M. ⚡ API live · OpenRouter. → Source
Ember 1 — Fireworks Research's specialized reasoning model built on Kimi K3, with 1M context and $3/$15 pricing. Released Sep 24 · 1M context · $3.00/$15.00 per 1M. ⚡ API live · OpenRouter. → Source
Space Bunny Alpha — Anonymous large model with fast inference, strong coding, and native multimodal input — currently free. Released Sep 23 · 1M context · FREE. ⚡ API live · OpenRouter. → Source
📊 Surfaced 7 · scanned 12 today
Thursday, September 24, 2026
Robotics gets its own foundation-model moment while API providers ship throughput variants of existing frontier models.
━━━ SPECIALIZED 🎯
Flux 3 Action SO101 — Black Forest Labs' first robotics action model, targeting single-object manipulation with a 6.9B-parameter architecture. Released Sep 22 · 6.9B params. 📦 Open weights · HF. → Source
Flux 3 Action Droid — Companion robotics model for broader manipulation tasks, same 6.9B footprint as SO101. Released Sep 22 · 6.9B params. 📦 Open weights · HF. → Source
Flux 3 Action Base — Base robotics model from Black Forest Labs, parameter count undisclosed but positioned as the foundation for task-specific variants. Released Sep 22. 📦 Open weights · HF. → Source
━━━ CLOSED FRONTIER 🔒
GLM 5.3 Prime — High-throughput variant of GLM-5.3 with 1.5–2× faster output via inference acceleration, same 1M context. Released Sep 23 · 1M context · $2.80/$8.80 per 1M. ⚡ API live · OpenRouter. → Source
Qwen3.8 Max Prime — Higher-throughput SKU of Qwen3.8 Max accepting text, image, and video at a premium price point. Released Sep 23 · 1M context · $4.00/$12.00 per 1M. ⚡ API live · OpenRouter. → Source
Ember 1 — Fireworks Research's specialized reasoning model built on Kimi K3, with 1M context and $3/$15 pricing. Released Sep 24 · 1M context · $3.00/$15.00 per 1M. ⚡ API live · OpenRouter. → Source
Space Bunny Alpha — Anonymous large model with fast inference, strong coding, and native multimodal input — currently free. Released Sep 23 · 1M context · FREE. ⚡ API live · OpenRouter. → Source
📊 Surfaced 7 · scanned 12 today
🤖 ModelBytes Digest
Friday, September 25, 2026
Today's pattern: frontier labs keep cutting prices on SOTA models while open-weight RL and ASR variants ship same-week.
━━━ OPEN FRONTIER 🔓
MiMo V2.6 Pro RL — RL-tuned follow-up to MiMo V2.6 Pro; reasoning-focused RL on the 1024B backbone for builders needing stronger agentic outputs. Released Sep 25 · 1024.2B total params · MIT. 📦 Open weights · HF. → Source
━━━ SPECIALIZED 🎯
Audio8 ASR Infinite — 4.1B streaming ASR for infinite-length audio; Apache-2.0 makes it embeddable without licensing friction. Released Sep 25 · 4.1B total params · Apache-2.0. 📦 Open weights · HF. → Source
📊 Surfaced 2 · scanned 4 today
Friday, September 25, 2026
Today's pattern: frontier labs keep cutting prices on SOTA models while open-weight RL and ASR variants ship same-week.
━━━ OPEN FRONTIER 🔓
MiMo V2.6 Pro RL — RL-tuned follow-up to MiMo V2.6 Pro; reasoning-focused RL on the 1024B backbone for builders needing stronger agentic outputs. Released Sep 25 · 1024.2B total params · MIT. 📦 Open weights · HF. → Source
━━━ SPECIALIZED 🎯
Audio8 ASR Infinite — 4.1B streaming ASR for infinite-length audio; Apache-2.0 makes it embeddable without licensing friction. Released Sep 25 · 4.1B total params · Apache-2.0. 📦 Open weights · HF. → Source
📊 Surfaced 2 · scanned 4 today
🤖 ModelBytes Digest
Saturday, September 26, 2026
Today's releases push builders toward multimodal decision-making at the edge: open 0.8B–4B image-text models for cheap visual decisions, plus a closed embodied-reasoning API for physical agents.
━━━ SPECIALIZED 🎯
Intern Decision 4B — Apache-2.0 image-text-to-text model for visual decision tasks; 4B params. Released Sep 26 · Apache-2.0 · 4B params. 📦 Open weights · HF. → Source
Intern Decision 2B — Apache-2.0 image-text-to-text model for visual decision tasks; 2B params. Released Sep 26 · Apache-2.0 · 2B params. 📦 Open weights · HF. → Source
Intern Decision 0.8B — Apache-2.0 image-text-to-text model for visual decision tasks; 0.8B params. Released Sep 26 · Apache-2.0 · 0.8B params. 📦 Open weights · HF. → Source
Perceptron Mk1.5 — Embodied reasoning model for physical agents; accepts text, image, video, audio and returns text plus structured annotations. Released Sep 25 · 37k context · Closed/API · $0.15/$1.50 per 1M. ⚡ API live · OpenRouter. → Source
📊 Surfaced 4 · scanned 6 today
Saturday, September 26, 2026
Today's releases push builders toward multimodal decision-making at the edge: open 0.8B–4B image-text models for cheap visual decisions, plus a closed embodied-reasoning API for physical agents.
━━━ SPECIALIZED 🎯
Intern Decision 4B — Apache-2.0 image-text-to-text model for visual decision tasks; 4B params. Released Sep 26 · Apache-2.0 · 4B params. 📦 Open weights · HF. → Source
Intern Decision 2B — Apache-2.0 image-text-to-text model for visual decision tasks; 2B params. Released Sep 26 · Apache-2.0 · 2B params. 📦 Open weights · HF. → Source
Intern Decision 0.8B — Apache-2.0 image-text-to-text model for visual decision tasks; 0.8B params. Released Sep 26 · Apache-2.0 · 0.8B params. 📦 Open weights · HF. → Source
Perceptron Mk1.5 — Embodied reasoning model for physical agents; accepts text, image, video, audio and returns text plus structured annotations. Released Sep 25 · 37k context · Closed/API · $0.15/$1.50 per 1M. ⚡ API live · OpenRouter. → Source
📊 Surfaced 4 · scanned 6 today
🤖 ModelBytes Digest
Sunday, September 27, 2026
Today's only fresh release is a 144M CPU decision model, a counterweight to frontier scale for builders who need on-device inference.
━━━ LOCAL 🏠
Julia 1 — A 144.3M-param open-weight model built to run decision workloads on CPU, not GPU clusters. Released Sep 26 · 144.3M params · open weights. → Source
📊 Surfaced 1 · scanned 1 today
Sunday, September 27, 2026
Today's only fresh release is a 144M CPU decision model, a counterweight to frontier scale for builders who need on-device inference.
━━━ LOCAL 🏠
Julia 1 — A 144.3M-param open-weight model built to run decision workloads on CPU, not GPU clusters. Released Sep 26 · 144.3M params · open weights. → Source
📊 Surfaced 1 · scanned 1 today
🤖 ModelBytes Digest
Monday, September 28, 2026
Today's pattern is open multimodal and RL-tuned weights landing alongside a managed Grok, which means builders get cheaper local options for vision/OCR and a high-capacity API fallback in the same week.
━━━ OPEN FRONTIER 🔓
Holo4 27B — Open image-text-to-text generalist for non-commercial multimodal work. 27B total · CC BY-NC 4.0 · Released Sep 24. 📦 Open weights · HF. → Source
Holo4 35B A3B — Apache-licensed sibling with 35B total params for commercial multimodal use. 35B total · Apache-2.0 · Released Sep 24. 📦 Open weights · HF. → Source
MiMo V2.6 Flash RL — RL-tuned text-generation variant of MiMo V2.6 Flash. 310.8B total · MIT · Released Sep 28. 📦 Open weights · HF. → Source
━━━ CLOSED FRONTIER 🔒
Grok 4.6 — High-capability xAI model now available as a managed model on Google Cloud's Gemini Enterprise Agent Platform. ⚡ API live · Google Cloud. → Source
━━━ SPECIALIZED 🎯
TeleOCR — Small 1.4B OCR specialist for extracting text from images. 1.4B total · Apache-2.0 · Released Sep 28. 📦 Open weights · HF. → Source
📊 Surfaced 5 · scanned 9 today
Monday, September 28, 2026
Today's pattern is open multimodal and RL-tuned weights landing alongside a managed Grok, which means builders get cheaper local options for vision/OCR and a high-capacity API fallback in the same week.
━━━ OPEN FRONTIER 🔓
Holo4 27B — Open image-text-to-text generalist for non-commercial multimodal work. 27B total · CC BY-NC 4.0 · Released Sep 24. 📦 Open weights · HF. → Source
Holo4 35B A3B — Apache-licensed sibling with 35B total params for commercial multimodal use. 35B total · Apache-2.0 · Released Sep 24. 📦 Open weights · HF. → Source
MiMo V2.6 Flash RL — RL-tuned text-generation variant of MiMo V2.6 Flash. 310.8B total · MIT · Released Sep 28. 📦 Open weights · HF. → Source
━━━ CLOSED FRONTIER 🔒
Grok 4.6 — High-capability xAI model now available as a managed model on Google Cloud's Gemini Enterprise Agent Platform. ⚡ API live · Google Cloud. → Source
━━━ SPECIALIZED 🎯
TeleOCR — Small 1.4B OCR specialist for extracting text from images. 1.4B total · Apache-2.0 · Released Sep 28. 📦 Open weights · HF. → Source
📊 Surfaced 5 · scanned 9 today
🤖 ModelBytes Digest
Tuesday, September 29, 2026
Closed labs shipped point upgrades while open tooling added a 36B math verifier — builders get cheaper Sonnet-class API work and a new eval checker in the same window.
━━━ CLOSED FRONTIER 🔒
Claude Sonnet 5.5 — Anthropic's Sonnet-class upgrade for well-scoped everyday work, direct successor to Sonnet 5. 1M context, closed API, $2/$10 per 1M. Released Sep 28. ⚡ API live · OpenRouter. → Source
Astra — Gemini's September update, appearing in the latest cross-lab IQ comparison. → Source
Fable 5.1 — Anthropic's Fable line point release, shipping alongside Sonnet 5.5. → Source
Flash — Gemini Flash update, the lighter model in the September comparison. → Source
━━━ SPECIALIZED 🎯
AdvancedMathBench AutoVerifier — 36B image-text-to-text model built to auto-verify math benchmark solutions. 36B params. Released Sep 29. 📦 Open weights · HF. → Source
📊 Surfaced 5 · scanned 6 today
Tuesday, September 29, 2026
Closed labs shipped point upgrades while open tooling added a 36B math verifier — builders get cheaper Sonnet-class API work and a new eval checker in the same window.
━━━ CLOSED FRONTIER 🔒
Claude Sonnet 5.5 — Anthropic's Sonnet-class upgrade for well-scoped everyday work, direct successor to Sonnet 5. 1M context, closed API, $2/$10 per 1M. Released Sep 28. ⚡ API live · OpenRouter. → Source
Astra — Gemini's September update, appearing in the latest cross-lab IQ comparison. → Source
Fable 5.1 — Anthropic's Fable line point release, shipping alongside Sonnet 5.5. → Source
Flash — Gemini Flash update, the lighter model in the September comparison. → Source
━━━ SPECIALIZED 🎯
AdvancedMathBench AutoVerifier — 36B image-text-to-text model built to auto-verify math benchmark solutions. 36B params. Released Sep 29. 📦 Open weights · HF. → Source
📊 Surfaced 5 · scanned 6 today
🤖 ModelBytes Digest
Wednesday, September 30, 2026
Today's releases split between incremental closed upgrades and open multimodal weights, while OpenAI's shelved GPT-6.1 Astra signals safety gates tightening at the top.
━━━ OPEN FRONTIER 🔓
AREX 2 — BAAI's 27.4B image-text-to-text model under Apache 2.0, a general multimodal with open weights for builders who need vision-language without a closed API. Released Sep 29 · 27.4B params · Apache 2.0. 📦 Open weights · HF. → Source
━━━ CLOSED FRONTIER 🔒
GPT-6.1 Sol — Upgrade to GPT-6 Sol, positioned below Astra, aimed at agentic coding, computer use, and document-heavy professional work; Pro variant sets reasoning.mode=pro for higher-quality complex tasks. Released Sep 29 · 1.1M context · $2.00/$10.00 per 1M. ⚡ API live · OpenRouter. → Source
━━━ SPECIALIZED 🎯
Mercury Voice — Inception's voice model, confirmed Sep 29, for audio/voice tasks. Released Sep 29. → Source
📊 Surfaced 3 · scanned 5 today
Wednesday, September 30, 2026
Today's releases split between incremental closed upgrades and open multimodal weights, while OpenAI's shelved GPT-6.1 Astra signals safety gates tightening at the top.
━━━ OPEN FRONTIER 🔓
AREX 2 — BAAI's 27.4B image-text-to-text model under Apache 2.0, a general multimodal with open weights for builders who need vision-language without a closed API. Released Sep 29 · 27.4B params · Apache 2.0. 📦 Open weights · HF. → Source
━━━ CLOSED FRONTIER 🔒
GPT-6.1 Sol — Upgrade to GPT-6 Sol, positioned below Astra, aimed at agentic coding, computer use, and document-heavy professional work; Pro variant sets reasoning.mode=pro for higher-quality complex tasks. Released Sep 29 · 1.1M context · $2.00/$10.00 per 1M. ⚡ API live · OpenRouter. → Source
━━━ SPECIALIZED 🎯
Mercury Voice — Inception's voice model, confirmed Sep 29, for audio/voice tasks. Released Sep 29. → Source
📊 Surfaced 3 · scanned 5 today
🤖 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
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
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
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
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