The Inference Report
The announcements reveal three distinct competitive theaters, each organized around a different constraint. The first is latency and cost at scale: OpenAI is cutting GPT-5.6 pricing and shipping real-time voice agents that process queries in milliseconds, while MiniMax pushes 1M context windows and Xiaomi targets 1000 tokens per second inference. These aren't marginal improvements. A 92% satisfaction rate on 30,000 retail interactions in two weeks signals that the speed-cost tradeoff has crossed into production viability. The second theater is compression. PrismML's entire announcement volume centers on running models on phones at 1-bit and ternary precision, while Liquid AI publishes research on tokenizer expansion, state reduction, and low-discrepancy sequences. This isn't academic exercise. If a 27B model runs locally without cloud calls, the economics of API pricing collapse. The third is agent orchestration and reasoning: MiniMax announces agent teams for long-running tasks, Google DeepMind ships multi-robot collaboration via Gemini Robotics ER 2, and Anthropic publishes red team findings on cybersecurity. The pattern across all three is that labs are no longer competing primarily on benchmark scores. They're competing on whether you can actually deploy their technology without vendor lock-in, without latency penalties, and without per-token billing.
What's conspicuously absent is any announcement about new model capabilities that require cloud infrastructure to access. OpenAI's move to lower pricing and enable on-device alternatives, PrismML's focus on edge deployment, and Xiaomi's push on inference speed all point toward a market where the marginal value of a cloud API call is being priced out. Google DeepMind's robotics announcements and MiniMax's agent frameworks suggest the next frontier is task completion rather than token generation. The volume of compression research and edge deployment work from Liquid AI and PrismML, contrasted against the relative silence on frontier model scaling, indicates that labs believe the competitive advantage has shifted from model size to model efficiency and operational control. For enterprises, this means the era of pure API dependency is ending. For labs, it means the race for proprietary capability is being replaced by a race for deployment flexibility and inference cost. The winners will be the ones who can deliver performance where the user is, not where the server is.
Sloane Duvall
- Fixed-Point Reasoners: Stable and Adaptive Deep Looped Transformers
- Neural Low-Discrepancy Sequences
- SwitchCraft: A Programmatic Framework for Designing State-Switching Proteins
- MeMo: Model as a Memory
- Preconditioned DeltaNet: Curvature-aware Sequence Modeling for Linear Recurrences
- The Key to State Reduction in Linear Attention: A Rank-based Perspective
- Tokenizer Expansion: Upgrading a Model's Tokenizer in Place
- Reducing Doom Loops with Final Token Preference Optimization
- IFStruct: Measuring structured-output compliance
- MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Evolutionary Search
- MiniMax M3: Frontier Coding, 1M Context, Native Multimodality — All in One Model
- MiniMax Agent Team: Built for Long-Running Tasks and Continuous Evolution
- Why Can't the MiniMax LLM Say "Ma Jiaqi"? Internal Investigation of Sparse Token Forgetting
- MiniMax M2.7: Early Echoes of Self-Evolution
- Forge: Scalable Agent RL Framework and Algorithm
- How enabling two settings tripled our scores on the ARC-AGI-3 benchmark
- Accelerating scientific discovery with ChatGPT for Academic Researchers
- How GPT-5.6 fuses frontier intelligence with frontier efficiency
- Advancing the price-performance frontier with GPT-5.6
- How avatarin built a 24/7 retail agent with GPT-Realtime
- Announcing Bonsai 27B: The First 27B-Class Model to Run on a Phone
- Introducing 1-bit and Ternary Bonsai Image 4B: Image Generation for Local Devices
- Introducing Ternary Bonsai: Top Intelligence at 1.58 Bits
- Announcing 1-bit Bonsai: The First Commercially Viable 1-bit LLMs
- PrismML Announces 1-bit Bonsai 27B – The First 27B Model to Run on a Phone
- PrismML Releases Bonsai Image 4B
- PrismML Introduces Ternary Bonsai Model Family
- Caltech Researchers Claim Radical Compression of High-Fidelity AI Models
- PrismML Launches World's First 1-Bit AI Model to Redefine Intelligence at the Edge
A curated reference of models from major AI labs, with open/closed weight status, input modalities, and context window size. American labs tend towards closed weights models and Chinese labs tend toward open weights models.
AllenAINone
- AllenAI: Olmo 3 32B ThinkText66K
Amazon- Amazon: Nova 2 LiteTextVisionVideoFiles1M
- Amazon: Nova Premier 1.0TextVision1M
None
Anthropic- Anthropic: Claude Fable 5TextVisionFiles1M
- Anthropic: Claude Haiku 4.5TextVisionFiles200K
- Anthropic: Claude Opus 4.1VisionTextFiles200K
- Anthropic: Claude Opus 4.5FilesVisionText200K
- Anthropic: Claude Opus 4.6TextVisionFiles1M
- Anthropic: Claude Opus 4.7TextVisionFiles1M
- Anthropic: Claude Opus 4.7 (Fast)TextVisionFiles1M
- Anthropic: Claude Opus 4.8TextVisionFiles1M
- Anthropic: Claude Opus 4.8 (Fast)TextVisionFiles1M
- Anthropic: Claude Sonnet 4.5TextVisionFiles1M
- Anthropic: Claude Sonnet 4.6TextVisionFiles1M
- Anthropic: Claude Sonnet 5TextVisionFiles1M
- Claude Opus 5TextVisionFiles1M
- Claude Opus 5 (Fast)TextVisionFiles1M
None
Arcee AINone
- Arcee AI: Trinity Large ThinkingText262K
Deep CogitoNone
- Deep Cogito: Cogito v2.1 671BText128K
Google- Google: Gemini 3 Flash PreviewTextVisionFilesAudioVideo1M
- Google: Gemini 3.1 Flash LiteTextVisionVideoFilesAudio1M
- Google: Gemini 3.1 Flash Lite PreviewTextVisionVideoFilesAudio1M
- Google: Gemini 3.1 Pro PreviewAudioFilesVisionTextVideo1M
- Google: Gemini 3.1 Pro Preview Custom ToolsTextAudioVisionVideoFiles1M
- Google: Gemini 3.5 FlashTextVisionVideoFilesAudio1M
- Google: Gemini 3.5 Flash LiteTextVisionVideoFilesAudio1M
- Google: Gemini 3.6 FlashTextVisionVideoFilesAudio1M
- Google: Lyria 3 Clip PreviewTextVision1M
- Google: Lyria 3 Pro PreviewTextVision1M
- Google: Nano Banana (Gemini 2.5 Flash Image)VisionText33K
- Google: Nano Banana 2 (Gemini 3.1 Flash Image Preview)VisionText66K
- Google: Nano Banana 2 (Gemini 3.1 Flash Image)VisionText131K
- Google: Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image)VisionText66K
- Google: Nano Banana Pro (Gemini 3 Pro Image Preview)VisionText66K
- Google: Nano Banana Pro (Gemini 3 Pro Image)VisionText131K
- Google: Gemma 4 26B A4B VisionTextVideo262K
- Google: Gemma 4 31BVisionTextVideo262K
IBMNone
- IBM: Granite 4.0 MicroText131K
- IBM: Granite 4.1 8BText131K
Inception- Inception: Mercury 2Text128K
None
Meta (Muse)- Meta: Muse Spark 1.1TextVisionVideoFilesAudio1M
None
Nous ResearchNone
- Nous: Hermes 4 405BText131K
- Nous: Hermes 4 70BText131K
NVIDIANone
- NVIDIA: Nemotron 3 Nano 30B A3BText262K
- NVIDIA: Nemotron 3 SuperText1M
- NVIDIA: Nemotron 3 UltraText512K
OpenAI- OpenAI: GPT AudioTextAudio128K
- OpenAI: GPT Audio MiniTextAudio128K
- OpenAI: GPT-5TextVisionFiles400K
- OpenAI: GPT-5 ImageVisionTextFiles400K
- OpenAI: GPT-5 Image MiniFilesVisionText400K
- OpenAI: GPT-5 MiniTextVisionFiles400K
- OpenAI: GPT-5 NanoTextVisionFiles400K
- OpenAI: GPT-5 ProVisionTextFiles400K
- OpenAI: GPT-5.1VisionTextFiles400K
- OpenAI: GPT-5.1 ChatFilesVisionText128K
- OpenAI: GPT-5.1-CodexTextVision400K
- OpenAI: GPT-5.1-Codex-MaxTextVision400K
- OpenAI: GPT-5.1-Codex-MiniVisionText400K
- OpenAI: GPT-5.2FilesVisionText400K
- OpenAI: GPT-5.2 ChatFilesVisionText128K
- OpenAI: GPT-5.2 ProVisionTextFiles400K
- OpenAI: GPT-5.2-CodexTextVision400K
- OpenAI: GPT-5.3 ChatTextVisionFiles128K
- OpenAI: GPT-5.3-CodexTextVisionFiles400K
- OpenAI: GPT-5.4TextVisionFiles1M
- OpenAI: GPT-5.4 Image 2VisionTextFiles272K
- OpenAI: GPT-5.4 MiniFilesVisionText400K
- OpenAI: GPT-5.4 NanoFilesVisionText400K
- OpenAI: GPT-5.4 ProTextVisionFiles1M
- OpenAI: GPT-5.5FilesVisionText1M
- OpenAI: GPT-5.5 ProFilesVisionText1M
- OpenAI: GPT-5.6 LunaFilesVisionText1M
- OpenAI: GPT-5.6 Luna ProFilesVisionText1M
- OpenAI: GPT-5.6 SolFilesVisionText1M
- OpenAI: GPT-5.6 Sol ProFilesVisionText1M
- OpenAI: GPT-5.6 TerraFilesVisionText1M
- OpenAI: GPT-5.6 Terra ProFilesVisionText1M
- OpenAI: gpt-oss-120bText131K
- OpenAI: gpt-oss-20bText131K
Perceptron- Perceptron: Perceptron Mk1TextVisionVideo33K
None
Perplexity- Perplexity: Sonar Pro SearchTextVision200K
None
PoolsideNone
- Poolside: Laguna S 2.1Text1M
- Poolside: Laguna XS 2.1Text262K
Reka AINone
- Reka EdgeVisionTextVideo16K
RelaceNone
- Relace: Relace Apply 3Text256K
- Relace: Relace SearchText256K
Thinking MachinesNone
- Thinking Machines: InklingTextVisionAudio1M
xAI- xAI: Grok 4.20TextVisionFiles2M
- xAI: Grok 4.20 Multi-AgentTextVisionFiles2M
- xAI: Grok 4.3TextVisionFiles1M
- xAI: Grok 4.5TextVisionFiles500K
- xAI: Grok Build 0.1TextVisionFiles256K
None
Mistral AI- Mistral: Mistral Medium 3.1TextVisionFiles131K
- Mistral: Codestral 2508TextFiles256K
- Mistral: Devstral 2 2512TextFiles262K
- Mistral: Ministral 3 14B 2512TextVision262K
- Mistral: Ministral 3 3B 2512TextVision131K
- Mistral: Ministral 3 8B 2512TextVision262K
- Mistral: Mistral Large 3 2512TextVisionFiles262K
- Mistral: Mistral Medium 3.5TextVisionFiles262K
- Mistral: Mistral Small 4TextVision262K
- Mistral: Voxtral Small 24B 2507TextAudioFiles32K
AI21 LabsNone
- AI21: Jamba Large 1.7Text256K
Alibaba (Qwen)- Qwen: Qwen3 Coder FlashText1M
- Qwen: Qwen3 Coder PlusText1M
- Qwen: Qwen3 MaxText262K
- Qwen: Qwen3 Max ThinkingText262K
- Qwen: Qwen3.6 FlashTextVisionVideo1M
- Qwen: Qwen3.6 Max PreviewText262K
- Qwen: Qwen3.6 PlusTextVisionVideo1M
- Qwen: Qwen3.7 FlashTextVisionVideo1M
- Qwen: Qwen3.7 MaxText1M
- Qwen: Qwen3.7 PlusTextVision1M
- Qwen: Qwen3 30B A3B Thinking 2507Text82K
- Qwen: Qwen3 Coder 30B A3B InstructText262K
- Qwen: Qwen3 Coder NextText262K
- Qwen: Qwen3 Next 80B A3B InstructText262K
- Qwen: Qwen3 Next 80B A3B ThinkingText262K
- Qwen: Qwen3 VL 235B A22B InstructTextVision262K
- Qwen: Qwen3 VL 235B A22B ThinkingTextVision131K
- Qwen: Qwen3 VL 30B A3B InstructTextVision262K
- Qwen: Qwen3 VL 30B A3B ThinkingTextVision262K
- Qwen: Qwen3 VL 32B InstructTextVision131K
- Qwen: Qwen3 VL 8B InstructVisionText262K
- Qwen: Qwen3 VL 8B ThinkingVisionText131K
- Qwen: Qwen3.5 397B A17BTextVisionVideo262K
- Qwen: Qwen3.5-122B-A10BTextVisionVideo262K
- Qwen: Qwen3.5-27BTextVisionVideo262K
- Qwen: Qwen3.5-35B-A3BTextVisionVideo262K
- Qwen: Qwen3.5-9BTextVisionVideo262K
- Qwen: Qwen3.6 27BTextVisionVideo262K
- Qwen: Qwen3.6 35B A3BTextVisionVideo262K
ByteDance- Seed: Seed 1.6VisionTextVideo262K
- Seed: Seed 1.6 FlashVisionTextVideo262K
- Seed: Seed-2.0-LiteTextVisionVideo262K
- Seed: Seed-2.0-MiniTextVisionVideo262K
None
DeepSeekNone
- DeepSeek: DeepSeek V3.1Text164K
- DeepSeek: DeepSeek V3.1 TerminusText164K
- DeepSeek: DeepSeek V3.2Text164K
- DeepSeek: DeepSeek V3.2 ExpText164K
- DeepSeek: DeepSeek V4 FlashText1M
- DeepSeek: DeepSeek V4 Flash 0731Text1M
- DeepSeek: DeepSeek V4 ProText1M
MiniMax- MiniMax: MiniMax M2-herText66K
- MiniMax: MiniMax M2Text205K
- MiniMax: MiniMax M2.1Text205K
- MiniMax: MiniMax M2.5Text205K
- MiniMax: MiniMax M2.7Text205K
- MiniMax: MiniMax M3TextVisionVideo1M
Moonshot AINone
- MoonshotAI: Kimi K2 0905Text262K
- MoonshotAI: Kimi K2 ThinkingText262K
- MoonshotAI: Kimi K2.5TextVision262K
- MoonshotAI: Kimi K2.6TextVision262K
- MoonshotAI: Kimi K2.7 CodeTextVision262K
- MoonshotAI: Kimi K3TextVision1M
XiaomiNone
- Xiaomi: MiMo-V2.5TextAudioVisionVideo1M
- Xiaomi: MiMo-V2.5-ProText1M
Z.aiNone
- Z.ai: GLM 4.5VTextVision66K
- Z.ai: GLM 4.6Text205K
- Z.ai: GLM 4.6VVisionTextVideo131K
- Z.ai: GLM 4.7Text205K
- Z.ai: GLM 4.7 FlashText203K
- Z.ai: GLM 5Text205K
- Z.ai: GLM 5 TurboText203K
- Z.ai: GLM 5.1Text205K
- Z.ai: GLM 5.2Text1M
- Z.ai: GLM 5V TurboVisionTextVideo203K
Upstage- Upstage: Solar Pro 3Text128K
None