Each report examines a model in depth: test results by discipline, strengths-and-weaknesses profile, Political Compass, and a clear recommendation for production use. The analyses are based exclusively on reproducible CrucibleMark benchmarks.
GLM-5.3 is Z.AI’s Frontier coding flagship from August 14, 2026, with 744 billion total and 40 billion active parameters. All improvements over GLM-5.2 stem from post-training: Terminal-Bench 3.0 rose from 4.6 to 28.3 percent, ExploitBench more than doubled. Reasoning is mandatorily active, context one million tokens. Weights available since 08/28/2026 under the bespoke GLM-5.3 license (MIT-like, $10B MaaS security review gate); Chinese cloud infrastructure.
Claude Opus 5.5 is Anthropic’s new Frontier model: as of September 22, 2026, it replaces Opus 5 and is claimed by the manufacturer to achieve Fable-5.1-level performance at 40 percent lower typical workload costs. Adaptive Thinking is active by default and controllable via effort level; the context window holds one million tokens with 128,000 output tokens. The reasoning classifier errors from Opus 5 have been fixed; only the metacognition refusal remains.
As an open preview of the Qwen4 architecture, Alibaba introduces Qwen3.8-Flash-Next — an experimental MoE model that NVIDIA has prepared as an NVFP4 quantization for local inference. Of approximately 180 billion parameters on disk, only around 6 billion activate per token, with text, image, and video input and 262,000 tokens of native context. License: combined NVIDIA and Qwen license. Important: This preview is not the hosted API Qwen3.8-Flash.
This EXL3 quantization of Z.AI’s GLM-5.3-Flash runs on TensorFold, a young open inference engine with lossless speculative decoding: accelerated draft tokens correspond exactly to the result of serial decoding. The MoE activates around 18 billion of 320 billion parameters per token, processes text, image, and video, and the weights are available under the MIT license.
Claude Opus 4.8 is a proprietary, multimodal foundation model by Anthropic. It is available as a cloud API and is distinguished by its agentic capabilities and an extremely large context window of one million tokens. The model can process both text and images and demonstrates high performance when using external tools. It is designed for complex, multi-step tasks and the orchestration of agents.
Swift Qwen 3.8 27B is UkisAI’s reasoning-efficiency fine-tune on Qwen 3.8 27B: up to 58 percent fewer thinking tokens at under one percent performance loss and roughly twice the throughput on reasoning tasks. NVFP4 quantization with 262,000 tokens of context, MTP head for speculative decoding, and documented tool use — license with a commercial ARR threshold.
320 billion total, 18 billion active parameters, and MIT license: GLM-5.3-Flash is the open mid-tier variant of Z.AI’s 5.3 family for coding and agentic workloads, with native image and video understanding and one million tokens of context. Hybrid attention keeps the inference footprint moderate despite the overall model size. Reasoning is mandatorily active, weights run locally — the sovereignty risk of the cloud is eliminated.
This abliterated community variant of Qwen 3.8 27B removes safety Refusals from the weights, making it usable for security research and red-teaming — at an MMLU loss of around two points according to the developer. Locally operable under Apache-2.0, with a 262,000-token context and image and video input.
Kimi K3 has been Moonshot AI’s Frontier flagship since July 2026, with 2.8 trillion total parameters making it the largest Open Weights model in the world to date. The Stable-LatentMoE architecture activates only 16 billion parameters per token from 896 experts; the context window spans 1,000,000 tokens for long-horizon agentic workflows. Kimi Delta Attention and Attention Residuals are designed to significantly improve scaling efficiency over the predecessor model.
Claude Opus 5 is Anthropic’s Opus flagship as of July 24, 2026, positioned as an everyday model between Opus 4.8 and the more expensive Fable 5. The cloud-only model under US jurisdiction offers 1 million tokens of context, 128,000 tokens of output, and adaptive reasoning control with five effort levels (low/medium/high/xhigh/max). Mid-conversation tool switching without cache loss and a Fast Mode with 2.5× speed round out the offering.
Xiaomi MiMo V2.5 Pro is Xiaomi’s flagship model with 1.02 trillion total and 42 billion active parameters, designed for frontier reasoning and agentic workflows. Its hybrid attention architecture significantly reduces KV-cache memory, and the context window spans one million tokens. Natively omnimodal for text, image, video, and audio, and fully commercially usable under the MIT license.
With 1.02 trillion total parameters, MiMo-V2.6-Pro-RL by Xiaomi is currently the largest model with fully open weights. Approximately 42 billion parameters activate per token, trained in a single mixed reinforcement learning run across coding, agent, and safety tasks; the accompanying live dashboard of the training process provided unusual transparency. Omnimodal for text, image, video, and audio, context up to 1 million tokens, MIT license.
Qwen3.8 Flash is Alibaba’s cloud-only multimodal reasoning model with a one-million-token context window and pricing of $0.16 / $0.47 per million tokens. It processes text, image, and video with tool calling, accessible via Alibaba Cloud and OpenRouter. The open base Qwen3.8-Flash-Next is available, but the tested build is cloud-only. Architecture and parameter count are not disclosed, and data is routed through Chinese jurisdiction.
Qwen3.8-27B is Alibaba’s dense 27.8-billion-parameter variant of the Qwen3.8 family (August 14, 2026), natively multimodal with a vision encoder for text, image, and video under the Apache 2.0 license. The 64-layer hybrid architecture of Gated DeltaNet plus Gated Attention blocks plus Multi-Token Prediction delivers 262,144 tokens of context (extensible to approximately one million via YaRN), configurable reasoning (xhigh/medium/low), and an optional no-thinking mode.
Muse Spark 1.2 is Meta’s proprietary Frontier model from August 5, 2026, designed for coding and agentic workflows, released by Meta Superintelligence Labs alongside the terminal agent Muse Code, with which it was co-trained. The cloud-only model under US jurisdiction (CLOUD Act) processes text, image, video, and audio with a context of 1,048,576 tokens (max. 131,072 output) and offers configurable reasoning effort up to ‘xhigh’.
Where the Opus class is too expensive, Claude Sonnet 4.6 steps in: coding, computer use, and agentic workflows at near-Opus level, at the lower Sonnet price. The model operates with adaptive thinking in three effort levels, processes text, images, and PDF documents, and offers a context window of one million tokens, generally available since March 2026.
Kimi K2.6 is Moonshot AI’s multimodal model for agentic tasks, coding, and tool-assisted workflows, with native input support for text, image, and video. The MoE architecture activates only 32 billion of the total one trillion parameters per token; the context window spans 256,000 tokens. Available as an Open Weights model locally or via cloud API, with Chinese jurisdiction as a material cloud risk factor.
Claude Haiku 5.5 has been Anthropic’s fastest model in the 5.5 family since early October 2026, built for high-volume tasks such as summarization, classification, browser use, and subagents. Context grows from 200,000 to one million tokens, output to 128,000, and average runtime costs are around 75 percent below Haiku 4.5. New to the Haiku class: adaptive reasoning with effort control. Text and image as input, proprietary, API-only.
MiniMax M3 is a multimodal MoE model with a context window of one million tokens, focused on agentic workflows, coding, and tool use. Of 428 billion total parameters, only 23 billion are active per token; the model processes text, image, and video as input. Its Chinese origin requires a separate data privacy risk assessment when used via cloud.
Gemini 3.5 Flash is Google’s fastest Frontier-class model, delivering near Pro-level performance at Flash pricing. With Dynamic Thinking across four configurable levels, a one-million-token context window, and full multimodality for text, images, audio, video, and PDF, the model is well-suited for agentic workflows, coding, and high-throughput workloads.
Claude Sonnet 5 has been the new default for free and pro users since late June 2026. The successor to Sonnet 4.6 brings Adaptive Thinking as default behavior, speed close to Opus, and processes text and images with a 1,000,000-token context window. Safety is significantly improved, with better resistance to prompt injection and fewer hallucinations.
GPT-5.5 is OpenAI’s Frontier model for complex professional workloads and agentic coding, with a context window of 1.05 million tokens. The model uses internal chain-of-thought reasoning that is not visible in the API response, and is designed for research, coding, and demanding productivity tasks. Available exclusively via the OpenAI API.
Kimi K2.5 is Moonshot AI’s flagship model featuring active chain-of-thought reasoning, multimodal input for text and images, and a focus on reasoning and agentic tasks. The MoE architecture activates 32 billion of the total one trillion parameters per token; the context window spans 128,000 tokens. Available as an Open Weights variant locally or via cloud, with Chinese jurisdiction as a cloud risk factor.
Qwen3.8-2.4T-A95B is Alibaba’s first Open Weights release at Qwen-Max level (August 12, 2026), a fine-grained Mixture-of-Experts model with 2.4 trillion total and 95 billion active parameters per token. License: proprietary ‘Qwen3.8-Max License’. The model processes text in a 262,144-token context (expandable to approximately one million) with a mandatory reasoning mode (low/high/xhigh) and hybrid attention combining Gated-DeltaNet and Gated-Attention.
GLM-4.6 is Zhipu AI’s Frontier language model with a focus on Chinese and English language proficiency. The model operates with a context window of 128,000 tokens and supports tool use for agentic workflows. Due to the Chinese manufacturer jurisdiction, a separate privacy assessment is required, and commercial use is subject to restrictions.
Ornith 1.5 35B-A3B has been the mid-tier model in DeepReinforce’s open Ornith family since August 19, 2026. The MoE activates only around 3B of 35B parameters per token, yet according to the manufacturer it significantly outperforms the similarly sized Qwen 3.6-35B on all coding and agentic benchmarks. Trained with a closed self-improvement loop that jointly optimizes its own tasks, scaffolds, and solutions. License: MIT, fully open and commercially usable.
Qwen 3.7 Max is Alibaba’s proprietary flagship model of the Qwen 3.7 series, focused on agentic coding workflows and autonomous operation of up to 35 hours. The model features a one-million-token context window, configurable thinking mode, and native tool-use support. Available exclusively via cloud APIs; Chinese jurisdiction applies.
What FP8 block quantization delivers with Ornith-1.0-35B-FP8: an Open Weights MoE with only around 3 of 35 billion active parameters per token runs on a single GPU and brings 262,144 tokens of context, native thinking, and tool calling. DeepReinforce trained the model to learn its own agentic approach rather than working with a fixed rule set. MIT license, commercial use, and fine-tuning without restrictions.
Gemini 3.8 Flash is Google’s most intelligent Flash model (GA since September 2, 2026), built on 3.7 Flash for long-horizon software engineering, autonomous agents, and complex enterprise workflows. Three thinking levels (low, medium, high, default medium) control reasoning depth; one million tokens of context and 65,536 output tokens fit entire repositories in a single request. Introductory pricing of 0.75 / 3.75 USD per million tokens through year-end, then 1.50 / 7.50.
Qwen 3.6 27B is Alibaba’s dense Open Weights model with 27.8 billion active parameters, which according to the manufacturer outperforms its own MoE sibling model 35B-A3B in coding, agent, and vision tests. Released on April 22, 2026 under Apache 2.0 for local deployment, the architecture combines hybrid attention with native multi-token prediction for noticeable throughput gains. This Card covers the NVFP4 variant — a performance update for vLLM on NVIDIA Blackwell: NVFP4 is a 4-bit floating-point format with two-level micro-block scaling (E4M3 per 16-value block plus FP32 per-tensor scale), reducing memory requirements by ~3.5× compared to FP16 and by ~1.8× compared to FP8 at under 1% accuracy loss. Predecessor profile without NVFP4: qwen3_6-27B-pre025 (Card qwen3_6-27B-pre025–VSPK.json, tests 2026-07-09, vLLM prior to 0.25.1).
GLM-5.2 is Z.AI’s current flagship with 744 billion total and 40 billion active parameters in a MoE architecture, optimized for complex engineering workflows and long-running coding tasks. The context window spans one million tokens; the weights are available as an Open Weights model under the MIT license.
Qwen 3.6 Plus is Alibaba’s proprietary flagship model of the Qwen 3.6 series, featuring a hybrid MoE architecture with a focus on agentic coding and multimodal processing. With a one-million-token context window, configurable thinking mode, and native agentic capabilities, the model targets demanding production applications. Available exclusively via cloud APIs; Chinese jurisdiction applies.
Solar Pro 4 is Upstage’s agentic flagship model, released on August 6/10, 2026, with an undisclosed parameter count. It is available as a proprietary cloud service (additionally as dedicated/on-premises deployment for enterprise customers) and is designed for multi-step agentic workflows across documents, terminals, and tool calls. The model offers a context window of 524,288 tokens with up to 131,072 output tokens, supports English, Korean, and Japanese for input and output, and allows a configurable ‘Reasoning Effort’ (high for deep analysis, low for real-time chat speed).
One million tokens of context for text, image, audio, and video in a single model: Qwen3.8-Omni-Flash (available September 18, 2026) is Alibaba’s first omni-modal model with an agentic focus — it understands multi-hour audio and video material, plans tasks, and executes them via tool calling. At 0.15 / 0.47 USD per million tokens, Alibaba claims it reduces audio costs by more than 98 percent compared to its predecessor. Output is text only; the weights remain closed.
Occamy 1.0 by Accio-Lab is an agentic derivative of the Qwen3.6-35B-A3B checkpoint, focused on long-horizon co-work sessions with tools, structured APIs, and persistent state tracking. The NVFP4 quantization is selective: only the routed experts are quantized, while attention, router, embeddings, and output head remain in BF16. The 35-billion-parameter MoE activates only 3 billion parameters per token and supports 262,000 tokens of context. Apache 2.0 license and documented provenance with recipe, data, and validation artifacts.
The smaller sibling of the MiMo-V2.6 series prioritizes efficiency over maximum size: MiMo-V2.6-Flash-RL by Xiaomi activates around 15 billion of 309 billion parameters per token, matching the flagship Pro on agent and coding tasks at roughly one-third of the API price. Omnimodal for text, image, video, and audio, context up to 1 million tokens, Open Weights under the MIT license.
Google DeepMind deliberately sets Gemma 4 26B-A4B Instruct apart from earlier Gemma generations: released under a genuine Apache 2.0 license, meaning without restrictive Gemma terms of use. This Open Weights MoE activates only approximately 3.8 of 25.2 billion parameters per token and supports Multi-Token Prediction for faster decoding. Multimodality for text and images, a 262,144-token context window, native function calling, and a configurable thinking mode round out the profile.
Optimized for speed: Claude Haiku 4.5 is Anthropic’s compact model for real-time interactions and high-throughput API pipelines. It processes text and image inputs with a context window of 200,000 tokens, supports tool calls, and delivers responses at a price of $1 per million input tokens and $5 per million output tokens. Anthropic has announced retirement for October 15, 2026.
Mistral 3 Large is the open Frontier model from Mistral AI’s third generation, featuring native text and image input and a context window of 256,000 tokens. The Sparse MoE architecture combines 675 billion total parameters with 41 billion active parameters per token. Available under the Apache 2.0 license, from a European provider environment with GDPR compliance.
DeepSeek V4.1 Flash is the compact variant of the DeepSeek V4.1 family, with 552 billion total parameters and 16 billion active parameters per token. The model is released as Open Weights under the MIT license with image and text processing capabilities. Its 1-million-token context window is designed for long agentic workflows, but due to its size requires server infrastructure for local deployment.
Grok 4.6 is xAI’s Frontier model from August 12, 2026, designed for coding, long agent sessions, and knowledge work — proprietary, cloud-only, under US jurisdiction (CLOUD Act). The model processes text and images with a context of 500,000 tokens and offers four reasoning levels (low/medium/high/xhigh). An optional Priority Processing Service Tier doubles API costs in exchange for lower latency.
Eight of 552 billion backbone parameters active during reads, sixteen during writes: DeepSeek-V4.1-Flash shifts compute to where agents touch it most. This card describes a community quant as an EXL3 variant of the official MIT-licensed model from September 10, 2026 — multimodal for text and image, one million tokens of context, operable on shared server memory thanks to quantization.
Ministral 3 14B is the top model of the Ministral 3 family for local assistant and analysis workloads. 13.9B dense parameters, 256,000 tokens of context, multimodal input for text and image, native function calling and JSON output. Licensed under Apache 2.0 and fully operable locally as an Unsloth GGUF variant — the most capable Ministral to date without any cloud dependency.
Gemini 2.0 Flash is a fast, cost-efficient, and highly scalable multimodal model from Google. It was designed for high-frequency, low-latency tasks and features a context window of one million tokens. The model processes text, images, audio, and video, and supports the use of external tools, making it versatile for agentic applications.
Occamy 1.0 by Accio-Lab is an agentic derivative of the Qwen3.6-35B-A3B checkpoint, focused on long-horizon co-work sessions with tools, structured APIs, and persistent state tracking. The NVFP4 quantization is selective: only the routed experts are quantized, while attention, router, embeddings, and output head remain in BF16. The 35-billion-parameter MoE activates only 3 billion parameters per token and supports 262,000 tokens of context. Apache 2.0 license and documented provenance with recipe, data, and validation artifacts.
GPT-OSS 120B has been OpenAI’s largest Open Weights model since August 2025, released under the Apache 2.0 license with free commercial use. The MoE bundles 116.8 billion total parameters at only 5.1 billion active parameters per token and runs on a single high-memory GPU thanks to native MXFP4 quantization. Three reasoning levels and native tool use in the Harmony format round out the profile.
Qwen 3.5 35B-A3B is a multimodal MoE model by Alibaba with 35 billion total and 3 billion active parameters on a hybrid architecture. This Q4 quantization by Unsloth enables efficient local operation; the context window spans 262,000 tokens. Features an optional thinking mode, native tool use, and vision capability via a separate multimodal projector file.
Kimi K2.7 Code is the coding-specialized agentic model from Moonshot AI within the K2 family, optimized for long-horizon software engineering workflows. The MoE architecture activates 32 billion out of a total of one trillion parameters per token, with a context window of 256,000 tokens. Multimodal input for text, image, and video, always-on thinking mode, and tool use support. Available as an Open Weights model under a Modified MIT license.
Mistral Medium 3.5 is Mistral AI’s Frontier model focused on agentic workflows and coding. With 128 billion parameters, the model operates with a context window of 256,000 tokens and supports multimodal inputs for text and image. Available as an Open Weights model under a Modified MIT license for local use or via cloud API, from a European provider environment with GDPR compliance.
Qwen 3.6 35B-A3B is Alibaba’s MoE model with 35 billion total and approximately 3 billion active parameters per token, released on April 22, 2026 under Apache 2.0 with open weights for local deployment. The hybrid attention architecture combines classic attention with a linear variant; Multi-Token Prediction noticeably accelerates generation.
Qwen 3 Coder Next is a coding-specialized Open Weights MoE model by Alibaba with 80 billion total and 3 billion active parameters. Q4 quantization significantly reduces memory requirements for local inference; the context window spans 262,000 tokens. Deployable locally on Workstation hardware under the Apache 2.0 license, optimized for coding agents and large codebases.
GLM-4.7 is Zhipu AI’s flagship model with 355 billion total and 32 billion active parameters in a MoE architecture, optimized for agentic coding, reasoning, and bilingual tasks in Chinese and English. The model supports a switchable thinking system and is available as an Open Weights variant for local deployment or via cloud interfaces.
Grok 4.5 has been xAI’s current flagship since July 2026, with native real-time access to web and X data, a 500,000-token context window, and multimodal text and image input. Reasoning runs server-side at all times and is controlled via configurable reasoning effort levels, with no visible chain-of-thought tags in the response text. Function calling, structured outputs, and prompt caching round out the profile for coding and agentic workflows.
GPT-5.6 Luna is the cheapest and fastest tier of OpenAI’s three-tier GPT-5.6 series (Sol, Terra, Luna) for high-volume, latency-sensitive tasks — available since July 30, 2026 at $0.20 / $1.20 per million tokens, roughly 80 percent below Sol. The 1-million-token context variant with 128,000 output tokens delivers frontier-adjacent agentic performance according to OpenAI, but falls off noticeably against its larger siblings on context recall beyond 512,000 tokens.
This ARA-Abliterated variant removes the safety filters from Google’s Gemma 4 26B-A4B MoE and delivers the model as a Q5-GGUF on Workstation hardware. The architecture remains efficient at 25 billion total and approximately 4 billion active parameters per token; 256,000 tokens of context, Apache 2.0 license. Thinking status and multimodal capabilities have not yet been cleanly verified in this variant — intended for research and red-teaming, not as an end-consumer assistant.
GLM-5.1 is Z.AI’s post-training upgrade with 754 billion total and 40 billion active parameters in a MoE architecture, optimized for long-horizon agentic coding workflows with up to eight hours of autonomous execution. The context window spans 200,000 tokens, and the weights are available as an Open Weights model under the MIT license.
Mistral Small 4 is Mistral AI’s compact Open Weights model for general and agentic tasks. The MoE architecture activates only 6.5 billion of the total 119 billion parameters per token, the context window spans 256,000 tokens, and the model processes text and image inputs. Available under the Apache 2.0 license for local use or via the Mistral API, from a European provider environment.
Signal 3.8 27B by AgentionAI is a token-efficiency fine-tune on Qwen 3.8 27B: via self-distillation, the model generates roughly 57 percent fewer response tokens and thinking chains half the length, at equal or better quality. Licensed under Apache 2.0, available in a mid-range Q5 quantization with a 262,000-token context and local inference.
Gemma 4 31B as a cloud API variant from Google DeepMind, with native support for text, image, audio, and video inputs and no local hardware requirements. The weights are also available for local deployment under the Apache 2.0 license, but this variant describes cloud usage with a context window of 128,000 tokens.
Google’s Frontier reasoning model with sparse MoE architecture and configurable Extended Thinking. Gemini 2.5 Pro operates with a one-million-token context window, natively processes text, images, audio, and video, and is exclusively accessible via the Google Cloud API. The focus is on complex reasoning and demanding coding tasks.
Qwen 3.5 397B A17B is Alibaba’s first natively multimodal Open Weights Frontier model, processing text, image, and video within a single model. Its hybrid architecture combining Gated DeltaNet and Sparse MoE activates only 17 billion of the total 397 billion parameters per token, with a context window of 256,000 tokens. Fully commercially usable under the Apache 2.0 license.
MiniMax M2.7 is a Chinese Frontier generalist model with a context window of 205,000 tokens for large-scale documents and multilingual applications. The MoE architecture delivers high performance for general language and reasoning tasks; the model is available as a cloud variant and designed for productive applications. When used via cloud, Chinese jurisdiction applies with the corresponding data privacy implications.
Instruct profile of the Q6_K_XL GGUF distribution of Gemma 4 12B Instruct (Unsloth) for local deployment on the DGX Spark: identical weights to the base profile, but with reasoning disabled server-side (–reasoning off). 12 billion dense parameters, 256,000-token context, Apache 2.0 license. The profile exists as a replacement run under the coverage rule of the Political Compass module, because the thinking run of the base profile suffered from truncations; it runs as a standalone benchmark entry.
DeepSeek V4 Flash is the efficiency-optimized variant of the V4 family: a hybrid attention MoE with 284 billion total parameters, of which only 13 billion are active per token. The model operates with a context window of one million tokens, supports three reasoning modes, and is locally deployable as an Open Weights model under the MIT license. The Chinese vendor jurisdiction requires a separate assessment for cloud usage.
A Q6_K_XL GGUF distribution of Gemma 4 12B Instruct for local deployment: 12 billion dense parameters, encoder-free unified multimodal architecture with text, image, audio, and video input. Apache 2.0 license, 256,000 token context, more compact than Q4 variants at moderate additional memory. Audio and video support depend locally on a compatible multimodal projector.
DeepSeek V4 Pro is the flagship of the V4 line, designed for reasoning, coding, and agentic workflows. The hybrid attention MoE architecture combines 1.6 trillion total parameters with 49 billion active parameters per token and a context window of one million tokens. The model is available as an Open Weights model under the MIT license, though Chinese jurisdiction in cloud deployments requires a separate privacy assessment.
NVIDIA Nemotron 3 Ultra is NVIDIA’s Frontier reasoning model with 550 billion total and 55 billion active parameters on a hybrid Mamba-Transformer-MoE architecture with LatentMoE routing and MTP layers. The context window spans one million tokens, and reasoning is configurable. Native tool calls and agentic orchestration are supported; available as an Open Weights model under the NVIDIA Open Model License.
At 350 output tokens per second, Gemini 3.5 Flash-Lite is the fastest model in Google’s 3.5 series, generally available since July 21, 2026. It targets high-volume, latency-sensitive workloads: translation, classification, and light agentic tasks at $0.30 / $2.50 per million tokens. One million tokens of context, multimodal input across text, image, video, and audio, native tool support. Hard reasoning tasks are not its domain.
What most Edge models can’t do: process text, image, audio, and video in a single weight package without a separate multimodal projector file. Gemma 4 E4B uses per-layer embeddings for 4.5 billion effective parameters and runs on Edge hardware with around 5 gigabytes of VRAM. Configurable thinking modes and a 128,000-token context under the Apache 2.0 license round out the profile.
This community fine-tune variant of Qwen 3.6 35B-A3B removes the safety filters and delivers unfiltered responses without Refusals. Of the 35 billion total parameters in the MoE architecture, only 3 billion are active per token; the context window spans 262,000 tokens. Operable locally at near-full quality with Q8 quantization under the Apache 2.0 license, with multimodal processing for text, image, and video.
Muse Glimmer 30B (August 10, 2026) is Meta Superintelligence Labs’ first open model under the Apache 2.0 license, running without an EU exclusion clause for local deployment. The dense 29.6-billion-parameter model with an additional 1.8-billion-parameter vision encoder processes text and images in a 131,072-token context and delivers up to 233 tokens/second on a consumer GPU via DFlash Speculative Decoder — Workstation-class with genuine Desktop capability.
Ministral 3 8B by Mistral AI is the middle member of the Ministral 3 family, positioning itself as an Edge model with genuine agent potential. 8.8B dense parameters, 256,000 tokens of context, multimodal input for text and image, native function calling and JSON output — Apache 2.0, runnable locally as an Unsloth GGUF.
Grok 4.7 is xAI’s Frontier flagship as of September 21, 2026, built on a larger base than Grok 4.6 and trained with a focus on multi-hour tasks. 500,000 tokens of context, text and image input, four reasoning levels from low to xhigh at unchanged pricing of 2 / 6 USD per million tokens. According to xAI, nearly doubled performance on long terminal tasks and a new safeguard stack against jailbreaks.
Qwen 3.5 27B is Alibaba’s dense 27-billion-parameter variant of the Qwen3.5 family (February 2026), an open Apache-2.0 release with native multimodality for text, image, and video. The hybrid architecture of Gated-DeltaNet and Gated-Attention blocks delivers 262,144 tokens of context (extensible to approximately one million via YaRN), up to 65,536 tokens of output, and strong coding performance (SWE-bench Verified 72.4) — the only dense variant in the Qwen3.5 family.
Xiaomi MiMo V2.5 is a natively omnimodal MoE model with 310 billion total and 15 billion active parameters. The model processes text, image, video, and audio within a single architecture; the context window supports up to one million tokens. Fully commercially usable under the MIT license, from a Chinese manufacturer jurisdiction with a corresponding assessment for cloud deployment.
Devstral 2 is Mistral AI’s code agent for demanding software engineering tasks. With 123 billion parameters, the model operates on codebase exploration, multi-file changes, debugging, and legacy modernization, supports a context window of 256,000 tokens, and processes both text and image inputs. Available as an Open Weights model under a modified MIT license, from a European provider environment with GDPR compliance.
GPT-5.4 Mini is the compact GPT-5.4 variant for fast and cost-efficient everyday tasks. With a context window of 272,000 tokens and multimodal input for text and image, the model targets applications requiring low latency with solid output quality. Available exclusively via the OpenAI API.
Qwen 3.5 4B is a compact Open Weights model from Alibaba featuring a Thinking-Optional architecture and Unsloth Dynamic quantization at Q6. The Q6 tier offers higher fidelity than the Q4 variant at a moderate VRAM overhead, with a context window of 128,000 tokens. Deployable locally on resource-constrained hardware under the Apache 2.0 license.
NVIDIA Nemotron 3.5 Lightning is an open 30-billion-parameter MoE with 3 billion active parameters per token (August 11, 2026), distilled from Nemotron 3 Ultra and specialized for the execution layer of always-on agents. The hybrid Mamba-2 + MoE + Attention architecture under the OpenMDW-1.1 license offers up to 1 million tokens of context and up to 4× output speed through Multi-Token Prediction and Speculative Decoding.
o4-mini is OpenAI’s compact reasoning model with native vision input for images, diagrams, and screenshots. The model processes text and image, operates with a context window of 200,000 tokens, and offers three adjustable reasoning levels for balancing response depth and latency. Full tool use including parallel tool calling for lightweight agentic workflows.
GPT-5.4 is the larger variant from OpenAI’s 5.4 generation for demanding workloads, delivering higher response quality than the Mini versions. The model operates with a context window of 272,000 tokens, processes text and image inputs, and is available exclusively via the OpenAI API. Proprietary and designed for productive, complex tasks.
Laguna S 2.1 has been Poolside’s third consecutive open coding model since July 2026, a MoE with 117.6 billion total and only 8.5 billion active parameters per token. The Linux Foundation’s OpenMDW-1.1 license permits free commercial use and modification. The model supports 1,024,000 tokens of context, interleaved thinking with tool calls within the chain of thought, and native tool calling for long-horizon agentic workflows.
Hermes 4 70B is an open instruct and reasoning model by Nous Research from the Hermes 4 family with 70 billion parameters. The model combines optional thinking with advanced tool use and structured outputs, trained for high steerability and reduced Refusal rates. Available as an Open Weights model under a Modified MIT license for local or server-side deployment.
Gemma 3 12B Instruct as a Q4 quantization, optimized for local inference on resource-constrained hardware. The model processes text and image inputs with a context window of 128,000 tokens, is designed for direct task execution, and operates without an external cloud connection. Licensed under the Google Gemma Terms of Use, which permit commercial use.
Qwen 3.5 9B is an Open Weights model by Alibaba for general language and reasoning tasks, featuring a thinking-optional architecture. The Unsloth Dynamic Q6 quantization delivers near-lossless quality at significantly reduced VRAM requirements; the context window spans 128,000 tokens. Runs under the Apache 2.0 license on local hardware without cloud dependency.
Hermes 4 405B is a high-performance instruct and reasoning model from Nous Research with 405 billion parameters, designed for complex reasoning tasks and agentic workflows. The model supports optional thinking, precise tool calls, and structured outputs. Trained for high steerability and reduced Refusal rates. Available as a Restricted Weights model under the Meta Llama Community License.
NVIDIA Nemotron 3 Nano 30B A3B is an efficient hybrid model from the Nemotron-3 series, combining Mamba-2 with Transformer layers. With 31.6 billion total parameters, the model activates only 3.2 billion per token; the context window supports up to one million tokens. Optional thinking mode with configurable budget, native tool calls, and agentic capabilities out of the box. Available as an Open Weights model under the NVIDIA Open Model License.
Hermes 4 14B Abliterated is a locally deployable Open Weights variant by NousResearch based on Qwen 3 14B, with safety mechanisms deliberately removed. With 14 billion parameters and a 128,000-token context window, the model targets unfiltered response behavior, creative writing workflows, and scenarios requiring maximum openness. Fully commercially usable under the Apache 2.0 license.
Gemma 4 E2B is the most compact Gemma 4 model and the entry point into the new family under the Apache 2.0 license. PLE architecture with 5.1B total and 2.3B effective parameters, 128,000 tokens of context, multimodal input for text, image, and audio, plus built-in reasoning and function calling. Runs locally as Unsloth GGUF on Edge- and phone-class hardware.
GPT-5.4 Nano is OpenAI’s most affordable GPT-5.4 variant for high-volume standard tasks such as classification, extraction, and ranking. With a context window of 272,000 tokens and up to 128,000 tokens of output, the model is well-suited for batch processing and sub-agent routing. Available exclusively via the OpenAI API at low cost.
o3-mini is OpenAI’s compact reasoning model with internal chain-of-thought, specialized in math, coding, and STEM tasks. The model operates with a context window of 200,000 tokens and offers three adjustable reasoning levels for balancing response depth, latency, and cost. Available exclusively via the OpenAI API.
Hermes 4.3 36B (NVFP4) is the compressed variant of NousResearch’s Open Weights model based on ByteDance Seed-OSS-36B, fully decentralized post-trained via the Psyche network. Fireworks compressed the BF16 checkpoint to NVFP4, significantly reducing the memory footprint of the weights. 36 billion dense parameters, 512,000 tokens of native context, optional thinking mode, and an Apache 2.0 license for free commercial use.
Qwen 3 14B is Alibaba’s open-weights model for general language tasks and reasoning with an optional thinking mode. The Q6 quantization is designed for efficient local operation without a cloud connection; the context window covers 128,000 tokens. Fully commercially usable under the Apache 2.0 license.
Hermes 4 14B as a Q4 quantization of the NousResearch distribution based on Qwen-3, optimized for local assistance and agentic tasks. With 14 billion parameters and a 128,000-token context window, the model runs on resource-constrained hardware and supports hybrid reasoning modes. Fully commercially usable under the Apache 2.0 license.
Hermes 4.3 36B (NVFP4) is the compressed variant of NousResearch’s Open Weights model based on ByteDance Seed-OSS-36B, fully decentralized post-trained via the Psyche network. Fireworks compressed the BF16 checkpoint to NVFP4, significantly reducing the memory footprint of the weights. 36 billion dense parameters, 512,000 tokens of native context, optional thinking mode, and an Apache 2.0 license for free commercial use.
Ornith 1.0 9B is DeepReinforce’s compact first-generation reasoning model, built on a Qwen-3.5 base and delivered as an Unsloth-GGUF for local Edge setups. 9.4B dense parameters, 262,000 tokens of context, MIT license — a reasoning-first profile for agentic coding workflows rather than a classic chat generalist.
Trained for code tasks, not for general use: Codestral 25.08 is Mistral AI’s specialized developer model from August 2025, optimized for fill-in-the-middle and a broad range of programming languages. With 22 billion parameters, it supports 128,000 tokens of context, runs either locally or via the Mistral API, but is subject to a restrictive license with limited commercial use.
What most Nano models don’t offer: native multimodal input and tool calling right out of the box. Ministral 3 3B by Mistral AI delivers exactly that in the 3B class, with 256,000 tokens of context, an Apache 2.0 license, and local Unsloth GGUF distribution.
What most OpenAI models can’t do: GPT-OSS 20B is OpenAI’s first Open Weights release since GPT-2 (August 5, 2025) under the Apache 2.0 license. The MoE with 21 billion total and 3.6 billion active parameters runs on a single consumer GPU with only around 16 GB of memory thanks to native MXFP4 quantization, supports tool use via the Harmony format, and offers 131,072 tokens of context as well as configurable reasoning intensity (low/medium/high).
This 8B GGUF artifact carries a Llama-3.3 name, but Meta’s official Llama-3.3 is the 70B text-only model — the 8B version listed here is an unofficial community construction with no verified upstream source. 8B dense parameters, 128,000-token context, locally runnable as an Unsloth GGUF under the Llama Community License. Upstream provenance remains unconfirmed pending verification.
Gemma 3 4B is the most unusual Nano-class variant in the CrucibleMark portfolio: a multimodal Google DeepMind model with text and image processing at just 4 billion parameters. 128,000 tokens of context, Gemma terms of use, locally deployable as Unsloth-GGUF — the most compact multimodal member of the Gemma 3 family.
Four billion parameters, Apache 2.0 license, and an optional thinking mode: Qwen 3 4B is designed for mobile applications and edge setups where every watt counts. Locally deployable under Q6 quantization, with a 128,000-token context window — generously sized for a Nano-class model. The manufacturer’s jurisdiction of China occasionally results in censored or evasive responses on politically sensitive topics.
Hermes 3 8B is an uncensored fine-tune by NousResearch based on Meta’s Llama 3.1 8B. With eight billion parameters and a 128,000-token context window, the model targets instruction following, tool use, and creative or ambiguous requests. Deployable locally under the Llama 3.1 Community License, the reduced refusal rate is a defining characteristic of this distribution.
3.8B dense parameters, synthetic training data, and a pronounced reasoning specialization: Phi-4-mini is Microsoft’s compact model for math, logic, and structured output. MIT license, 128,000 tokens of context, locally deployable as an Unsloth GGUF — not a broad generalist, but a specialist at Nano scale.
The strongest variant in the DeepSeek-R1-Distill line: DeepSeek-R1-Distill-Qwen-14B brings 14.8B dense parameters on a Qwen-2.5-14B base with full R1 reasoning structures into the Desktop range. 128,000 tokens of context, MIT license, locally as Unsloth-GGUF — the reasoning compromise between Edge suitability and Workstation capacity.
A Q6_K-GGUF distribution of Qwen 2.5 Coder 7B for local coding: 7.6 billion dense parameters, Apache-2.0 license, specialized in code generation, debugging, and repair. The family supports 128,000 tokens of context; in GGUF setups, only 32,000 are natively available without long-context configuration. Fully commercially usable, compact and efficient on Workstation hardware.
3.21B dense parameters, 128,000 tokens context: Llama 3.2 3B is Meta’s compact text-only variant of the Llama 3.2 family for local tasks such as summarization, paraphrasing, and instruction-following. Unsloth GGUF build, Llama 3.2 Community License, fully operable offline.
Llama 3.2 1B is Meta’s smallest Llama 3.2 model and a text-only baseline for edge setups. 1.23B dense parameters, 128,000 token context, locally deployable as Unsloth GGUF under the Llama 3.2 Community License. Suitable as a compact on-device reference, not as a quality anchor.
What the small distill models lack: the complete reasoning pattern of the R1 line in an edge-suitable footprint. DeepSeek-R1-Distill-Qwen-7B brings 7.6B dense parameters on a Qwen-2.5-Math-7B base, 128,000 tokens of context, an MIT license, and runs locally as an Unsloth GGUF on consumer hardware.
One and a half billion parameters, distilled on a Qwen 2.5 base: DeepSeek-R1-Distill-Qwen-1.5B brings the reasoning behavior of the large R1 models into the Nano class. MIT license, locally operable as an Unsloth GGUF, designed for reasoning tasks rather than chat polish.
With 270 million parameters, Gemma 3 270M is the smallest Gemma 3 model and a text-only model for local latency baselines and embedded setups. The Unsloth GGUF variant runs fully offline under the Gemma terms of use, but is not suitable as a general-purpose quality anchor.
