2.LLM Industry Overview
👉 #AI #LLM
1. 2026 AI Power Rankings — Core Tech and Ecosystem Mapping
📅 Mon. 2026-04-06 🕐 09:31
This note is based on the latest April-2026 industry tech architecture and gives an in-depth dissection of the "technical soul" of today's mainstream LLMs.
1.1. Cross-Model Comparison Table (2026.04)
| Vendor | Flagship model | Released | Core trait (one-liner, in depth) | Pricing |
|---|---|---|---|---|
| Anthropic | Claude 4.6 Opus | 2026-02 | Environment manipulation master: the only model with native "computer use" + complex Agent loops in perfect alignment. | Pro ($20/mo) / Enterprise API |
| OpenAI | GPT-5.4 | 2026-03 | Multimodal reasoning hub: deepest cross-modal reasoning with OS-level automation control. | Plus ($20/mo) / metered API |
| Gemini 3.5 Pro | 2026-02 | Native infinite-context streaming: 2M+ context, the benchmark for native real-time A/V stream processing. | AI Premium / Vertex AI | |
| Perplexity | Pro Search V3 | 2026-Q1 | Real-time knowledge synthesizer: not a single model but a multi-model orchestration that produces optimal search reasoning. | Pro ($20/mo) |
| DeepSeek | DeepSeek V4 | 2026-02 | Cost-efficient reasoning king: Engram memory architecture; reasoning rivaling top models at extremely low cost. | Very low-cost API / open weights |
| Mistral AI | Mistral Small 4 | 2026-03 | Controllable open intelligence: an MoE that lets users switch "fast response" vs. "deep thought". | API / free open download |
| Microsoft | Copilot (2026) | continuous | Ecosystem parasite: core value is deep permissions on Office 365 data and enterprise compliance. | Enterprise subscription / M365 bundles |
| Meta | Llama 4 (Scout/Maverick) | 2025/2026 | Privacy and freedom fortress: ceiling of locally-run performance, no cloud dependence, ideal for minimalists. | Free open source (Apache 2.0) |
1.2. Core Models — Deep Technical Notes
- Anthropic: Claude 4.6 (code + environment alignment)
- Claude's core advantage is the evolution of Constitutional AI.
- Version 4.6 introduces Native Computer Use weights — not a simple API wrapper, but specialized attention heads in the underlying Transformer trained on GUI pixel recognition and coordinate mapping.
- With OpenClaw assistance, it uses Chain-of-Action (CoA) to break long-horizon goals into dozens of backtrackable sub-tasks, exhibiting very low "instruction drift" across multi-turn dialogue.
- OpenAI: GPT-5.4 (general reasoning + long-horizon prediction)
- GPT-5.4 retains its dominance in reasoning depth.
- It introduces Mid-Response Course Correction (MRCC) — when generating long code or essays, internal verification branches detect and fix logical drift in real time.
- Its Tool Search mechanism, optimized with vector indexing, cuts token waste by 40%+ during complex Agent invocations, making it shine in multi-Agent collaboration frameworks like OpenCrow.
- Google: Gemini 3.5 Pro (native multimodal + long context)
- Gemini's killer feature is its Ring Attention algorithm, enabling stable processing of up to 2 million tokens without recall loss.
- Unlike other models that bolted on long context after the fact, Gemini was natively trained for it.
- It can take an entire HD video as input (using Video-to-Token compression) and align semantics at the pixel level — currently the best choice for handling large technical document libraries.
- DeepSeek: DeepSeek V4 (Engram memory architecture)
- The most disruptive model of 2026.
- DeepSeek V4 uses an Engram (conditional storage) architecture: "static knowledge" lives in something resembling an O(1) hash lookup, while Transformer compute focuses on "dynamic reasoning".
- This Knowing–Thinking decoupling lets it match or exceed GPT-5 on high-frequency logical tasks like code generation at a tiny fraction of the compute cost — the preferred backend for lightweight tools like Kiro.
- Mistral AI: Mistral Small 4 (granular MoE)
- Mistral perfectly executes Granular Mixture-of-Experts (G-MoE).
- It has 128 expert nodes but activates only 4 per inference; this extreme weighted sparsity gives it unmatched inference speed.
- Its Reasoning Effort protocol lets users manually tune reasoning depth, switching dynamically between "fast instruction execution" and "deep thinking (O1-style)".
- Meta: Llama 4 (Maverick line)
- A blessing for local enthusiasts.
- Extreme optimization on RoPE (Rotary Position Embedding) dramatically improves stability of long-text inference on local GPUs with limited VRAM.
- Retains classics like SwiGLU activation and RMSNorm but optimizes operators for 2026 AI-PC hardware — currently the performance peak when running locally with Ollama via Ghostty terminal.
- Perplexity: Pro Search V3 (RAG peak)
- Perplexity's edge is not the base model but its Reasoning-over-Search (RoS) architecture.
- It auto-decides search depth based on the question's entropy and uses Multi-Query Expansion to gather real-time information from multiple angles in parallel.
- It is no longer just Q&A — Pages uses Constrained Generation to turn fragmented search results into structured, rigorously cited professional reports.
- Microsoft: Copilot (2026 enterprise agentic layer)
- Copilot has evolved into an Agentic layer.
- Its core is the Work IQ engine, which auto-connects to the Dataverse enterprise database.
- It uses Hybrid Search (combining vector search with traditional relational queries), giving the AI access to all your company's emails, documents, and even ERP data, while maintaining strict tenant-level isolation for privacy.
- It is currently the deepest integration of an LLM with complex business processes (like financial close).
1.3. Tool Ecosystem Distribution by Model
- Claude (Anthropic) zone
- Claude Code: official CLI; directly controls the terminal for fully automated dev loops.
- OpenClaw (primary alignment): supports multiple models, but its Computer Use node performs most reliably under Claude.
- Artifacts: the killer code-preview feature in the web UI; dramatically reduces user "verification cost".
- Gemini (Google) zone
- Antigravity: Google's own custom IDE (a VS Code fork) with great support for Gemini Live audio debugging.
- NotebookLM: knowledge-management workhorse; supports 2M-context-level document analysis.
- AI Studio: official advanced developer platform; great for fine-tuning long prompts.
- GPT (OpenAI) zone
- Canvas: an immersive workspace for long-form writing and code refactoring.
- ChatGPT Search: a real-time-info-stream tool that replaces parts of traditional search.
- Perplexity zone
- Pages: automated technical-wiki generator.
- Sonar: Perplexity's in-house lightweight high-speed search model.
- DeepSeek zone
- Kiro (recommended backend): the VS Code fork you use; pairing it with DeepSeek's low-latency API yields excellent real-time code completion.
- DeepSeek Coder: a dedicated web interface focused on code logic.
- Mistral zone
- Le Chat: Multi-Agent collaboration platform that can orchestrate multiple expert models simultaneously.
- Pixtral: vertical app specialized for image understanding and chart analysis.
- Local / Llama zone
- Ollama / LM Studio: the "heart" of running local models.
- Ghostty: your terminal — invokes the local Llama 4 model directly via Local Agent CLI.
- Cross-platform integration tools
- Surf (Obsidian Plugin): your core knowledge-base plugin; recommended config — Gemini for retrieval, Claude for logical organization, DeepSeek for low-cost error correction.
1.X. Questions
- Given your STEM background and lifelong-learning mindset, have you tried using an MCP server to expose your Obsidian vault as long-term memory directly to Claude or Gemini?
- Now that you use Kiro and an MCP server, which model's API is your local Academic Agent currently driven by? Have you hit token-limit or recall problems on long-form PDF papers?
- For the Academic Agent project, have you tried using DeepSeek V4 as a "pre-screening model" to filter useless literature first, then doing 2M-context deep analysis with Gemini 3.5? This hot/cold tiering can significantly cut API cost.
2. Global LLM Industry Overview — Three Competitive Dimensions
📅 Sun. 2025-11-09 🕐 10:21
A comprehensive overview and deep analysis of the global LLM industry. We classify all 13 core players in the AI ecosystem into three categories and compare them across capability, pricing, public-market performance, and profitability outlook.
2.0. Introduction
The LLM industry has evolved from a pure technology race into a comprehensive war of ecosystem, application integration, and cost-effectiveness. Core competitors fall into three groups: foundation model providers (core technology), platform integrators (application & productivity), and infrastructure / specialty service providers (enablers). This report cross-analyzes the 13 main players.
2.1. Part 1: Core Foundation Model Providers
These companies are the source of AI technology. They train and release the strongest LLMs and serve developers and enterprises globally via APIs. Competition focuses on raw performance, long context, and multimodal capability.
(1) Capability, pros/cons, and pricing
| Player | Pros | Cons | Top-tier API price (input/output per 1M tokens) |
|---|---|---|---|
| OpenAI (GPT-5) | All-around leader: best general reasoning, coding, multimodal; most complete ecosystem and consumer platform (ChatGPT). | Closed, expensive, limited data control. | ≈ $10.00 / $40.00 (GPT-5) |
| Google (Gemini 2.5 Pro) | Ecosystem integration: deeply tied to Google Workspace/Cloud; very long context (1M+ tokens), native multimodal. | Slightly behind on some general benchmarks at the top tier. | ≈ $12.50 / $35.00 (Gemini 2.5 Pro) |
| Anthropic (Claude 3.7 Opus) | Safety + ethical alignment: fits regulated industries; long-doc understanding and clear, natural output are excellent. | Opus output is expensive; safety limits sometimes overly conservative. | ≈ $15.00 / $75.00 (Claude 3.7 Opus) |
| Meta (Llama 3) | Open-weights leader: free for commercial use; highest privacy and customizability; massive developer community. | Lower out-of-the-box readiness; enterprises must invest in deployment + tuning; raw performance ceiling lower than top closed models. | Free (you pay deployment + inference cost) |
| Mistral AI (Mistral Large) | High value-for-money: MoE delivers excellent perf/cost ratio; strong coding. | Lacks a large consumer ecosystem; brand awareness still growing. | ≈ $8.00 / $24.00 (Mistral Large) |
| Grok (xAI) | Real-time data: direct access to X (Twitter) data stream; unique humor and unconstrained persona. | Limited availability (X Premium+); reliability and seriousness below mainstream models. | Subscription (X Premium+ bundle) |
| Cohere (Command R+) | Enterprise-grade RAG: focused on enterprise retrieval-augmented generation and multilingual. | Lower visibility and weaker general performance than the big three. | ≈ $2.50 / $10.00 (Command R+) |
| AI21 Labs (Jamba) | Innovative MoE architecture: hybrid architecture for high efficiency and large context window. | Smaller market share; mainly a tech-driven challenger. | ≈ $3.50 / $14.00 (Jamba Large) |
(2) Public-market performance and profitability outlook (core LLM)
| Player (ticker) | Business model / outlook | Market view (end-2025) |
|---|---|---|
| OpenAI (private) | High-margin API + ChatGPT subscriptions. | Highest private-market valuation; strong profitability; focus on long-term moat and future IPO timing. |
| Google (GOOGL) | AI-empowered Search/Cloud/Workspace; Gemini drives Cloud + subscription growth. | Financially solid; heavy AI investment, but market favors data + distribution advantages for long-term profit. |
| Anthropic (private) | Premium B2B/B2G API, focused on safety and quality. | Backed by Amazon, Google; very high valuation; targets high-margin enterprise + safety + compliance market. |
| Meta (META) | Ad revenue; Llama is an ecosystem defense strategy. | Very profitable; Llama solidifies its AI position and is seen as a long-term growth catalyst. |
| Mistral / xAI (private) | High-growth challenger; raises capital via tech and unique positioning. | Massive growth potential; favored European/American AI darlings; short-term focus on market share over profit. |
2.2. Part 2: Platform Integrators & Application Layer
These companies turn LLM tech into end-user productivity or info tools. Competition focuses on UX, integration depth, and feature specialization.
(1) Capability, pros/cons, and pricing
| Player | Pros | Cons | Pricing model |
|---|---|---|---|
| Microsoft Copilot | Ecosystem integration: deep Microsoft 365 integration; default AI assistant for hundreds of millions of enterprise users. | Performance depends on OpenAI; limited outside Microsoft. | Subscription (≈ $30/user/month for M365 edition) |
| Perplexity | High-transparency search: focused on real-time retrieval with summaries that include precise citations. | As a pure chatbot, less creative than GPT-4. | Subscription (≈ $20/month Pro) and free tier |
(2) Public-market performance and profitability outlook (application layer)
| Player (ticker) | Business model / outlook | Market view (end-2025) |
|---|---|---|
| Microsoft (MSFT) | Platform subscription revenue; Copilot lifts M365 margins, Azure underpins OpenAI. | Extremely bullish; Copilot regarded as one of the fastest-growing enterprise products ever, central to revenue and market cap. |
| Perplexity (private) | Pro subscriptions; offers ad-free, high-quality search as an alternative to traditional search engines. | High-growth unicorn; near-term focus on share + users; significant IPO or acquisition potential. |
2.3. Part 3: Infrastructure & Global Challengers
These companies play key roles in the AI supply chain or dominate specific geographic markets.
(1) Capability, pros/cons, and pricing
| Player | Pros | Cons | Pricing model |
|---|---|---|---|
| NVIDIA | Hardware monopolist: H100/A100 GPUs are required for all LLM training and inference; strong software stack (CUDA, Triton). | Does not build general LLMs; profits concentrated in hardware sales, vulnerable to supply-chain and tech cycles. | High hardware prices; software services (AI Enterprise) are subscription (≈ $4,500/GPU/year) |
| AWS (Bedrock) | LLM aggregator platform: powerful cloud infrastructure + Bedrock (hosts Anthropic, Cohere, Meta, etc.). | Own Titan models have minor influence; competitive edge is in platform + services. | Multi-model on-demand or provisioned throughput, complex pricing depending on the model. |
| Databricks (DBRX) | Data intelligence platform: focus on lakehouse-based AI; efficient LLM deployment + tuning. | Niche; mainly serves enterprises with large data assets. | Cloud-compute usage / resource-based pricing |
| Alibaba (Qwen) | Asia LLM giant: standout in Chinese and multilingual processing; deep Alibaba Cloud integration. | Less global influence than US giants; mainly Asia. | Pay-as-you-go API (Alibaba Cloud) |
| Baidu (ERNIE Bot) | Chinese NLP heritage: dominant in China search and apps; aggressive low-price/free strategy. | Geographic limits and limited global reach. | Free tier (ERNIE 3.5); paid sub (ERNIE 4.0 Pro ≈ $8.2/month); PAYG API (very competitive — ERNIE 4.5 ≈ $0.55 / $2.2 per 1M tokens) |
(2) Public-market performance and profitability outlook (infrastructure / global)
| Player (ticker) | Business model / outlook | Market view (end-2025) |
|---|---|---|
| NVIDIA (NVDA) | AI hardware and data-center revenue. | Extremely high valuation; biggest winner of the AI era; profits very strong and consistently above expectations; near-term volatility but long-term position solid. |
| AWS (AMZN) | Cloud-services revenue; Bedrock drives high-margin cloud growth. | Financially solid; AWS growth is Amazon's biggest profit driver; Bedrock cements its place in enterprise AI deployment. |
| Alibaba (BABA) | Cloud + e-commerce; Qwen / Alibaba Cloud drive cloud growth. | Constrained by domestic competition and macro environment; AI is the key strategic lever for cloud recovery and growth. |
| Baidu (BIDU) | Search + cloud; ERNIE Bot is the AI upgrade of Baidu's ecosystem. | Benefits from domestic AI adoption and cost advantage; AI is the main hope for growth and valuation re-rating. |
2.4. 🔑 Industry Insights — Summary
- Performance convergence + cost war: top closed models (GPT, Claude, Gemini) are narrowing the raw-performance gap; competition is shifting to cost-effectiveness (e.g., Mistral's MoE) and specialization (e.g., Cohere's RAG).
- Platform + integration wins: for end users, application-layer integration matters more than raw model performance; Microsoft Copilot proves that embedding AI in existing workflows is the biggest profit opportunity.
- Open-weight and closed balance: open-weights models like Llama and Mistral are rapidly improving and meet enterprise needs for privacy, customization, and on-prem; they exert serious pressure on closed-API markets.
- Decisive role of infrastructure: NVIDIA and AWS are critical; NVIDIA controls the AI "oil" (GPUs), AWS controls the "refinery" (cloud); both are long-term, steady winners.
2.X. Appendix
(1) Pricing categories
LLM pricing typically falls in two categories: subscription for consumer access (chatbots) and pay-as-you-go for API access (developers).
- Subscription pricing (direct-to-consumer)
| Model | Service | Price | Notes |
|---|---|---|---|
| OpenAI | ChatGPT Plus | $20.00/mon | Access to GPT-4o, GPT-5, browsing, custom GPTs, image generation (DALL-E 3). |
| Google One AI Premium | $19.99/mon | Access to Gemini 2.5 Pro (Advanced), 2 TB Google One storage, integration with Workspace apps. | |
| Anthropic | Claude Pro | $20.00/mon | Access to Claude 3.7 Opus, higher usage limits, early access to features. |
- API pricing (developer / enterprise — per 1 million tokens) API pricing is tiered by model capability. Input tokens are for the prompt you send; Output tokens are for the model response. Output tokens are usually significantly more expensive.
| Model variant | Input price (/1M tokens) | Output price (/1M tokens) | Speed/cost niche |
|---|---|---|---|
| GPT-4o Mini | $0.15 | $0.60 | Best value for high-speed, general tasks. |
| Gemini 2.5 Flash | $0.75 | $1.50 | Cost-effective for bulk processing and speed. |
| Claude 3 Haiku | $0.25 | $1.25 | Fastest, most affordable Claude for simple tasks. |
| GPT-5 | $10.00 | $40.00 | Premium intelligence and reasoning. |
| Gemini 2.5 Pro | $12.50 | $35.00 | Premium context and deep multimodal analysis. |
| Claude 3.7 Opus | $15.00 | $75.00 | Highest output price, used for the most complex, high-value tasks. |
3. Open-Source Local LLMs
3.1. Google — Gemma
- A new model family for different compute scenarios:
- E2B & E4B (edge models)
- Geek-grade models optimized for smartphones and IoT.
- 128K context with native on-device multimodal interaction.
- 9B & 12B (desktop-grade)
- Excellent value-for-money: code-Copilot engines and local RAG bases.
- 26B A4B (MoE)
- Efficiency monsters using the MoE mechanism; complex-reasoning performance far above same-parameter peers.
- 31B (dense flagship)
- Exclusive 256K context; replaces the previous 27B as the flagship base model for private deployment.
3.2. Alibaba — Qwen
For your machine (M4 chip + 24 GB unified memory), let's first acknowledge the hardware advantage: macOS unified memory lets the GPU dynamically draw from system RAM. Typically up to about 70% can be used as VRAM (≈ 16-18 GB usable VRAM).
That means you can step out of the 6B/7B trap and enjoy mid-size (14B-class) models with strong reasoning, and even push 32B-class quantized models in extreme cases.
Below is a curated list of small-to-medium Qwen models that fit your machine, organized by use case.
(1) Daily workhorse (the sweet spot): 14B class
This is the golden sweet spot for your machine. 14B models are a real qualitative leap in logical reasoning, long-text understanding, and multilingual ability — and the quantized variants run smoothly.
- Qwen 2.5 (14B)
- Position: an all-around daily base model.
- VRAM (INT4 quant): ~9-10 GB.
- Experience: a perfect fit for 24 GB; leaves ample memory for OS and IDE (Cursor / VS Code), and easily handles 8K or 16K context.
- Command: ollama run qwen2.5:14b
(2) Coding-specific: the Coder line
If you primarily use AI to assist coding, debugging, or analyzing project structure, the Qwen Coder line is the undisputed T0 in its class.
- Qwen 2.5 Coder (7B)
- Position: extremely fast code completion and lightweight refactoring; beats older, larger models on many coding benchmarks.
- VRAM (INT4 quant): ~4.5-5 GB.
- Experience: lightning fast; ideal as a background service for local IDE plugins (called via Ollama API), without crowding out your dozens of Chrome tabs.
- Command: ollama run qwen2.5-coder:7b
- Qwen 2.5 Coder (14B) (or any newer Qwen 3 Coder)
- Position: complex architecture design and deep system-level code review.
- Command: ollama run qwen2.5-coder:14b
(3) Multimodal vision-language: the VL series
You saw the recommendation qwen2.5vl:7b; that's a great pick because Qwen's VL models have a strong reputation in the open-source world.
- Qwen 2.5 VL (7B)
- Position: image parsing, web-UI screenshot-to-code, complex chart extraction.
- VRAM (INT4 quant): ~5 GB for the model; HD images add 2-3 GB peak via the visual encoder.
- Experience: 16 GB+ of available VRAM handles its multimodal reasoning comfortably.
- Command: ollama run qwen2.5-vl:7b (when testing images at the CLI, you can drag and drop the image file onto the terminal to provide the path).
(4) Pushing the limits: 32B class
Although you asked about "small models", I strongly recommend trying Qwen's 32B version when you don't need to run heavy software (large games, heavy video editing).
- Qwen 2.5 (32B)
- Position: very high-IQ complex logical reasoning.
- VRAM (Q4_K_M quant): about 19-20 GB.
- Experience: bleeding-edge fit; a 24 GB Mac may need to use a small amount of swap. Inference drops from "instant" 14B speed to "steady typewriter" (~15-25 tokens/s), but the answer quality is impressive.
- Command: ollama run qwen2.5:32b
💡 Configuration recommendation
For an M4 with 24 GB, the most comfortable daily combo is Qwen2.5:14B (or Qwen2.5-Coder:14B) as your default large model. Their IQ far surpasses 7B-class while staying within ~10 GB VRAM — local, uncensored, zero-latency, high-IQ.
3.3. High-Flyer — DeepSeek (code / math)
📅 2026-04-06 21:55 CDT; Gemini Pro 3.1 📎 DeepSeek Official Models | HuggingFace
(1) Overview
- DeepSeek is a series of open-source models from an AI lab founded by Chinese quant giant High-Flyer.
- Design intent: extreme cost-efficiency and algorithmic innovation; aims to break the "scaling-law brute force" approach by using smarter architectures to do more with less.
- Pain points solved
- Inference cost: through underlying architecture redesign, API inference cost is driven to industry-low levels (about 1/dozens that of GPT-4o).
- Math & coding: specifically reinforced for structured logic tasks, addressing open-source weaknesses in complex code construction.
- Reasoning: the R1 series successfully replicated and open-sourced OpenAI o1-style chain-of-thought.
- Core features
- Extreme MoE: more granular expert splitting (Fine-grained Experts) than other MoE designs.
- Memory efficiency: original architecture significantly cuts memory usage during long-context inference.
- RL-First: in the R1 era, it proved that pure reinforcement learning alone can elicit strong reflection and reasoning.
- Use cases
- Code generation: extremely strong code refactoring, debugging, and system architecture design.
- Complex mathematics: olympiad-level logical proofs and derivations.
- Agentic core: very low cost makes it ideal as the brain for multi-step, multi-Agent workflows.
- Competitors
- OpenAI o1/o3: DeepSeek-R1 is currently the only open-source contender that holds its own on deep-reasoning blind benchmarks against OpenAI's flagship.
- Qwen: Qwen wins on broad parameter-size coverage and bilingual general knowledge; DeepSeek wins on STEM (code/math) single-point breakthroughs.
(2) Concept, components, & architecture
DeepSeek's global impact comes from several disruptive low-level components.
1. MLA (Multi-head Latent Attention)
- Pain: when handling 128K long text, traditional KV-cache eats massive memory and triggers OOM.
- Breakthrough: MLA uses low-rank compression to compress the huge KV matrix into a latent state.
- Benefit: KV-cache footprint drops by 90%+ during inference — a godsend for your 24 GB Mac, allowing longer context without VRAM blowout.
2. DeepSeekMoE — fine-grained expert architecture
- Mechanism: traditional MoE (e.g., Mixtral) has 8 large experts with 2 active per pass; DeepSeek splits into up to 256 micro-experts with 8 active, plus a few always-on shared experts to cover general knowledge.
- Benefit: routing is more accurate; with very few active parameters it matches the performance of large dense models.
3. DeepThought (reasoning paradigm)
- The R1 model first does extensive self-correction, hypothesis testing, and logical reasoning inside a <think> tag before producing the final answer.
- This built-in System-2 slow-thinking lets it "design first, then code" on complex coding problems.
(3) DeepSeek for Mac M4 (24 GB) — recommendations
With your M4 + 24 GB unified memory, the official 671B flagship is out of reach, but the ecosystem has tools tailor-made for your machine.
1. DeepSeek-Coder-V2-Lite (16B)
- Position: a code-specialized MoE built for desktop.
- Hardware fit: 16B total parameters but MoE activates only ~2.4B per pass; INT4 quantization fits in under 10 GB on your M4.
- Feel: very low latency and very high code-pass rates — the best replacement for GitHub Copilot for offline local coding.
2. DeepSeek-R1-Distill-Qwen (14B)
- Position: distilled deep-reasoning model.
- Background: DeepSeek officially distilled R1's chain-of-thought into Qwen-architecture small models.
- Hardware fit: same size as the 14B Qwen we discussed (~9-10 GB VRAM).
- Feel: combines Qwen's solid base with DeepSeek's high-IQ slow thinking — an absolute "golden sweet spot" on your machine.
3. DeepSeek-R1-Distill-Qwen (32B)
- Position: the limit-pushing local reasoning model.
- Hardware fit: ~19-20 GB; just barely fits in 24 GB.
- Feel: slow (typewriter speed), but its <think> process is genuinely impressive when you ask it to design a complex microservice architecture or debug deep system bugs.
ollama run deepseek-r1:14b # daily very-high-IQ reasoning (recommended primary)
ollama run deepseek-coder-v2 # pure local programming assistant engine
ollama run deepseek-r1:32b # squeeze the M4 to its logical limits
- API & Client integration
- DeepSeek-R1's
<think>tag needs front-end UI support to be collapsed for readability. - Recommended clients: if running locally, pair with an open-source client like Chatbox or AnythingLLM — they natively parse and collapse DeepSeek's chain-of-thought output.