THE OPEN-SOURCE ECOSYSTEM

Layer: Foundation Models · Dossier: open-weights landscape · Status: draft for review · 2026-07

Lede

In January 2025, a Chinese hedge-fund spinoff released a reasoning model you could download, and Nvidia lost roughly $600 billion of market value in a day ⚑ unverified. That was the moment the open-weights world stopped being a hobbyist sideshow. DeepSeek R1 matched the closed frontier on hard reasoning, published its methods, priced its API at a fraction of Western rates, and handed out the weights under MIT. Everything since — OpenAI shipping its first open weights in six years, Meta's license contortions, a release cadence out of China that Western labs cannot match — is downstream of that day.

Here is the honest shape of the market in mid-2026: closed frontier models are still the smartest things you can rent. But the gap has compressed to months, the open models are past the threshold where most real workloads notice the difference, and open gives you three things no closed vendor will ever sell you — weights you control, marginal costs that approach electricity, and the certainty that no deprecation email can kill your product. This dossier maps the families, decodes what "open" actually means (it is rarely what the press release implies), tells you where to run these things, and gets you from zero to a local model tonight.

---

The landscape at a glance

Read the country column. In 2023 open-weights meant Meta and a French startup. In 2026 the release velocity, and increasingly the leaderboard top, is Chinese — Qwen, DeepSeek, Kimi, GLM, MiniMax ship major open models on a cadence Western labs reserve for closed products. That is a strategic fact about the industry, not a footnote.

---

What "open" actually means

This is the section vendors hope you skip. "Open source AI" as a phrase does three different jobs, and only one of them is honest.

Open weights means you can download the trained parameters — the multi-gigabyte file of numbers — and run them on hardware you control. This is what every family above offers. It is genuinely valuable: no API dependency, no data leaving your machines, no per-token meter. But it is not open source in the software sense. You get the compiled artifact, not the build. You cannot inspect what it was trained on, and you could not reproduce it.

Open license is the separate question of what you may legally do with those weights. The spread is wide:

• MIT / Apache 2.0 (DeepSeek, most Qwen, gpt-oss, Mistral's open line, Gemma 4 ⚑ unverified): do anything — commercial use, fine-tuning, redistribution, building competing products. Apache adds a patent grant. These are real open licenses, the same ones underneath half the internet.
• Llama Community License (all Llama models): free for almost everyone, but with teeth. Companies with more than 700 million monthly active users on the release date need a separate Meta license ⚑ unverified — a clause aimed at maybe a dozen firms, but a clause nonetheless. Derivative models must carry "Llama" in the name and display "Built with Llama" ⚑ unverified. The Open Source Initiative has said plainly that this is not an open-source license, and they are right. For a normal company it functions like one; for lawyers it does not.
• Research-only and bespoke terms: still common on cutting-edge releases (some MiniMax and video/multimodal models ship this way ⚑ unverified). Read the license file. Every time. It takes four minutes and it is the cheapest legal review you will ever do.

Open data and training code — the full recipe — almost nobody offers. AI2's OLMo line and a few academic efforts are the honorable exceptions ⚑ unverified. Everyone else, including the MIT-licensed Chinese labs, publishes weights and papers but not the corpus. The industry standard is transparency about everything except the one thing that would let you check it.

The practical takeaway: for choosing a model, "open" collapses to two questions. Can I run it where I need to run it? And does the license permit my use case? Apache/MIT models answer both with yes. Llama answers yes with an asterisk. Everything else, read the file.

---

The families, in depth

Llama — the incumbent that made the market

Meta created this ecosystem. The original LLaMA leaked in March 2023, Llama 2 arrived commercially licensed that July, and Llama 3.1 405B in mid-2024 was the first open model to look a closed frontier flagship in the eye. Every tool in this dossier — llama.cpp is named for it — exists because Meta decided open weights were its wedge against OpenAI and Google.

Then came Llama 4, April 2025: a MoE family — Scout, with an advertised 10-million-token context window ⚑ unverified, and Maverick — that landed to the worst reception of any major release in the ecosystem's history. A specially tuned variant topped a chatbot leaderboard while the released weights performed worse; the long-context claims wilted under independent testing; and the "Behemoth" flagship was delayed and, per multiple reports, never shipped in its promised form ⚑ unverified. Meta reorganized its AI division around Superintelligence Labs in mid-2025, poached aggressively, and signaled that its next frontier work may not be open at all ⚑ unverified.

Where that leaves Llama in 2026: an enormous installed base, first-class support in every tool ever written, and a fine set of workhorse models (Llama 3.3 70B remains a deployment staple), but no longer the quality leader at any size point. You choose Llama today for its ecosystem inertia and enterprise comfort — the name your compliance team already approved — not because it wins evaluations. And you remember the license is Community, not Apache.

Qwen — the ladder

Alibaba's Qwen is the most complete product line in open AI, closed or open. From sub-billion-parameter edge models to the Qwen3-235B-A22B MoE flagship ⚑ unverified, through dedicated coder models (Qwen2.5-Coder and successors — the de facto standard for local coding assistants), vision-language models, embedding models, and audio models — nearly all Apache 2.0. Qwen 3.5 landed in early 2026 continuing the line ⚑ unverified.

Qwen's superpower is that whatever your VRAM budget and task, there is a Qwen that fits, and it is usually at or near the top of its weight class. The numbers back the breadth: Qwen is the most-derived-from family on Hugging Face, with more fine-tunes and merges built on it than on Llama ⚑ unverified. If the open ecosystem has a default vendor in 2026, it is Alibaba, which would have been a strange sentence to type three years ago.

The hesitation some Western enterprises voice — Chinese-origin weights — deserves a precise answer rather than a vibe: weights are static numbers; they cannot phone home; served on your hardware they exfiltrate nothing. The legitimate concerns are subtler — training-data opacity (true of every lab) and alignment choices on politically sensitive topics, which are documented and mostly irrelevant to code and business workloads. Evaluate on your task and decide like an engineer.

DeepSeek — the cost disruption

DeepSeek is the reason your API bills fell. The story has three beats. V3, December 2024: a 671B-parameter MoE with ~37B active per token, trained — per the paper — for about $5.6M in compute ⚑ unverified, a number that detonated the assumption that frontier training required hundreds of millions. R1, January 2025: reasoning via large-scale reinforcement learning, openly published, MIT-licensed, priced at a few percent of OpenAI's o1 ⚑ unverified — the release that cratered Nvidia's stock for a day and forced every closed lab to cut prices or justify them. V4, spring 2026: the current MoE flagship, still MIT, with aggressive context caching that drops cache-hit input pricing to fractions of a cent per million tokens ⚑ unverified.

DeepSeek's structural contribution outlasts any single model: it proved efficient training methods work at the frontier and published enough detail that everyone copied them, and its distilled small models (R1's reasoning distilled into Qwen and Llama checkpoints) put credible reasoning on consumer laptops. Its API is the cheapest frontier-adjacent inference money can buy ⚑ unverified; its weights are the most permissively licensed at the top of the leaderboard. The trade-offs: the hosted API is Beijing-jurisdiction (self-host or use a Western host like Together or Fireworks if that matters), and the full-size models are far too large for local hardware — DeepSeek is an open model you mostly run through someone else's GPUs.

Mistral — the split strategy

The French lab that proved a startup could play. Mistral 7B (2023) and Mixtral 8x7B (the MoE that mainstreamed the architecture in open weights) built its reputation; the business model since has been a split: open Apache-2.0 weights at the small-and-medium tier — Mistral Small 3.x, the Devstral coding line, Magistral Small for reasoning ⚑ unverified — with the Medium and Large flagships closed and API-only. Its enterprise pitch is European sovereignty: EU jurisdiction, on-prem deployment deals, and distance from both US and Chinese supply chains, a pitch that has landed real government and defense contracts ⚑ unverified.

For a builder, Mistral's open models are dependable mid-weights with clean licenses and particularly strong European-language performance. The honest note: at equal parameter count, the current Qwen and DeepSeek releases usually edge them on benchmarks ⚑ unverified, and Mistral's best work is behind its API. You choose open Mistral for the license, the latency-friendly sizes, the multilingual strength, or the jurisdiction — all legitimate reasons.

gpt-oss — OpenAI comes back

In August 2025, OpenAI released open weights for the first time since GPT-2: gpt-oss-120b (117B total, ~5B active) and gpt-oss-20b (21B total, ~3.6B active), Apache 2.0 ⚑ unverified. The sizing is surgical — the 120b runs on a single 80GB GPU, the 20b in 16GB of memory ⚑ unverified — and both are MoE reasoning models with adjustable effort and real tool-use training. Performance lands roughly at o3-mini/o4-mini level ⚑ unverified: not the frontier, but respectable, and the 20b is one of the best things you can run on a laptop.

Read it as strategy and it is more interesting: a hedge against DeepSeek and Qwen owning the open tier entirely, and a funnel — learn the OpenAI way of building (its harmony prompt format, its tool conventions), then graduate to the paid API. It works in both directions, though; gpt-oss also taught a lot of OpenAI-native developers that local inference is good enough now.

The new wave — and the pace

The most important 2026 development is not any single model but the cadence. GLM-5 from Z.ai posts the strongest open coding-agent results (high-70s on SWE-bench Verified ⚑ unverified); Moonshot's Kimi K2 line runs a trillion total parameters with 32B active ⚑ unverified under a modified MIT license and specializes in long-horizon agentic work; MiniMax M3 (June 2026) claims frontier coding, a 1M context, and native multimodality in one open release ⚑ unverified. Whatever is top of the open leaderboard when you read this will be something newer. The families above are the stable reference points; check a current leaderboard (LMArena, Artificial Analysis, SWE-bench) the week you choose.

---

Where you run them

Open weights are inert until something serves them. Four venues, in ascending order of commitment.

Your own machine — Ollama, llama.cpp, LM Studio

Ollama is the front door: one install, ollama pull, ollama run, and an OpenAI-compatible local API on port 11434 that nearly every AI tool can point at. It wraps llama.cpp, manages model files, and reduced local inference from a weekend project to a five-minute one. llama.cpp is the engine underneath — the C++ inference runtime that made CPU-and-consumer-GPU inference viable and defined the GGUF file format; go direct to it when you need maximum control or minimum footprint. LM Studio is the graphical option — model browser, chat UI, local server — the right answer for anyone who does not live in a terminal.

What local costs you: electricity, VRAM anxiety, and speed. A 32B model on strong consumer hardware generates roughly 20–30 tokens per second ⚑ unverified — comfortable for chat, sluggish for agentic loops that chain dozens of calls. What it buys you: total privacy (nothing leaves the machine — the reason law firms, clinics, and anyone handling privileged material should be running local), zero marginal cost, offline operation, and models that never get deprecated, "safety-updated," or repriced out from under you.

Cloud inference — someone else's GPUs, per token

The middle path: open models, hosted APIs, closed-model convenience at open-model prices.

• Together AI — the broadest catalog of open models behind one OpenAI-compatible API, plus fine-tuning and dedicated endpoints. The default first stop.
• Fireworks AI — speed-focused serving of the major open models, strong on function calling and structured output; popular under production agents.
• Replicate — pay-per-second model zoo, strongest for images/video/audio alongside language models; the fastest way to demo anything with weights.
• Groq — custom LPU silicon serving open models at hundreds of tokens per second ⚑ unverified, several times GPU-serving speed. When latency is the product, Groq is the answer.

Per-token prices for a 70B-class open model on these platforms run well under a dollar per million ⚑ unverified — typically 5–20x cheaper than closed flagships. This venue is also the sanctioned answer to jurisdiction worries: DeepSeek or Kimi weights served from US or EU datacenters under a Western provider's terms.

Your own GPUs — rented or owned

When volume is high and steady, per-token pricing loses to per-hour. GPU clouds — RunPod, Lambda, Vast.ai, plus the hyperscalers — rent you an A100/H100 or a consumer 4090 by the hour; you run vLLM or SGLang (the production serving engines — continuous batching, paged attention, many concurrent streams) and the economics flip: a single rented H100 at a few dollars an hour ⚑ unverified serving a 70B model to hundreds of concurrent users beats any per-token price at sustained load. Owned hardware pushes further: a workstation with 64–128GB of unified or GPU memory amortizes to near-zero per token and runs 70B-class models entirely in-house. This is the tier where "AI cost" becomes a capex line you control instead of an opex meter you fear.

Quantization — the trick that makes all of this fit

Raw weights are 16-bit: two bytes per parameter, so a 70B model is ~140GB — datacenter territory. Quantization rounds those weights to fewer bits, and the surprise of the past three years is how little quality it costs: 4-bit quantization (the ecosystem default, usually the GGUF format's Q4_K_M variant) cuts memory ~75% for a quality loss most users cannot detect in normal use ⚑ unverified. Below 4-bit, degradation gets noticeable; Q4 is the sweet spot, Q5/Q6 the cautious upgrade, Q8 near-lossless.

What fits where, at Q4_K_M, weights only — add roughly 1–4GB+ for context, more for long windows ⚑ unverified:

MoE models complicate the table pleasantly: a mixture-of-experts model must fit in memory at its total size but computes only its active parameters, so a 120B MoE runs at the speed of a ~5B model. That is why the 2026 open flagships are almost all MoE — it is the architecture that makes big cheap.

---

The honest trade

Strip away the ideology on both sides and the decision is four factors.

Capability. Closed frontier models are still better — at the hardest reasoning, the longest agentic chains, the messiest instructions. The gap in 2026 is months, not years ⚑ unverified, and for the broad middle of real workloads — summarization, extraction, classification, drafting, ordinary coding — current open models cleared the "good enough" bar some time ago. But if your product's entire edge is maximum intelligence, closed still wins, and pretending otherwise is cope.

Cost. Open wins, at every venue, usually by 5–20x per token ⚑ unverified, approaching zero marginal cost when self-hosted at scale. The counterweight is engineering time: closed vendors sell you a phone number to call when it breaks; open sells you the whole stack, including the 2 a.m. parts.

Control. Entirely open's column, and it is worth more than most teams price in. Weights on your disk cannot be deprecated, repriced, re-aligned, or rate-limited. Your prompts train nobody. Your privileged data never leaves your jurisdiction. You can fine-tune to your domain — closed vendors' fine-tuning offerings are narrow and non-portable. Every closed-API product carries platform risk that reprices without notice; open weights are the only full hedge.

Accountability. The murkiest column. Closed vendors give you terms of service, indemnities, and someone to sue. Open weights arrive as-is: training data unknown, biases yours to discover, compliance yours to certify. For regulated industries this cuts both ways — self-hosting satisfies data-residency rules no API vendor can, while model provenance becomes your audit problem.

Who should go open: anyone with privileged or regulated data that cannot leave the building; anyone whose per-token volume makes API pricing a P&L problem; anyone whose product depends on a model behaving identically in five years; anyone who needs deep fine-tuning; and every developer who wants to understand this technology rather than rent it. Who should not: teams whose one differentiator is frontier intelligence, teams with no ops capacity and no volume, and anyone for whom a $50/month API bill is not worth a weekend of setup.

The pattern winning in practice is not either/or — it is closed frontier for the 10% of calls that need it, open for the 90% that don't, behind a router. That architecture is the actual endgame, and it only exists because open weights forced the market to allow it.

---

How to start — tonight

Thirty minutes, no GPU required (8GB+ of RAM helps; Apple Silicon is ideal).

1. Install Ollama. ollama.com → download for macOS/Windows/Linux, or curl -fsSL https://ollama.com/install.sh | sh. One binary, no dependencies.
2. Pull a model matched to your machine. 8GB RAM: ollama pull qwen3:4b ⚑ unverified. 16GB: ollama pull gpt-oss:20b or qwen3:8b. 32GB+: ollama pull qwen3:32b or a 30B-class coder model. The download is a few GB; it happens once.
3. Run it. ollama run gpt-oss:20b and you are chatting with a reasoning model on your own silicon, offline, unmetered. Ask it something real from your work.
4. Point a tool at it. Ollama serves an OpenAI-compatible API at http://localhost:11434/v1. Drop that base URL into any AI client, coding assistant, or three lines of Python (openai SDK, api_key="ollama"), and every OpenAI-shaped tool you own now runs local.
5. When you outgrow the laptop, take the same model to a hosted API — Together, Fireworks, or Groq for speed — same weights, same prompts, someone else's electricity. When volume justifies it, rent the GPU itself — RunPod or Lambda — and serve with vLLM.

The whole ladder — laptop to hosted API to your own cluster — runs the same open weights. That continuity is the product. No closed vendor can sell it to you, which is precisely why this ecosystem exists.

---

Affiliate & sign-up surface (for production)

• Together AI — hosted open-model API, fine-tuning.
• Fireworks AI — fast open-model serving.
• Replicate — pay-per-use model zoo.
• RunPod / Lambda / Vast.ai — GPU rental by the hour.
• (Groq, Ollama, LM Studio have no paid sign-up worth tagging at draft time — confirm current programs before publish.)

Verification checklist before publication

All ⚑ unverified items: the Nvidia one-day loss figure; Llama 4 context/variant claims and license clause numbers; current Qwen/DeepSeek/GLM/Kimi/MiniMax flagship names, licenses, and benchmark figures; DeepSeek training-cost and pricing figures; gpt-oss specs; every row of the VRAM table against current GGUF file sizes on Hugging Face; hosted-inference and GPU-hour price ranges; current Ollama model tags. Sources: official model cards on Hugging Face, vendor pricing pages, ollama.com/library, and a current leaderboard (LMArena / Artificial Analysis / SWE-bench). This landscape turns over monthly; verify within a week of publish.