The open-weight tier right now: Llama 4, Mistral, Qwen, DeepSeek

"Open-weight has caught up." When this piece first ran in January, that take was premature. Five months later it's much closer to true, because the spring of 2026 handed the open tier a generational refresh: DeepSeek shipped V4 in April, Alibaba shipped Qwen3.6 the same month, and Mistral had already moved Large 3 to a clean Apache 2.0 license back in December. The honest version of the claim still needs the same split, though. Open weights have closed the gap on conversational use, on code generation, and on multilingual reasoning. They remain behind on long-context retrieval at extreme scale, and the agent-loop gap — while smaller than it was — hasn't fully closed. The four families covered here show the unevenness in different ways.

The lineup: Llama 4 in its two shipped configurations (the 400B-parameter mixture-of-experts Maverick with about 17B active, and the 109B Scout with its 10M-token context), Mistral Large 3 at 675B total / 41B active, Qwen3.6 in the 27B dense and 35B-A3B variants, and DeepSeek-V4 in Pro and Flash — same weights, two hosted sizes, with Pro at 1.6T total parameters and 49B active per token. Every hard number below comes from the provider's own model card or pricing page. If you want to run any of these on your own hardware, see running models on your own machine, and the open-weight guide maps the whole tier.

One caveat before the lineup. The headline benchmark jumps on DeepSeek-V4 vs V3.1 are real and published on the official model card. Whether that translates into the same delta on your specific workload (finance modeling, hard reasoning under uncertainty) is the kind of thing only your own test can tell you. The leaderboard gap tends to look larger than what shows up at the workload level.

Llama 4

Meta released Llama 4 in April 2025, per Meta AI's Llama 4 announcement, in two main configurations. Maverick is the 400B-parameter mixture-of-experts model (about 17B active per token) intended for serious GPU deployment. Scout is the 109B mixture-of-experts variant that activates roughly 17B parameters per forward pass and runs on a single 80GB H100 with sensible quantization. Weights and downloads are available at llama.com.

More than a year after launch, Llama 4's claim on the tier has narrowed to one thing nobody else offers: Scout's 10M-token context window on hardware you control. Maverick is still a competent generalist, but the spring 2026 releases from DeepSeek and Qwen pushed past it on coding and reasoning — Maverick's official card shows 69.8% on GPQA Diamond, well under Qwen3.6-27B's 87.8% and DeepSeek-V4-Pro's 90.1%. Scout is the workhorse, trading peak capability for a single 80GB H100 with sensible quantization and that giant window.

The license is the Llama 4 Community License: permissive for almost everyone, with a clause forbidding use by services with more than 700M monthly active users. That clause is irrelevant to a small team or a solo developer. At a large company, read the license carefully against the specific deployment context.

Mistral Large 3

Released December 2, 2025 at 675B total parameters with 41B active, per Mistral's announcement and the official model page. The headline change from Large 2 isn't capability — it's the license. Large 3 ships under Apache 2.0, dropping the old Research License and its separate commercial terms entirely. You can deploy it commercially today without talking to anyone.

This is still the open-weights lab with the strongest house style: a clean preference for structured output and a willingness to commit to opinions instead of hedging endlessly. Its European-language work remains clearly stronger than the alternatives, and the hosted API is cheap at $0.50 / $1.50 per million tokens. The context window is 256K. Where it trails the new DeepSeek and Qwen releases is raw coding and math throughput — pick it for language quality and license cleanliness, not leaderboard position.

Qwen3.6

Alibaba's current series, released April 2026, with model cards on Hugging Face under the Qwen organization. The two confirmed open-weight variants are the 27B dense model and the 35B-A3B mixture-of-experts that activates just 3B parameters per token. The 27B is the one to take seriously: its official card lists 77.2% on SWE-bench Verified and 87.8% on GPQA Diamond — numbers that would have been frontier-class a year ago, from a model that fits on one GPU. Qwen remains the strongest open family on Chinese-language work and one of the better ones on Arabic and Japanese.

Both variants ship under Apache 2.0, the cleanest license in the lineup. The instruction-following drift that dogged earlier Qwen generations in long conversations is improved but not gone, and it's still the thing to test first if your workload is agentic. On single-shot or short-conversation use, the quality per parameter is the best in the open tier.

DeepSeek-V4

Released April 24, 2026, with the announcement and pricing at api-docs.deepseek.com. V4 comes in two hosted sizes built on the same open MIT-licensed weights: Flash at $0.14 / $0.28 per million tokens, and Pro — 1.6T total parameters, 49B active — at $0.435 / $0.87. Both carry a 1M-token context window and an unusually large 384K max output. DeepSeek has kept the most aggressive open-weights story of any current lab: detailed technical reports, model cards that publish the numbers instead of marketing language, and hosted pricing far below the Western alternatives.

The official card puts V4-Pro at 80.6% on SWE-bench Verified, 90.1% on GPQA Diamond, and 93.5% on LiveCodeBench — the strongest published set in the open tier, and within a few points of closed flagships that cost ten times more per token. The weaknesses haven't moved: tool use inside long agent loops is less polished than the closed alternatives, and the safety tuning is lighter than what Western users may expect from a frontier model.

The license is MIT for the weights. Permissive and clean; read the hosted-API terms separately if your deployment touches anything sensitive.

Three places where open has caught up

Three categories where the open-weight tier is close enough to closed models that capability shouldn't decide it. License, cost, and deployment preferences should.

General knowledge and conversational reasoning at typical lengths. The top open-weight models are within striking distance of the closed frontier on chat-style use, factual questions, and structured reasoning that fits in a single context window. The leaderboards capture this accurately, even if they miss the categories further down. For more on the leaderboard problem, see why benchmarks stopped telling you anything.

Code generation is the second, and it's no longer limited to isolated tasks. DeepSeek-V4-Pro's 80.6% on SWE-bench Verified is a repository-level score — real multi-file fixes, not toy functions — and it sits within striking distance of closed flagships. The gap that remains shows up at architectural scale, on the design decisions that span a production codebase. For most day-to-day coding, the open models are simply good enough now.

The third is multilingual capability in high-resource languages. The top open models compete strongly across European languages, Chinese, Japanese, and increasingly Arabic, and Qwen3.6 specifically pushes the Chinese frontier ahead of any closed model you can buy. For organizations doing serious multilingual work, the open-weight tier has become a genuine first choice rather than a fallback.

The capabilities open weights are slowest to match are exactly the ones the closed labs poured the most engineering into. The hardest gaps to close are the ones worth the most money.

Two places where closed still wins

Two categories where the open-weight tier is clearly behind the closed alternatives. For serious production deployments here, stick with closed.

The first is long-context retrieval at extreme scale. The closed models (Claude Opus 4.8, GPT-5.5, Gemini 3.5 Flash, Gemini 3.1 Pro) have put enormous engineering effort into making their million-token contexts usable: recall stays high, hallucinations stay low, and the model will quote rather than summarize when asked. Open-weight models with similar nominal windows still show drops past the 500K-token mark — though DeepSeek publishes its own long-context numbers now (83.5% on MRCR at 1M for V4-Pro), which is more transparency than this category had a year ago.

The second is reliable tool use and agent behavior, and this is the gap that moved most since January. DeepSeek tuned V4 specifically for long-running agentic work, and it shows in the published Terminal-Bench scores. But the closed labs spent that same period tuning too, and in production the open models still need more scaffolding to stay on task and recover from tool failures without getting stuck. The gap is now a lead, not a chasm. For the highest-stakes multi-step workflows, closed still wins; for everything else, test before you assume it does.

1. Feb 2024 Mistral Large

   First serious open-weight competitor to GPT-4.

2. Jul 2024 Llama 3.1 405B

   Meta's first frontier-class open model.

3. Dec 2024 DeepSeek-V3

   Open MoE that closed the cost gap.

4. Apr 2025 Llama 4 Maverick / Scout

   Meta's MoE generation; Scout brings the 10M-token context.

5. Dec 2025 Mistral Large 3

   675B MoE — and the move to a clean Apache 2.0 license.

6. Apr 2026 Qwen3.6 + DeepSeek-V4

   The generational refresh: 77-81% SWE-bench Verified in the open tier.

The comparison table

Open-weight frontier models, benchr survey, June 2026

Model	Parameters	License	Best at	Avoid for
DeepSeek-V4-Pro	1.6T (49B active)	MIT	Code, math, top open tier	Safety-critical applications
DeepSeek-V4-Flash	same weights, smaller serve	MIT	Cheapest capable API ($0.14 in)	Peak accuracy — use Pro
Qwen3.6-27B	27B dense	Apache 2.0	Single-GPU quality, multilingual	Long agentic conversations
Mistral Large 3	675B (41B active)	Apache 2.0	European languages, structure	Leaderboard-level code and math
Llama 4 Scout	109B (17B active)	Llama 4 Community	10M-token context on one H100	Anything needing top accuracy
Llama 4 Maverick	400B (17B active)	Llama 4 Community	General use in Meta's stack	Code and reasoning — V4 and Qwen3.6 lead

Granite (IBM's openly-licensed line) and Phi (Microsoft's small-model family) aren't in this survey. Granite is solid for enterprise text work but doesn't compete at the frontier. Phi gets its own piece in the small-model review.

The decision rule

If you're building anything that has to run inside a regulated environment with no data leaving your network, open weights are effectively the only option on the table. Whatever capability gap exists is worth absorbing to avoid the compliance problem of sending data to a closed API.

If the unit economics of your workload are dominated by per-token cost (high-volume inference, batch document processing, anything serving thousands of requests per minute), DeepSeek-V4-Flash at $0.14 per million input tokens beats the closed alternatives by an order of magnitude on dollars per query. Run the math yourself in the cost calculator.

If your workload depends on the model maintaining coherence across hundreds of thousands of tokens of retrieval, or on flawless multi-step tool use where a stuck loop costs real money, stay on closed. The gap shrank this spring, but it's still there, and it's exactly where production failures hurt most.

When there's no strong prior either way, prototype on a closed model for development speed, then re-test the production path on DeepSeek-V4 or Qwen3.6-27B before scaling. Often the open model will work fine and save you money that adds up over time. Often enough, you'll hit a specific failure mode that justifies the closed-model premium. Which way it breaks depends on the use case more than on any rule of thumb.

Open-weight models in mid-2026 are good enough to be the right answer for most workloads. That wasn't true when this piece first ran in January, and the difference is the April releases: repository-level coding scores in the high 70s and low 80s, published long-context numbers, and two of the four families now on Apache 2.0 or MIT. The bread-and-butter work — conversation, code, multilingual reasoning — is settled in open's favor on cost grounds alone.

The categories where closed still leads happen to be where most production money goes, and that overlap is no coincidence. The closed labs have prioritized the workflows that generate the highest-value revenue, and the open-weights labs have followed at a shrinking distance. DeepSeek's V4 closed more of that distance in one release than the rest of the tier managed in a year.

If you have to pick one open-weight model today, the default is DeepSeek-V4-Pro on a hosted endpoint, or Qwen3.6-27B if the weights have to run on your own hardware. Others win on specifics: Mistral Large 3 on European languages and clean structure, Llama 4 Scout when nothing else fits the context requirement. Let the workload pick, not the brand.