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

The spring of 2026 materially changed the open-weight candidate set: DeepSeek shipped V4 in April, Alibaba shipped Qwen3.6 the same month, and Mistral had already moved Large 3 to Apache 2.0 in December. Public scores show gains, but they do not establish that open weights have universally “caught up” in conversation, coding, multilingual work, long retrieval, or agents. The useful question is narrower: which candidate meets your measured quality threshold under your license, privacy, latency, and cost constraints?

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.

Mistral positions the model for multilingual and structured-output work. Its documented hosted price in this snapshot is $0.50 / $1.50 per million tokens and its context window is 256K. The cited sources do not prove that its style or European-language output is stronger than every alternative. Include it when those dimensions and Apache 2.0 matter, then compare language fidelity, structured-output validity, and task accuracy on the same held-out set.

Qwen3.6

Alibaba's current series, released April 2026, has 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 card reports 77.2% on SWE-bench Verified and 87.8% on GPQA Diamond. Those are provider-reported benchmark signals, not proof of production performance or language leadership. Teams working in Chinese, Arabic, or Japanese should use fluent reviewers and a local held-out set.

Both variants ship under Apache 2.0. Review the current license and deployment dependencies for your use case. For agentic work, test instruction retention and recovery over long conversations; for short requests, measure quality, throughput, and memory on your own hardware. The cited material does not establish a universal quality-per-parameter winner.

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 reports V4-Pro at 80.6% on SWE-bench Verified, 90.1% on GPQA Diamond, and 93.5% on LiveCodeBench. These vendor-reported benchmark numbers are useful shortlist evidence, but cross-provider comparisons may use different settings and do not establish production reliability, agent behavior, or safety. Test repository tasks, tool failures, abuse cases, and total serving cost under your own configuration.

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

Three areas worth testing with open weights

Public results make three categories credible targets for open weights. Whether the quality is close enough depends on your workload. Measure it before license, cost, or deployment preference decides the pick.

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 areas that need extra validation

Long-context retrieval and multi-step tool use are especially sensitive to setup and failure handling. Public benchmark results alone do not justify a blanket open-versus-closed rule.

The first is long-context retrieval at extreme scale. A maximum context window does not guarantee recall or low hallucination rates for either closed or open models. DeepSeek reports 83.5% on MRCR at 1M for V4-Pro, but that vendor result does not predict your corpus. Create answer-key documents at several depths, measure citation and abstention behavior, and compare candidates under equal retrieval settings.

The second is reliable tool use and agent behavior. Providers publish agent benchmarks, but production results depend on prompts, tool schemas, retry policy, permissions, and observability. Run the same multi-step tasks with injected tool failures; score completion, unsafe actions, recovery, latency, and cost. Use human approval for consequential actions regardless of model family.

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	Published signal	Validate locally
DeepSeek-V4-Pro	1.6T (49B active)	MIT	Provider coding and reasoning scores	Safety and multi-step tools
DeepSeek-V4-Flash	same weights, smaller serve	MIT	$0.14/M input snapshot	Quality and total serving cost
Qwen3.6-27B	27B dense	Apache 2.0	Compact multilingual candidate	Long agentic conversations
Mistral Large 3	675B (41B active)	Apache 2.0	Multilingual positioning and structure	Language, code, and math quality
Llama 4 Scout	109B (17B active)	Llama 4 Community	Advertised 10M-token context	Recall and task accuracy
Llama 4 Maverick	400B (17B active)	Llama 4 Community	General-purpose Meta model	Code, reasoning, and hardware

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 data cannot leave your network, self-hosted open weights are a relevant candidate. They are not automatically compliant: review the license, model provenance, security controls, logging, infrastructure, and your sector's requirements with the responsible legal and security teams.

If token price dominates your workload, the documented $0.14-per-million input rate for DeepSeek-V4-Flash makes it worth modeling. Per-token price is not total cost: include output, caching, retries, hosting, latency, review, and failure rates in the cost calculator.

If your workload depends on retrieval across hundreds of thousands of tokens or multi-step tools, require an acceptance test before deployment. Compare open and closed candidates under the same corpus, tool failures, retry limits, and human-approval policy instead of choosing by model family.

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 credible production candidates for many workloads, but “good enough” must be defined locally. April releases added provider-reported repository coding scores, long-context results, and permissive license options; none of those facts alone settles conversation, code, or multilingual production quality.

Closed and open providers optimize different product surfaces, and public tables do not quantify a universal distance between them. Track the failure modes and economics that matter to your deployment rather than inferring strategy or production value from benchmark movement.

For an initial shortlist, include DeepSeek-V4-Pro when its published coding results and hosted price fit, Qwen3.6-27B for a smaller self-hosted candidate, Mistral Large 3 for multilingual and structured-output evaluation, and Llama 4 Scout when its advertised context is relevant. Let measured workload results—not the brand or this editorial list—choose.