Qwen
The best licence-to-capability ratio in open weights: Apache-licensed for most sizes, strong instruction following, and a small-model range that makes a genuinely private pilot affordable for a mid-sized firm.
Tier B — conditional, and the model we would put in front of a firm that wants a private deployment with the fewest licence complications. Its weaknesses are behavioural rather than structural: it reasons past silence and it is terse. Both are manageable with prompt discipline, both should be tested on your own matters before rollout.
Specifications, as published
| Publisher | Alibaba Cloud |
| Family | Qwen3 generations, including small dense models and large mixture-of-experts releases |
| Parameters | Spans small dense models suited to a single workstation through to large mixture-of-experts releases |
| Context | Long context across the range; small models hold up better than their size suggests on retrieval-grounded tasks |
| Licence | Apache 2.0 for the majority of releases — the most straightforward commercial licence in this index |
| Weights | Downloadable, with strong quantised community support |
| Release | Continuous release cadence; version pinning is essential |
| Licence posture | Apache 2.0 for the majority of releases — the most straightforward commercial licence in this index |
Specifications are as published by the publisher at the review date and change frequently. Confirm them in your own evaluation before you procure.
Releases and variants
| Release | Size | Context | Serving footprint | What it is for |
|---|---|---|---|---|
| Qwen3.5-0.8B | 0.8B dense, vision-language | 262,144 tokens natively | ~0.5GB at 4-bit quantisation; ~1GB at 8-bit | The publisher describes this size as being for prototyping and fine-tuning, and its documented tendency to enter thinking loops means it belongs in experiments and classification pipelines rather than anywhere near client material. |
| Qwen3.5-2B | 2B dense, vision-language | 262,144 tokens natively | ~1.5GB at 4-bit quantisation; ~2.5GB at 8-bit | Small enough to sit inside a workstation's spare memory for tidy extraction and translation work, but a firm of 10–50 fee-earners should use it as one component in a pipeline rather than as the model answering a fee-earner's question. |
| Qwen3.5-9B | 9B dense, vision-language | 262,144 tokens natively, extensible to 1,010,000 tokens | ~6GB at 4-bit quantisation; ~10GB at 8-bit | The smallest release in this family we would point at real documents: one 24GB card gives a small firm credible extraction and summarisation, subject to the judgement ceiling that a 9B model implies. |
| Qwen3.6-35B-A3B | 35B total / 3B active, 256 experts | 262,144 tokens natively, extensible to 1,010,000 tokens | ~18GB at 4-bit quantisation; ~36GB at 8-bit | The throughput play in this family — only 3B parameters active per token, so a single 24GB card keeps a busy team served on extraction and triage where a dense 27B would queue. |
| Qwen3.8-27B | 27B dense, vision-language, thinking mode on by default | 262,144 tokens natively, extensible to 1,000,000 tokens | ~15GB at 4-bit quantisation; ~28GB at 8-bit | Our recommended default: a firm of 10–50 fee-earners can serve it at 8-bit on one 48GB card, and it combines current-generation capability with the most straightforward Apache 2.0 terms in this index. |
| Qwen3.5-122B-A10B | 122B total / 10B active, 256 experts | 262,144 tokens natively, extensible to 1,010,000 tokens | ~63GB at 4-bit quantisation; ~125GB at 8-bit; ~250GB for the published BF16 weights | A realistic step up for a firm that has outgrown the 27B release — an 80GB card runs it at 4-bit and the 10B active parameters keep latency civil — but the quality gain has to be proved on your own matters before you buy the card. |
| Qwen3.8-Flash-Next | 125B total / 6B active, plus 51B n-gram embedding and 4B multi-token-prediction module | 262,144 tokens natively, extensible to 1,000,000 tokens | ~65GB at 4-bit quantisation for the MoE weights, plus ~50GB of n-gram embedding tables that the publisher notes can be offloaded to host memory | The low-latency, long-context release built for high-volume serving on an 80GB card or more, and the one to read the licence on: Qwen Community Licence 1.0 permits a firm's internal use but not offering the model's capabilities to third parties. |
| Qwen3.8-2.4T-A95B | 2.4T total / 95B active, 512 experts (10 routed plus 1 shared) | 262,144 tokens natively, extensible to 1,010,000 tokens | ~1.2TB at 4-bit quantisation; ~4.9TB for the published BF16 weights | Capable and impractical for a 20-partner firm: 2.4T parameters is a data-centre deployment under a bespoke Qwen3.8-Max licence, and a firm of this size should rent capacity or use a hosted endpoint rather than plan to serve it. |
Sizes, context windows and licences are as published by the publisher at the review date. The variant you pick matters more than the family name: a small dense release that fits one workstation and a large mixture-of-experts release that needs a multi-GPU server are not the same product, whatever the marketing says.
How it behaves on legal work
For legal work Qwen's most useful trait is instruction obedience. Give it a schema, a tone, a length limit and a refusal rule, and it will follow all four — which sounds unremarkable until you have watched other models quietly ignore the fourth. Its extraction behaviour is the strongest of the models we have assessed: asked to pull dates, parties, obligations and termination rights out of a contract set into a fixed structure, it produces clean output that a paralegal can validate rather than repair. Drafting is competent and slightly terse; letters come out shorter and less padded than Llama's, which many fee-earners prefer and some partners find abrupt. On research-style questions it is fast and structured but has a visible weakness: it will happily reason from its own general knowledge when retrieval returns nothing useful, and its confident phrasing in those moments is the risk. Handling of Chinese-language material is an obvious strength, which matters for firms with cross-border matters and is a genuine differentiator against most of this index. The multi-step reasoning mode is worth using for triage — allocation of documents to issues, or deciding what to read next — rather than for final legal conclusions.
Abstention is good when a refusal rule is explicit and the retrieved context is empty. It is average when the context contains adjacent-but-not-answering material: it will bridge the gap rather than flag it. Quote-with-reference instructions work well; ask for paragraph-level citation and it complies.
What we would use it for
- Contract extraction into a fixed schema at volume
- Disclosure triage and document allocation to issues
- Multilingual matter support, including Chinese-language documents
- Drafting from precedent where brevity is wanted
- A genuinely affordable first private pilot on a single GPU
What to watch
- Reasoning past the retrieved evidence when the corpus is silent
- Version churn — pin the exact release in your evaluation record
- Provenance questions from clients: be ready to explain the publisher's jurisdiction and terms
- Terse drafting that can read as blunt in client-facing correspondence
What it costs to run
We would serve Qwen3.8-27B at 8-bit on a single 48GB card: that preserves the long context the family advertises, leaves room for the KV cache of several concurrent requests, and keeps the deployment on plain Apache 2.0 terms. For a firm of 10–50 fee-earners that is comfortable interactive latency, because schema-bound extraction is largely input-bound and parallelises well, while long drafting runs come back nearer reading speed. Where throughput rather than quality is the constraint, Qwen3.6-35B-A3B at 4-bit runs on a 24GB card with 3B parameters active per token; the 122B and 2.4T releases are separate purchases with separate hardware consequences.
| Basis | GPU hours | Hourly (USD) | Monthly (USD) | When this is the right pattern |
|---|---|---|---|---|
| Always-on server (24/7) | 730 h | $0.85 – $2.85 | $625 – $2,080 | Firm-wide access, no cold starts, predictable latency |
| Business hours (10 h × 21 days) | 210 h | $0.85 – $2.85 | $180 – $600 | The realistic pattern for a firm of 10–50 fee-earners: power it up, use it, stop paying |
| Bursty / autoscaled endpoints | 60 h | $1.80 – $3.60 | $110 – $215 | Occasional analysis and pilots; you pay only for the seconds the model is working |
| Storage — weights, index and evaluation sets (~200 GB) | — | — | $30 | Billed whether the model is running or not — the quiet line on the invoice |
Indicative GPU class: 48GB class — L40S / RTX A6000. Every figure above includes a 50% buffer on the underlying cloud rates — for encrypted storage, egress, idle capacity between requests, cold starts, operational overhead, and the plain fact that these are estimates rather than quotes. Rates move weekly and vary by region, tier and commitment.
Qwen is the family where buying hardware is most defensible, because the 27B release we would serve fits a single 48GB-class workstation — indicatively $8,000–15,000 for the class — and the Apache 2.0 terms let you leave it in place for years without a licence conversation. Renting the same class remains the right first step if your usage amounts to a few hours a week, since idle capital is the largest avoidable cost in private AI. Those bands are indicative rather than quotes and take no account of the retrieval layer and evaluation time that will dominate your real spend; the 122B release above needs an 80GB-class server, which is a materially larger commitment.
Two rules of thumb that hold across the models we have deployed: renting beats buying until a firm is using the model more than about half of every working day, and stopping the instance matters more than the hourly rate — an idle server, and an idle storage volume attached to it, are where private AI budgets quietly go.
Sampling and prompt settings
Extraction: temperature 0, top_p 0.8, fixed schema, no creative latitude. Triage: 0.2–0.3. Drafting: 0.4. Research summaries: 0.2 with an explicit 'cite the passage' instruction. The small variants can be pushed to 0 for classification without collapsing into repetition.
Pin the exact model release in your evaluation record. Behaviour moves between point releases, and an evaluation that does not name a version cannot be reproduced.
Hardware and quantisation
| Deployment profile | What it fits | What to know |
|---|---|---|
| Single workstation (small dense, quantised) | Extraction and summarisation for one team | Cheapest credible private pilot in this index |
| One GPU server (mid-size) | Firm-wide extraction and drafting support | Good throughput per pound for schema-bound work |
| MoE release, larger deployment | Research-style tasks and heavier drafting volume | Evaluate against the mid-size model before paying for the step up |
Who it suits
Firms wanting Apache-licensed capability, cross-border and multilingual work, and high-volume extraction where an output schema matters more than elegant prose.
Firms that need warm client-facing drafting out of the box, or that cannot commit to pinning and re-evaluating model versions.
Review history
| Date | Change |
|---|---|
| Sep 2026 | First entry. |
Sources
- Qwen model releases and licences (Alibaba Cloud)
- Open source LLM landscape 2026
- Probative Co: AI readiness assessment