Z.ai Releases GLM-Legal: 19.17% Full-Pass Rate on Harvey LAB, API by Application Only with No Public Weights
According to a MarkTechPost report on October 13, 2026, Z.ai released the legal-domain model GLM-Legal, built on GLM-5.3-Flash, with a 19.17% full-pass rate on Harvey LAB—roughly triple the base model's 6.67%—and a Vals AI per-test cost of 0.15 USD; access is currently limited to an application-only API, with no published pricing or weights.
What changed
According to a MarkTechPost report dated October 13, 2026, Z.ai released GLM-Legal, a domain-specialized model for legal scenarios, targeting contract review, evidence analysis and compliance research. The model is built on GLM-5.3-Flash, which the source describes as a 320B total-parameter, 18B activated-parameter MoE with a 1 million token context and 128K maximum output, mixing sparse and linear attention, and able to read text, images, video and files. Training has two stages: mid-training on an expanded legal corpus, followed by reinforcement-learning post-training across tens of thousands of legal task environments; the source also mentions that the execution framework and the model co-evolved during later training. On evaluation, Harvey LAB is an open benchmark containing 1,250 tasks covering 24 practice areas, scored by more than 75,000 expert-written rubrics, with a task counted as passed only if all its rubrics pass; Vals AI ran the holdout set with two LLM judges under internet-disabled conditions. The report says GLM-Legal achieved a 19.17% full-pass rate, ranking 7th among 79 models, versus 6.67% for the base model, and a 92.5% per-rubric pass rate. The same table shows Kimi K3 at 12.92%, Claude Fable 5 at 11.25% and GPT-6 Astra at 5.42%, but the release chart omits the six higher-ranked models, with the top entry at 25.42%. On LegalBench, GLM-Legal scored 86.23%, behind Claude Fable 5's 88.56%; on Legal Research Bench it scored 41.83%, below the base model's 45.19%. Access is only via Z.ai's hosted API with an application required; as of the report date no pricing had been published and no corresponding repository existed on Hugging Face; the base model can be self-deployed via vLLM or SGLang, with open FP8 weights.
What it means for developers
For legal-tech and in-house legal teams, the value of this news is not the ranking itself but that a vertical model roughly tripled the base model's pass rate on agentic legal tasks at a relatively low per-test cost, suggesting that domain post-training plus RL may yield noticeably larger gains on long, tool-using workflows than general capability improvements would. But the deployment judgment needs to be unpacked. First, GLM-Legal currently has no public weights and is only available through an application-gated hosted API, which effectively hands part of data flow, audit and compliance responsibility back to the vendor—a precondition, not a side issue, for law firms handling client-confidential material. The open weights mentioned in the source refer to the FP8 weights of the base GLM-5.3-Flash, not to GLM-Legal itself being open source; weight availability and license rights are two different things, and the base model's openness does not imply the derivative can be self-deployed. Second, the cost advantage under the source's framing needs careful reading: the 0.15 USD figure is Vals AI's per-test cost computed under specific task lengths using 0.075 USD per million input tokens and 0.25 USD per million output tokens—it is not a vendor-published API price, and GLM-Legal's pricing has not been disclosed, so any cost model based on that number is an assumption. Third, capability gains are uneven: Harvey LAB improved by 12.5 percentage points, but Legal Research Bench was 3.36 percentage points below the base model, indicating the model leans toward multi-step tool-invocation tasks and may yield negative gains on open-ended legal retrieval and research. Fourth, each Harvey LAB task averaged 1 hour 12 minutes, slower than most models in the same table, so the trade-off between interactive and batch-offline scenarios would be entirely different. In addition, the source says Z.ai stated GLM-Legal will come to its desktop agent application AutoClaw, but no timeline was given; these are all reporting claims, and KimiSeek has not independently verified them.
Integration and evaluation
The following are advisory judgments, not verified facts. Decision one: choose the base model for prototyping or wait for GLM-Legal access. The report says the base GLM-5.3-Flash is available with open weights, while GLM-Legal is application-only, so if a project must start immediately or must process sensitive documents locally, building the pipeline on the base model is more realistic; if the tasks center on multi-step contract comparison and evidence-chain assembly that need tool calls, and data can be vendor-hosted, applying is worth it. Preconditions are whether access is approved and whether real client material may be processed. The evaluation criterion could be: on 50 to 100 of your own historical contract-review samples, compare the base model and, once approved, GLM-Legal on first-pass rate under the full-pass rubric, and require the latter to show an explainable improvement over the base; otherwise do not migrate. Decision two: factor latency into model selection rather than looking only at pass rate. The source shows GLM-Legal averaging 1 hour 12 minutes per task, clearly longer than other models in the same table, so a lawyer-facing real-time assistant and a nightly batch due-diligence pipeline should make different choices; if the former cannot accept that latency, keeping the base model or a faster general model in the front end and routing the high-accuracy model to an asynchronous queue is a conditional option. The precondition is that the task can be made asynchronous and users accept hour-level turnaround, and the evaluation criterion could be measuring P95 end-to-end latency and per-task cost under realistic concurrency against budget. Decision three: do not finalize on a single leaderboard. The report gives two opposing results—a large Harvey LAB gain and a Legal Research Bench score below the base model—so before adoption, split your evaluation set by business distribution, tally pass rates separately for retrieval-and-research tasks and process-execution tasks, and record per-rubric pass rates rather than only task pass rates. The evaluation criterion could be: the direction of the delta versus the base model on the two task types must match their weight in your business; if retrieval and research dominate, do not switch solely because of the Harvey LAB number. Finally, since pricing is undisclosed, any cost accounting should first confirm billing terms with the vendor and recalculate using your own token distribution, rather than carrying over a third-party per-test cost.
Limits and open questions
All benchmark scores, costs and durations in this report come from the Vals AI run described by the source as of October 12, 2026, and KimiSeek has not independently reproduced them. This version of GLM-Legal is not currently in the KimiSeek model catalog; the API availability, application method and pricing status mentioned refer to the vendor Z.ai, not this site; Z.ai has not published GLM-Legal's API price, so the per-test cost in the source cannot be treated as an actual billing standard. The base GLM-5.3-Flash has a same-named glm-5.3-flash entry in the KimiSeek catalog, but that is a different model from GLM-Legal, and it cannot be used to infer that GLM-Legal is available or self-deployable. The source did not disclose GLM-Legal's independent parameter count, nor provide specific figures for the Chinese PLawBench, so Z.ai's claim that it is on par on LegalBench and PLawBench lacks checkable numbers. The names of the higher-ranked models on the Harvey leaderboard were omitted from the release chart, and this report merely relays the original text. Whether API access will be granted, the AutoClaw launch timeline and the actual license terms are all undetermined and should be subject to the vendor's later statements.