Skip to content
Blu SolutionsBlu Solutions
Our AI

Models we run on our own servers

We invested in the hardware to enable this, and we host the models ourselves — checkpoints most practices never run. On top of the published weights we add our own RAG. Human accountants use that for analysis and automation; a person still applies professional judgement.

  • Qwen3.8-Flash-Next — workhorse
  • n_ctx 262144 (256K native window)
  • RAG on our servers

Open weights, private retrieval

Named models, Hugging Face links

These are the published checkpoints. We serve them locally and attach our own RAG databases. See Blu OS.

On our servers

Qwen3.8-Flash-Next

Workhorse — day-to-day work on our servers

Open weights on Hugging Face. 176 billion total parameters. For general work on a Mac Studio, people recommend Qwen3.8-27B (~27B) — we run Flash-Next instead. We serve Qwen3.8-Flash-Next at n_ctx 262144 — the full 256K native context window on our servers.

The model we run locally for ordinary firm work.

Hugging Face: Qwen/Qwen3.8-Flash-Next · Qwen

Published benchmarks

Published scores, not tests we ran. Linked to the source on each card.

LiveCodeBench v6
91.9
SWE-bench Pro
62.5
SWE-bench Multilingual
81.0
DeepSWE 1.1
58.7
NL2Repo-Bench
48.1
CoWorkBench
73.9
JobBench
55.7

Qwen’s published table reports that Flash-Next exceeds Qwen3.8-27B, Qwen3.7-Plus and DeepSeek V4-Flash on several coding and agentic tests, and is ahead of Claude Opus 4.6 on SWE-bench Pro and CoWorkBench in that table. That does not prove universal superiority — it means the model is competitive with much larger and closed models on those selected evaluations. Model card.

What independent testers have said

  • DataCamp’s local-coding write-up called it a step up from smaller Qwen checkpoints for agentic coding: it plans the job before it writes. DataCamp
Adding for slow reviews

GLM-5.3-Flash

Slow reviews — careful second pass

Open weights on Hugging Face. 320 billion total parameters — this needs specialised hardware; it is not a model you run on a Mac. We are adding it on our servers for slower, more careful reviews. Client work stays in our office.

A high-speed second pass. DualEntry has it at 88.1% — above Claude Fable 5.1 — and 76.2% of tasks passed on all three runs. A person still signs off every figure.

Hugging Face: zai-org/GLM-5.3-Flash · Z.ai (zai-org)

Published benchmarks

Published scores, not tests we ran. Linked to the source on each card.

DualEntry accuracy
88.1%
DualEntry pass line (2 of 3 runs)
80%
DualEntry consistency (all 3 runs)
76.2%
Claude Fable 5.1 (DualEntry)
85.1%

Start with the human exams: SAICA’s 2026 IAC pass rate (both sittings) was 73%, and 61% passed the December 2025 APC. The IAC pass mark is 50% overall. DualEntry is a different test — 101 accounting cases — with a harder bar: a case needs a score of 0.8, and it only counts as passed if that happens in at least two of three runs. GLM-5.3-Flash scored 88.1%. That is not only above DualEntry’s 80% pass line; 76.2% of tasks passed in all three runs, so it is consistently clearing the bar, not scraping a single sitting. DualEntry also has Flash ahead of Claude Fable 5.1 (85.1%) and Fable 5.2 (81.2%). DualEntry does not sit the IAC or APC. A person still signs off. We host the weights locally so financial data stays in our POPIA-secured office. DualEntry benchmark.

What independent testers have said

  • DualEntry DualBench100: 101 cases, 80% (0.8) pass in at least two of three runs. GLM-5.3 Flash 88.1% accuracy, 76.2% of tasks passed in all three runs — ahead of Claude Fable 5.1 (85.1%) and Fable 5.2 (81.2%). DualEntry Accounting AI Benchmark
  • SAICA: 73% passed the 2026 IAC (January and June sittings combined). 61% passed the December 2025 APC (2 314 of 3 801). SAICA IAC June 2026 media release

Professional judgement. A person still signs the work. Human accountants use AI for analysis and automation. Software proposes; a person reviews every figure and applies their own professional judgement. Nothing is filed, signed, or posted to a ledger by a model.

For general work on a Mac Studio, people recommend Qwen3.8-27B. We run the 176B Flash-Next checkpoint instead — Mac Studio local AI vs what we run.

Slow reviews — DualEntry

CA(SA) pass rates, then GLM-5.3-Flash

Human exam pass rates first. Then DualEntry: GLM-5.3-Flash is not only above the 80% pass line — 76.2% of tasks passed in all three runs, so it is consistently passing.

MetricHuman CA(SA) candidatesGLM-5.3-Flash
Overall pass / accuracy73% IAC (2026, both sittings) · 61% APC (December 2025 sitting)88.1% DualEntry
Passing threshold50% minimum overall mark on the IAC80% (0.8) on a case; the case counts as passed in at least 2 of 3 runs
ConsistencySingle high-stakes exam sitting76.2% of tasks passed in all 3 runs

On the same DualEntry table, GLM-5.3-Flash (88.1%) is ahead of Claude Fable 5.1 (85.1%) and Fable 5.2 (81.2%).

DualEntry is 101 synthetic accounting cases — not the IAC or APC. A person still signs off every figure.

Our RAG, not a public chatbot

RAG here means retrieval: the model can look up text we store on our servers (templates, internal notes, approved procedures). It does not send your ledgers to a third-party AI API for training.

Blu OS is our accounting-firm operating system — not a SaaS app. It produces financial statements from the ledger. The models sit behind that workflow. They do not post to a ledger on their own.

What this is not

  • Not a SaaS app you log into. Blu OS is the firm operating system; these models are tools we run for the firm.
  • Not Bluubin. We are not affiliated with that company.
  • Not a promise that an AI files your SARS return unattended. A person signs off.

We invested in the hardware to enable this. We host and test the checkpoints most practices never run — Qwen3.8-Flash-Next (176 billion parameters) and GLM-5.3-Flash (320 billion). Our vision is a private cloud on servers we run; we are working towards that. Client financial data is processed with local AI on our own servers.

We still use the internet and enterprise SaaS. We also use cloud AI on enterprise accounts with no-training agreements — including Microsoft and other specialised tools — so those models are not trained on that work. One firm cannot recode every tool an accountant can practically use.

Frequently Asked Questions

Which model do you use every day?

Qwen3.8-Flash-Next is the local workhorse. We serve Qwen3.8-Flash-Next at n_ctx 262144 — the full 256K native context window on our servers. Weights: Qwen/Qwen3.8-Flash-Next.

What context window do you run?

We serve Qwen3.8-Flash-Next at n_ctx 262144 — the full 256K native context window on our servers.

What is GLM-5.3-Flash for?

Slower, more careful reviews. 320 billion total parameters — specialised hardware, not a Mac-local model. We are adding it on our servers. Weights: zai-org/GLM-5.3-Flash.

How does GLM-5.3-Flash compare with CA(SA) pass rates?

SAICA’s 2026 IAC pass rate was 73%; 61% passed the December 2025 APC. DualEntry’s bar is 80% on a case, passed in at least two of three runs. GLM-5.3-Flash scored 88.1% on that test and passed 76.2% of tasks in all three runs — not only above the pass line, consistently passing. DualEntry has it ahead of Claude Fable 5.1 (85.1%). DualEntry is not the SAICA exam. A person still signs off.

Is this the model people run on a Mac Studio?

No. Mac Studio write-ups keep recommending Qwen3.8-27B for general work. We run Qwen3.8-Flash-Next (176 billion parameters) on our servers. Comparison.

When can I become a client?

New client intakes open 1 March, at the start of the South African financial year. Apply now to become a client from 1 March.

Apply for 1 March

New client intakes open 1 March, at the start of the South African financial year. Apply now to become a client from 1 March.

CallWhatsAppApply for 1 March