OsaurusAI/Diplodocus-preview-JANG_6M

Diplodocus preview 4, the Osaurus assistant for finance, documents, spreadsheets and tool-assisted work, built on IFM/K2-Horizon-7B and trained on Osaurus' own workspace tools (workbooks and documents edited in place, the built-in calculate tool, reasoning on and off). This calibrated JANG affine mixed-precision bundle is 6.94 GB of weights (6.46 GiB; about 6.96 GB to download).

Osaurus 0.25.21 or newer is required. osaurus.json identifies this as model version 3. Download the complete bundle when upgrading from the preview-3 weights published under this name.

What preview 4 adds over preview 3

Trained rounds R8–R11 on top of preview 3: the Osaurus calculate tool (when to call it, chained money arithmetic, recovery from engine errors), workbooks and documents through Osaurus' in-place file_edit operations (bounded file_read windows, mode:"structure", set_cells with live formulas, never regenerating an existing workbook), PDF forms and signed-PDF handling, and a loop fix for reasoning-off answers. The source's own held-out evaluation moved end-to-end workbook/document tasks from 59/51 (on/off, of 100) to 70/70 and calculator-executed long-document money answers from 31 to 43 of 60 with reasoning on.

Quantization

Layer numbers are zero-based. The packed weights total 6,938,269,080 bytes. All 254 quantized module layouts were cross-checked against their packed tensor shapes. Precision is placed where it matters: every MLP module was measured individually at 4, 5 and 6 bits (the module substituted into the BF16 model and scored on held-in calibration text), and the byte budget was spent greedily on the modules with the best quality gain per byte. The same measurement showed that the 250k-row embedding and output head cost almost nothing at 6-bit, so their bytes went to the MLP instead; the resulting bundle is both smaller than an all-6-bit MLP and better than the preview-3 recipe at the same size.

Calibration uses importance statistics, AWQ and GPTQ over 3,295,622 tokens: Osaurus tool conversations and calculator chains, preview 4's own generated traces across 60 task families (workbooks, PDF/Word/PowerPoint documents, finance with the real calculator engine, math, agentic coding, structured output, translation, reasoning on and off) rendered with Osaurus' current compact workspace contract, and 30% generic text (coding, reasoning, agentic, terminal, security, long-context recall). The bundle uses MLX-compatible affine packing with per-module precision overrides. A plain uniform-bit MLX loader is insufficient: use Osaurus or a compatible JANG-aware vmlx-swift runtime.

Reasoning and tools

• High: reasoning on. Low: reasoning off, using the supplied template's empty prefilled reasoning block. Do not send medium.
• Preserve assistant history and tool history across turns, and use the included tokenizer and chat template.
• Tool calls use the native k2_horizon envelope; Osaurus parses them (a doubled <ifm|tool_calls> opener is treated as one).
• Default sampling is temperature 1.0, top_p 0.95.
• For money and exact arithmetic use reasoning on and let the model drive the calculate tool; it rounds to the document's rule itself and carries rounded figures into the next call.

Validation and limits

Held-out quality against the BF16 source on 646,927 positions (Osaurus tool conversations, calculator chains, preview 4's own held-out traces, reasoning, long-context recall): restricted top-128 KL median 0.0050 / mean 0.063, top-1 agreement 92.6%, call-vs-answer flips 1 of 430 decisions, think-close agreement 424 of 429, reasoning-off tag integrity 35 of 35. The preview-3 recipe built on the same source and calibration scores 0.0172 / 0.170 / 87.0%, and BF16 with one-ULP embedding noise (the measurement floor) scores 0.0013 / 0.020 / 96.2%. These are quantization diagnostics on training-distribution data, not an independent task-accuracy benchmark.

The bundle was exercised in the real vmlx-swift runtime (greedy, reasoning on and off): 14 of 14 probes finished cleanly with the Swift-rendered prompt identical to the reference tokenizer's, including workbook set_cells edits through the compact Osaurus tool contract, a PDF file_search, native file_read/file_write/calculate calls, the stale-tool and identity probes, and the reasoning-off tag check. A calculator-backed invoice ran as a real multi-turn tool loop against the Osaurus calculator engine: six model turns, five calculate calls, correct $283.71. Unaided mental arithmetic remains a model limitation (one of two unaided invoice probes landed on $283.70 after a digit slip in the tax line; BF16 preview 3 made the same class of error), so money goes through the calculator. See evaluation/swift-probes-summary.json and evaluation/invoice-loop.json.

Known model limits (from the source's own evaluation): new workbooks with live formulas sometimes end up with text formulas; change requests on signed PDFs with reasoning on are handled in 3 of 10 cases; money arithmetic with reasoning off is unreliable (use on); the model may identify itself as K2 and may attempt a tool from older conversation history. Use current tool definitions and validate consequential outputs.

Version 3 changes

Replaces the preview-3 weights with preview 4 (R11 u220) and a measured per-module precision plan; same tokenizer, template, reasoning toggle and tool protocol.

한국어 안내

Diplodocus 프리뷰 4의 JANG 혼합 정밀도 affine 양자화 모델입니다. 가중치 크기는 6.94 GB이며, 어텐션은 8비트, MLP는 모듈별 측정에 따라 6·5·4비트, 임베딩과 출력 헤드는 6비트입니다. Osaurus 0.25.21 이상을 사용하고, 추론은 high로 켜거나 low로 끌 수 있습니다. 금액 계산에는 추론을 켜고 calculate 도구를 사용하세요.

Quantized and validated by Jinho Jang — eric@osaurus.ai