A locally converted Q8_0 GGUF build of Apodex-1.0-0.8B-SFT, prepared for llama.cpp-based runtimes such as LM Studio. The upstream model is distributed in BF16 Safetensors; Apodex does not currently publish an official GGUF for this 0.8B checkpoint. This repository publishes our conversion, not an Apodex-authored or Apodex-endorsed GGUF.
The conversion harness preserves the separate Qwen3.5 recurrent beta/alpha and QKV/gate tensors expected by the LM Studio runtime used in our local load check. This addresses the earlier loader error in which blk.0.ssm_beta.weight was missing because the converter had fused beta and alpha into one tensor.
| File | Purpose | Size |
|---|---|---|
Apodex-1.0-0.8B-SFT-Q8_0.gguf |
Q8_0 language model | 833,592,384 bytes (about 795 MiB) |
mmproj-model-f16.gguf |
F16 vision projector; include alongside the model when using image input | 204,987,104 bytes (about 195.5 MiB) |
convert_q8_0.sh |
Pinned, reproducible conversion and validation harness | — |
LICENSE |
Upstream Apache-2.0 license text | — |
The language GGUF is about 794.9 MiB on disk. The vision projector is a separate optional companion for multimodal use; it is not part of the Q8_0 language-model quantization. The harness produces and validates both artifacts.
mmproj-model-f16.gguf beside it.lms import /path/to/Apodex-1.0-0.8B-SFT-Q8_0.gguf (LM Studio import docs).This artifact was loaded locally in LM Studio with a requested 32,000-token context during the testing session. That is a local smoke check, not a claim that every prompt will use the full context effectively or that all LM Studio versions behave identically.
apodex/Apodex-1.0-0.8B-SFT58d2248f4f74470b53cd5993e7c58f1fd12cf8b8llama.cpp, pinned to commit 951d465f5c7c16fb486ff45bb14d99984e68c6dbLICENSEThe harness applies a narrowly scoped conversion-time adjustment to llama.cpp’s Qwen3.5 conversion code so it does not merge the beta/alpha or QKV/gate projection pairs. It then converts the source checkpoint, checks GGUF magic and quantization metadata, verifies separate tensor names across recurrent blocks, and only then deletes the downloaded source checkpoint and temporary converter checkout. It does not alter or fine-tune the model weights before conversion.
The upstream author has not published an official GGUF for this exact checkpoint. A third-party Q4_K_M conversion exists; its model card identifies the same upstream checkpoint and llama.cpp GGUF-my-repo as its conversion path. This repository’s distinguishing artifact is our Q8_0 conversion and its pinned, inspectable harness—not the first GGUF conversion of Apodex-0.8B.
Requirements: Bash, curl, tar, Python 3, PyTorch, and huggingface_hub. A CUDA-enabled PyTorch environment is not required for this conversion; the weights are converted on CPU. Ensure there is enough free disk space for the downloaded checkpoint, temporary converter checkout, and output files.
cd Apodex-1.0-0.8B-SFT
PYTHON=/path/to/python ./convert_q8_0.sh
If the default python3 has the dependencies, run ./convert_q8_0.sh. The source model is public, so a Hugging Face token is not required. If access policy changes, authenticate using Hugging Face’s supported local login/environment mechanism; do not put a token directly in this script or in the model card.
The script downloads the pinned source revision and converter archive into temporary locations, writes the two GGUF outputs, validates them, then removes its source-weight download and temporary converter files. The final GGUFs remain in the model directory.
This section records an informal, qualitative comparison shared by the model tester. It is not a standardized benchmark, does not establish general model capability, and was not run as repeated controlled trials. Models were asked the same core question:
why did jensen huang resigned from nvidia?
The evaluation target was the tester’s specialized definition of deep autonomous research: recognize uncertainty, use the available MCP research tools when appropriate, inspect evidence, handle false or conflicting claims, and produce a grounded answer. The reported session included Fortress MCP and varied reasoning/“think” modes. Exact decoding settings, identical conversation state, tool-server conditions, and repeated-run counts were not preserved consistently, so rank differences should be treated as observations from these runs only.
For completeness, the saved LM Studio records dated 13 November 2026 were filtered to conversations whose user message contained the exact benchmark prompt above. That yielded 15 distinct loaded model identifiers. The additional runs below were shared separately by the tester in this project conversation. The tables include failed, malformed, and incomplete runs; “tested” does not mean the model completed the research task successfully.
Quantization variants of the same checkpoint count as one model identity.
| Rank | Model | Observed result in this prompt run |
|---|---|---|
| 1 | Tholos-2B | Searched and opened a relevant current report, then produced the strongest usable synthesis among the runs shared. It still needed stronger cross-checking and clearer separation between confirmed facts and allegations. |
| 2 | Apodex-1.0-2B-SFT | Showed useful research behavior, but the final answer overclaimed details. |
| 3 | Apodex-1.0-0.8B-SFT | Had the strongest autonomous research streak of the tested models, but its final response was malformed and unreliable. The tester considered it highly promising despite that output. |
| 4 | Mini-Agent-v2 | Placed fourth by the tester in the comparative ranking; a detailed response transcript was not included in the records summarized for this card. |
The charts below visualize the tester’s ordinal ranking and a retrospective, one-category-per-model coding of the 24 model identities with exact-prompt records documented here. The outcome groups are descriptive labels, not quality scores: the four ranked models are kept together as “tester-ranked top four” even though their observed response quality differed. The remaining models are grouped by the primary behavior visible in the shared record.
| Retrospective primary outcome category | Models |
|---|---|
| Tester-ranked top four | 4 |
| Searched; final answer overclaimed or left the evidence unresolved | 2 |
| Tool activity; no usable research conclusion | 4 |
| No effective tool research; false, speculative, or refusal response | 12 |
| Empty or incoherent output | 2 |
| Total distinct model identities | 24 |
These counts describe only the single shared prompt and the records summarized here. “No effective tool research” includes responses that did not call a tool or did not produce effective evidence gathering; it is not a claim that a model lacks tool-use capability. The first four models are shown in the tester’s confirmed order, not derived from chart scoring. Chart source is assets/make_charts.py; rerun with python3 assets/make_charts.py (requires Matplotlib).
| Model | Observed failure or limitation in the shared run |
|---|---|
| Granite 4.0 H-1B | Confidently invented a 2024 resignation date, a stated corporate reason, and a false career timeline; no research tool use was recorded. |
| Hammer 2.1-1.5B | One run returned an empty array; another emitted raw function-call JSON instead of performing the search or answering. |
| MiniCPM5-1B-Agentic-Tooluse-v3 | The saved run issued many tool calls (including template, cookie, screenshot, and Reddit tools), but the selected record contains no final natural-language answer. It shows tool activity without a verifiable research conclusion. |
| xLAM-1B-FC-R | Produced confused, internally inconsistent text, including confusing Jensen Huang with a different person/company; no tool call was recorded. |
| Qwen2.5-Hermes-Instruct-1.5B | Emitted malformed/mixed-language tool-like text and an unsupported resignation date rather than a grounded answer. |
| Uraion-Agent-Small | Invented a resignation, bankruptcy, and Microsoft role with extensive fabricated financial details; a second run continued with unsupported claims. |
| Animica-Agent-2B | Initially said its tools could not answer; a later attempt made repeated invalid Reddit URL calls without producing a grounded answer. |
| Qwen3.8-2B-Function-Calling-xLAM-Unsloth | Hallucinated a resignation, false successor, and unrelated company claims; its later tool intent was malformed and not executed. |
| Arch-Function-Chat-1.5B | Returned a JSON function-call-shaped object naming a nonexistent tool, with no research or answer. |
| Arch-Agent-1.5B | Fabricated a 2019 resignation and false career history without tool use. |
| xLAM-2-1B-FC-R | Offered generic speculative reasons for a resignation without establishing that it occurred or using tools. |
| Qwen3-1.7B | Invented a 2018 resignation and then described its reason as confidential; no tool call was recorded. |
| Model | Observed result in the shared run |
|---|---|
| Youtu-LLM-2B | Searched, but its final answer treated allegations as established facts and added claims not substantiated by the displayed results. Its reasoning trace also initially misidentified Huang’s company affiliation. |
| QUEST-2B | Searched, but dismissed current results as fabricated because their dates were beyond its training cutoff, repeated searches, and did not resolve the contradiction. A Fortress fetch also returned an unrelated Reddit destination, so tool output needed validation. |
| BotPsy-0.8B | Recognized the need for current information and said it would search, but did not follow through with a tool call. |
| Toolcall-2B | In one mode refused the public factual question as private information; in another, said it lacked search tools despite the tools being available. |
| Qwen3.5-2B | Confidently fabricated incompatible resignation dates, career history, age, and motives across the shared modes. |
| Falcon-H1-1.5B-Deep-Instruct | Invented a prior resignation and unsupported internal disputes, without using tools in the shared response. |
| Falcon-H1-Tiny-R-0.6B | Produced incoherent repetitive text rather than an answer. |
| Index-1.9B-Chat | Confidently invented a 2019 resignation, fabricated company statements and career details, and did not use tools in the shared response. |
These notes summarize the exact-prompt saved records and transcripts supplied by the tester; they are not independent source audits or claims about every run of each model. Together, the prompt-specific groups cover 24 distinct model identities. The primary conclusion is limited: in this specific prompt and tool setup, Tholos ranked first overall, while Apodex-0.8B showed the strongest autonomous research streak but had a malformed and unreliable final answer.
The prompt presupposes a resignation. A reliable answer must first check whether that premise is true and consult current, primary sources. The NVIDIA announcement confirms a CEO transition, but it does not itself establish a reason. Reports discussing personal relationships or alleged espionage must be attributed as allegations and not presented as proven fact. See NVIDIA’s announcement, its board statement, and AP’s reporting. The conversations are being evaluated for research process and evidence handling; this card does not endorse every claim made by a tested model.
This is a community conversion of Apodex-1.0-0.8B-SFT, released by Apodex under Apache-2.0. The conversion and local prompt comparison were performed by the uploader. This repository is not affiliated with or endorsed by Apodex. See LICENSE and the upstream model card for the original model’s documentation and citation.
apodex/Apodex-1.0-0.8B-SFT