模型调研报告

The user wants a thorough look at the SenseNova-Loop-9B-Preview-GSQ-RCO-GGUF model and its versions, then a polished PDF on how strong its Artificial Analysis intelligence score is and which frontier models it gets compared against. I'm also thinking about what I actually know about SenseNova, which is SenseTime's model family. I don't recognize this specific model name, so I can't speak to its details from memory, and I have no web or file tools here to research it. I'd rather not invent Artificial Analysis scores or benchmark comparisons, since a report that looks authoritative but is fabricated would mislead. The useful path is to say that plainly and offer a template for the report that the user can fill in from the actual sources. I'm planning a methodology for estimating the AA intelligence index, a checklist of sources to verify (the model card, the Artificial Analysis site, and the benchmarks that make up the index), and a PDF template the user can fill in with real data. I'll also flag that GGUF is the llama.cpp quantization format, while GSQ and RCO are unclear and should be treated as guesses. Working through the history of the Artificial Analysis index: the early versions leaned on MMLU-Pro, GPQA Diamond, Humanity's Last Exam, LiveCodeBench, SciCode, and AIME-style math, then later versions swapped in long-context, instruction-following, and agentic tasks. I'll need to check the exact composition of the current version before relying on any of it. I'm recalling the Artificial Analysis Intelligence Index's evolution: version 3 used ten evaluations including MMLU-Pro, GPQA Diamond, Humanity's Last Exam, and LiveCodeBench, while my vague recollection is that version 4 swapped several of those for agentic and real-world tasks. I'm not certain of the exact v4 lineup, so I'd want to check the Artificial Analysis site before relying on it. Now I'm comparing the rough v3 scores for the frontier models, which I'd have to verify before relying on them, and trying to place a 9B-class model against them. I'm weighing whether the scores I half-remember for the frontier models are trustworthy enough to repeat. My recall of the Intelligence Index numbers is shaky, and the scale changed between versions, so I shouldn't state them as fact. I'd present any figures as rough and mark them for checking against the current Artificial Analysis site. I'm noting that AA only publishes Intelligence Index numbers for models it has benchmarked itself, and a GGUF quant of a 9B preview model is unlikely to be on that list. Any "AA estimate" here would have to come from the user or the community, so the report needs to keep official figures and estimates clearly separate. Without tools, my plan is to lay out the limitations and a checklist for building the estimate myself. Now I'm sketching out how to estimate the score myself, since AA likely hasn't benchmarked a GGUF quant like this: gather the component benchmark results, combine them the way AA weights them, and account for the loss from quantization. I'm also planning a polished PDF template with placeholder fields for the user to fill in once they have real numbers, covering a cover page, summary, and a table of quantization variants. I'm weighing whether to produce a full PDF script now or keep it lighter, since any data I put in would be a placeholder. I'd lean toward a template with a single data block left empty and a "data to be verified" watermark, so nothing invented can slip into a polished document. I'm weighing whether to include a full PDF script or keep things lighter, since I'd have to leave the data fields empty anyway. A middle path seems right: be upfront that I can't verify this model, offer a data-filled-later template, and ask for the model card or benchmark numbers so the final report rests on real sources. I'm weighing the options for the answer: say plainly that I can't find this model, don't invent a report, and offer a methodology and a PDF template with placeholder fields rather than made-up numbers. I'm going to keep any reference scores out unless I can verify them, since the Intelligence Index has changed scale across versions. Next I'm weighing whether to name specific comparison models and their scores, since a polished PDF with wrong numbers would mislead. I'll keep the comparison at the level of categories: same-size open models on one side, frontier models on the other, with any figures marked as approximate until checked against the current Artificial Analysis site. I'm weighing how the score of a 9B model would likely compare. Under the v3 index, models this size usually fall in the 30s to 40s, and claims that a small model matches frontier systems tend to come from selective benchmarks like AIME rather than the composite index, which also covers agentic tasks. Now I'm recalling rough AA scores for small models, but I'm not confident in the exact numbers, since they shift between index versions. gpt-oss-20B, Qwen3 8B, and Nemotron Nano 9B are the comparison points that come to mind, and I'd need to check their current figures on the Artificial Analysis site before putting any of them in a report. I'm weighing whether to cite approximate scores for small models like Qwen3 4B, Qwen3 30B A3B, and EXAONE 4.0 32B, along with frontier systems like GPT-5, Gemini 3 Pro, and Claude 4.5 Sonnet. I don't trust my recall of these numbers enough to put them in a polished report, especially since the index has changed versions and scale. I'm leaving out specific AA scores for now, since the index has changed versions and scale, and anything I quoted from memory could be outdated or wrong. The user would get a polished PDF with numbers that look authoritative but aren't checkable. I'm checking whether my version-specific claims hold up: I think v4 changed the evaluation set and pulled the top scores down, but I'm not fully sure of the details. Since the scale shifts between versions, I shouldn't compare numbers across them. I'm wondering whether this model might be real. SenseTime has released several models on Hugging Face under the "sensenova" organization, but I don't recognize "SenseNova-Loop" itself, so it may be newer than what I know, and the GSQ and RCO parts of the name are unclear to me. I'm noting that GSQ and RCO don't match any standard llama.cpp quant types like Q4_K_M or IQ4_XS, so they're probably the publisher's own quantization or tuning scheme, and I'd need the model card to confirm what they mean. I'm checking whether this repo is legitimate, since the "-GGUF" suffix and tags like GSQ and RCO don't look like standard quant naming. The safest route is SenseTime's official HuggingFace, ModelScope, or GitHub pages, and I'm not certain of the exact org names. I'll keep any AA claim the user mentions separate from what's verifiable. I'm thinking about the "Loop" part of the name. If it's a looped model that reuses layers, like ByteDance's Ouro, then a standard GGUF conversion might not run correctly in mainline llama.cpp without custom support, which is one more reason to check the model card before trusting any claims. I'm weighing which tooling to use for the PDF. WeasyPrint from HTML and CSS looks like the easiest route to good typography, provided a CJK font is installed, and reportlab is the more verbose alternative that needs fonts registered by hand. Comparing a few PDF options now: an HTML page printed from the browser seems the least setup, with Typst and WeasyPrint as alternatives if I want more control over the layout. Writing HTML template code... Writing HTML template code... Still building the HTML report template... Now I'm noting that the AA site only scores models it has actually run, so a GGUF quant like this one probably won't have its own entry. Now I'm weighing how the AA index gets computed: I think v3 averaged its evaluations with equal weight, and v4 groups them into categories, but I'm not sure of the details, so the methodology page should be the reference. Now I'm sketching the comparison set: same-size peers on one side, frontier anchors on the other, with the expectation that a 9B dense model will usually sit well below frontier systems on composite indices, especially on agentic, knowledge, and long-context tasks. Now I'm deciding how the PDF itself gets built: a single self-contained HTML page that the user opens in a browser and prints to PDF, so no extra tools are needed. Now I'm checking the llama.cpp quant names: GSQ-RCO isn't among the standard types like Q4_K_M, Q8_0, or the IQ family, so it's likely the publisher's own scheme. Next I'm checking how the AA index is put together: it's a composite of several evaluations run by Artificial Analysis under consistent settings, and the versions keep changing. I also recall that AA scores models through provider APIs rather than community GGUF quants, so a quant like this one probably has no entry of its own. I'm weighing whether to mention the AA index scores for frontier and mid-size models, but I'm leaning against precise numbers since my recall of them is unreliable. A rough picture is more defensible: frontier closed models sit well above most 8–9B open models on composite indices, and I'd need the live AA page to confirm any specific range. I'm recalling rough AA v3 figures for small models, things like Qwen3 8B, Nemotron Nano 9B, and gpt-oss-20B, and they sit far below the frontier closed models. I'm not confident in the exact numbers, so I'd only state a broad range of something like 20–40 for the 8–9B class and keep it marked as unverified. I'm weighing whether to give actual AA numbers here, but my recall of them is unreliable, so I'd rather not. The more useful point is that the composite index leans on agentic work (Terminal-Bench Hard, τ²-Bench), broad knowledge (HLE, AA-Omniscience), and long-context tasks (AA-LCR). Small models are structurally weaker on those, since knowledge scales with parameter count and agentic use needs robust tool handling. So a 9B model claiming to match frontier systems on the composite would be unusual, and such claims usually rest on a few math or coding benchmarks. Now I'm working out how to word the AA caveat: if someone says the model scores near GPT-5 level, I'd check which AA version it is, which sub-evaluations are included, who ran the test, and whether it was a quantized build. I'll also ask the user for the figure they've already seen, since a release post may quote one. Now I'm filling in the data object and the layout for the HTML report, with the AA index score and estimate range as the core numbers. Now I'm setting the chart scale: the bars should be sized against the highest value in the data, so the target and peer scores stay comparable at a glance. Writing HTML template code... Still writing the report styles and layout... Still writing the report's CSS: the KPI cards and the section headers. Still writing the report's CSS: the bar-chart rows and the table styling. Still writing the CSS: the range markers and the bar labels. Still writing the report's CSS: the watermark overlay and the print rules for page breaks. Now I'm laying out the data object for the report, with the model specs left as unverified placeholders until I have the official model card. Still filling in the data object for the report, with the variant rows and the AA estimate fields. Now I'm filling in the rest of the DATA object: the peer rows and the sources list, with every score left null until I can pull it from the AA site for the same version. Now I'm working out how to place the user's model against the peers: I want to sort the peer scores and find the closest one just below and just above the estimate, so the KPI can say where it falls. Still writing the chart rows: the peer scores and the estimate go in, sorted from highest to lowest, with the range shown as a lighter band. Now I'm writing out the caveats for the methodology section: scores only count when they come from the same AA version, and anything AA hasn't published is a third-party estimate that has to state who ran it and how. Quantized builds need their bit width stated, since low-bit quants hit small models' reasoning and math harder, so I'd cap them against the BF16 or Q8_0 results. If only some sub-evaluations were run, it can't be called an AA index score, just a partial estimate. Now I'm working out how the report gets printed: the page margins come from the print rules, and in Chrome the "default" margin setting will pick those up. Now I'm listing the typical GGUF variants a repo would ship, like Q2_K up through Q8_0 and BF16, so the version comparison covers the full range of quants. Next I'm sizing the GGUF variants by bits per weight: Q8_0 at roughly 9.6 GB, Q6_K near 7.4 GB, Q5_K_M around 6.4 GB, Q4_K_M near 5.5 GB, Q3_K_M about 4.4 GB, and Q2_K somewhere around 3.5–3.8 GB. These are generic estimates, so I'm wondering whether to include them, since the real sizes depend on vocab and embedding layers. Working out the file-size math next: dividing a GGUF's size in GB by the parameter count in billions, then multiplying by 8, gives its effective bits per weight. For a 9B model at about 5.5 GB, that works out to roughly 4.9 bpw, which lines up with a Q4_K_M-class quant. Next I'm checking the file-size math against the GGUF tooling: `gguf-dump` from the gguf package should show the quantization type and file type in the metadata, which would confirm what a given file actually contains. Checking the GGUF metadata is the quickest way to confirm what the files actually contain, since the architecture and per-tensor quant types are both stored in the file header. Now I'm checking the AA methodology page for the current version's sub-evaluations and weights, then I'll see whether AA has already listed this model. If it has, that official score is the only one I can call an "AA index" figure. If not, I'd run as many of the sub-evaluations as I can on the BF16 or Q8_0 build, following AA's published settings, and mark anything I can't reproduce. Now I'm settling the rules for the report's status fields: it stays a draft while any model field is null, any variant is unverified, or no sources are listed. I'm also thinking about a small automatic touch for the third KPI card, placing the estimate between the nearest peer scores below and above it, and showing 待核实 when there's no estimate yet. Now I'm weighing whether smaller models really lag on knowledge-heavy and agentic tasks like Terminal-Bench and τ²-Bench; knowledge recall is clearly tied to parameter count, so I'll hedge that agentic gaps are usually larger too. Now I'm deciding which models to compare against: the frontier set and same-size open models, kept to whatever is currently listed on the AA leaderboard, since my examples may already be out of date. Still filling in the remaining data fields for the report. Now I'm getting into the report's structure: a name-decoding table, then how to build an AA estimate that holds up, then which models to compare against, and finally the PDF template. Now I'm sorting out the name pieces: SenseNova is SenseTime's model brand, and the rest needs checking. "Loop" isn't something I can pin down yet, so I'd look for the model card or the GGUF architecture field to see what it refers to. "9B" means roughly 9 billion parameters, though I still need to confirm whether it's dense or MoE. "Preview" means scores may shift once a final release lands. "GSQ-RCO" isn't a standard llama.cpp quant type like Q4_K_M or IQ4_XS, so it's probably a custom quantization or training-method abbreviation, and the model card should say. GGUF is the llama.cpp-ecosystem file format, and I'd check whether mainline llama.cpp supports the architecture. Now I'm checking whether "9B" means a dense model or an MoE, since the total and active parameter counts would change how the size math reads. For verification, I can open any .gguf file on the Hugging Face file page to see its metadata like the architecture name and per-tensor quant types, or run `gguf-dump` locally after installing the gguf package. Then I'm estimating effective bits per weight from file size times 8 divided by parameter count, so a 5.5 GB 9B file works out to about 4.9 bits and lines up with a Q4_K_M-class quant, which could reveal what GSQ/RCO actually contains. I'm on the AA estimation part now: the official AA Intelligence Index is a score AA computes itself from its own runs, so an unlisted model has no official number, and any figure I'd give is only a third-party guess. AA also usually tests the original-precision or hosted version, not community GGUF quants, so a quant's score would need its own measurement. Step one is pinning down the current index version and its sub-evaluations and weights on AA's methodology page, since a version change swaps the eval set and moves the scale, so scores from different versions aren't comparable. Now I'm checking the estimate steps: the official AA number only exists once AA has listed the model, so for anything else I'd self-test the BF16 or Q8_0 build as closely to AA's published settings as I can, mark the sub-evaluations I can't reproduce as missing, and not fill those gaps with other benchmarks. Now I'm pinning down the comparison set: the "对标" claim should be split into overall placement on the AA index and single-benchmark highlights, with peers chosen from whatever is currently on the AA leaderboard. Now I'm checking how the AA estimate should be framed: the 智能指数 is only an official number once AA has listed the model, so anything else is my own estimate, and it should be reported as a range with the missing sub-evaluations noted. Now I'm working on the HTML template for the report, starting with the document head and page setup. Still writing styles... Still writing styles... Still writing styles... Still writing styles... Still writing styles... Still writing styles... Still writing styles... Still writing the data block, the sources field and the closing lines of the config. Still writing the data block, the sources field, and the closing lines of the config. Still writing the ranking logic that places the estimate between the nearest peers. Still writing the position logic that compares the estimate against the sorted peer scores. Still writing the chart rows and the scale calculation. Still writing the bar renderer, which draws each row's band, fill and label. Still working on the band and fill layering: the estimate's fill should stay on the bottom, and the uncertainty range needs to sit on top so it remains visible. Still writing the band and label logic, with the estimate's uncertainty range drawn as an outlined box over the fill. Still writing the tables for the model fields and the variant rows. Still writing the table rows for the model fields, the variants and the components. Now I'm working out the KPI cards: the first shows AA's official index if it exists, otherwise "未收录/待核实", with the version noted underneath. The second shows my estimate with its range, and the third gives the position text in a smaller font so it fits. Still writing the notes block, the sources list and the closing method lines. Still writing the HTML report template, wiring the chart rows, tables, KPI cards and notes block into the page. Now I'm putting together the remaining sections of the report page: the model overview, the GGUF variant table, the AA sub-scores with their basis, and the start of the frontier comparison chart. Still writing the report's chart and legend markup, with the notes list and sources block following. Still writing the CSS for the watermark and the note block, with the band and legend styles set to sit above the table backgrounds. Next I'm escaping HTML special characters in the values, so a stray "<" in the data can't break the markup. Still writing the draft check, adding peers and components to the list of fields that mark the report as unverified. Still writing the draft check, adding peers and components to the list of fields that mark the report as unverified. Now I'm weighing whether to ship the template at all, since it gives the user most of the PDF they asked for and the watermark keeps unverified numbers from passing as fact. I can't browse here, so I'll say that outright, and note the model may be newer than my knowledge cutoff, so I'll point them to verifying it against the official AA listing. Still writing the version note, settling on wording that the index gets updated now and then, with its evals and weights changing, so scores from different versions aren't directly comparable. Now I'm settling on the wording for the name note: "Loop" may point to a looped architecture, where the same layers repeat several passes to deepen reasoning, but it could also just be a product or training-method label, so I'll present it as a guess. Now I'm going over the table entries for the model name parts, checking which ones I can actually confirm from the model card and the GGUF metadata. Now I'm checking how the GGUF export would handle the new architecture: if llama.cpp doesn't support it yet, the file may only run under a specific branch, so I'll note that in the report. Now I'm wrapping up the instructions: they'll save the template as an HTML file, fill in the DATA block, and print to PDF from Chrome or Edge with background graphics turned on. The watermark stays until every field is verified, so unchecked numbers can't pass as findings. Then I'll ask them to paste the model card text and the AA page data, since I can't open links. Still writing the HTML wrapper, with the page margins and base typography set. Still writing the CSS for the cover banner and its badges. Still writing the CSS for the KPI cards and their value text. Still writing the CSS for the report's layout grid and table rows. Still writing the CSS for the row layout and the bar tracks. Still writing the watermark styles, with the rotated overlay text in red. Still writing the DATA block, the title and subtitle are set, and I'm leaving the author field empty for now. Still writing the DATA block: the summary is left empty for now, and the model fields cover the publisher, parameter count, the meaning of the "Loop" and "GSQ-RCO" parts, context length, license, release date and llama.cpp support. The variant rows and the AA section come next, with the version, official score and estimate range. Still writing the escape helper, which turns the characters that would break the markup into HTML entities before any value reaches the page. Still writing the helper functions and the draft check, which flags the report as unverified if any key field is still missing. Still writing the ranking logic, picking the closest peers above and below my estimate to describe where it sits. Still building the bar rows, ranking the peers against my estimate and scaling the bars to the highest value. Still writing the notes block, with the AA band and the score labels wired into each row. Still writing the HTML template, wiring the notes list and the closing lines into the report. Still writing the KPI cards for the official index, the estimate and the overall positioning, with the version and the range shown under each value. Still writing the GGUF version matrix, with the file, quantization, size, and bit-width columns and the table rows being filled from the variant data. Still writing the table rows for the GGUF variants, with each file's quantization, size, bit-width and note shown, and the verified status marked beside it. Still writing the table rows and the bar chart, with the peer bars scaled against the top value and the target row marked with its estimate band. Still writing the report footer, which tells the reader that any figure marked as unverified should not be cited. Still working out the KPI cards, and I'm deciding how to show the official score when AA hasn't listed the model, so the card reads clearly whether or not there's a published number. Now I'm tidying the KPI cards so the official score reads as "—" when it's missing, with the index version noted underneath. Next I'm checking the bar fill, since an unrecognized type would leave the bar invisible, so I'll give it a neutral default color. Still writing the CSS for the bar fill, giving it a neutral default color so an unrecognized type still shows up, with the frontier, peer and target colors set afterward to override it. Still checking the layout details: the bar track is one fixed-width column next to the name, and on A4 the page leaves enough room so nothing gets cramped. Still checking the print settings: Chrome should print the cover gradient once background graphics are on, and I'll tell the user to tick that box to be safe. Now I'm double-checking the edge cases: a score of 0 still counts as present, the numbers get escaped like any other value, and the peer rows fall back to the neutral type when one is missing. Still writing the HTML template, and I'm renaming the score variable to `myScore` so it reads clearly against the chart rows. I'm deciding whether to pre-fill the peer list with frontier model names and leave the scores empty, so the user only has to add numbers. That would give a ready-made comparison set, but the names could go stale, so I'll keep them as suggestions rather than hard-coded data. I'm pre-filling the peer list with a few frontier and mid-size models, leaving their scores empty so the user only has to add numbers. The names could go stale, so I'm keeping them as suggestions to check against AA's current leaderboard. Still writing the peer list, checking the model names I'm pre-filling against what each release date implies, and leaving their scores open for the user to fill in. Still writing the peer list, pre-filling a few frontier and mid-size models with empty scores so the user only needs to add numbers. Still writing the peer list, pre-filling a few frontier and mid-size models with empty scores so the user only needs to add numbers. Still checking the opening caveats: I can't look up this model, I have no reliable information on it, and "AA intelligence score" only means something if AA has actually tested it. Still drafting the opening caveats: I can't verify this model, and an "AA intelligence score" only exists if Artificial Analysis has actually tested it. Now I'm checking how the quantization type and the bit-width figure relate: if a 9B file is 5.5 GB, that works out to about 4.9 bits per weight, which matches the Q4_K_M range. Still checking the GGUF metadata tips: the Hugging Face file page has a viewer for the .gguf metadata and tensor types. Still deciding whether to mention Ouro, a looped language model, as an example of the architecture. It's a real model family, but I'm not certain how llama.cpp handles it, so I'll describe "Loop" in general terms instead. Now I'm working through the expectation-setting paragraph: the AA sub-scores that track parameter count, the gap a sub-10B model usually shows on the composite, and the four things to check before trusting any claim. Now I'm working out how to write the AA estimate step, and I'm checking each point against the methodology page: the index version, the sub-scores it includes, and how they're weighted. Still filling in the peer list: I'm working out how to word the "对标" section so the composite positioning and the sub-score comparison come out as two separate points. Now I'm laying out the peer comparison in two parts: the overall composite placement, and the sub-scores where a few items come close to a particular frontier model. The "matches GPT-x" claims in marketing usually fall into that second group and don't mean overall strength is comparable. For reference, I'm drawing on models I know of, with current AA leaderboard entries taking priority. Same-size peers include Qwen3-8B, Nemotron Nano 9B v2, GLM-Z1-9B and the MiniCPM series, with gpt-oss-20B a bit larger but often used for local-deployment comparisons. The frontier anchors are GPT-5, Gemini 3 Pro, Claude Opus 4.5 and Sonnet 4.5, and Grok 4. Now I'm wrapping up the peer list and the usage steps for the HTML template: each number needs a source, and the print settings need background graphics turned on so the cover renders correctly. Still filling in the closing section of the report: I'm asking for the full model card, the file list with sizes, and the AA scores as pasted text or screenshots, so I can check the measurement basis, compute the effective bit width, and fill in the data fields. Now I'm laying out the table of name fragments: what each piece of "SenseNova-Loop-9B-Preview-GSQ-RCO-GGUF" can be stated about, and what still needs checking. "Loop" is the unclear one, and I'll describe it as a possible looped-depth architecture without asserting it. GSQ-RCO doesn't match a standard llama.cpp quant type, so I'll flag it as a custom scheme to confirm from the model card. Now I'm checking the closing notes on the GGUF tooling and whether the template's draft watermark logic holds up. Checking the edge cases in the JavaScript helpers now: the `every` checks on peers and components, and whether `ranked.filter(...).pop()` can come back empty when the list is non-empty. Still finishing the HTML template: I'm checking the layout and print-related CSS, and whether the body needs top padding so the cover isn't flush against the top edge on screen. Still tweaking the screen preview: a light gray background with a white, rounded panel around the report, so it looks finished in the browser. The print rules stay untouched. Still writing the HTML template, now on the screen-preview styles: a light gray background with a white, rounded panel around the report, while the print rules stay as they are. Still writing the CSS for the cover and badge styles... Still writing the CSS for the badges, headings and the three-column KPI cards... Still writing the CSS for the table cells and the value styling... Still writing the CSS, now the bar-row styling for the comparison chart... Still writing the CSS, now the legend and the floating watermark overlay. Still writing the CSS, now the legend and the floating watermark overlay. Now I'm putting together the JavaScript data block, with the title and subtitle set and a note that every number needs a source and that null values show as pending verification. Now I'm filling in the data block: the title and subtitle are set, and the comment reminds me that every number needs a source and that null fields will show as pending verification. I'm still working out the model metadata and the variants list, where each GGUF file gets its own row. Still writing the JavaScript data block: the AA section's version, official score and estimate fields are next, along with the components list. Still writing the JavaScript data block: the model metadata and the variants list come next, then the AA section's fields and the peer list. Still writing the JavaScript data checks: I'm defining the draft flags and the score helpers, and ranking the peers by their scores. Now I'm working out how to place the model against the peers: it lands between the closest lower and higher scores in the ranked list, and if it sits at an edge, it gets described as just below or just above the nearest reference. Still building the bar chart rows: each peer gets a fill sized to its score, and my own row gets a red band for the low-to-high range when it's known. Still finishing the JavaScript: I'm adding the band that marks the low-to-high range on my own row, sized against the shared scale. Still finishing the JavaScript: I'm checking the draft flags and the position helper, and making sure the score ordering and the low-to-high band on my own row line up with the shared scale. Writing the final answer now: two caveats up front about what I can't verify, then the report. Now I'm checking how AA scores get measured: their index is run by AA under its own settings, so a model they haven't evaluated has no official number, and community GGUF quants usually go unscored. Now I'm checking how the GGUF file size translates into bits per weight, which will show whether the quant level is plausible for a 9B model. Now I'm checking the GGUF file size against parameter count to estimate bits per weight, and confirming the publisher is the official SenseTime account before trusting any scores. Now I'm double-checking a claim I'd made about AA's methodology: they test open-weight models through API endpoints, either the creator's own or a third-party host, and those deployments may run at reduced precision like FP8, so "original precision" isn't guaranteed. Now I'm checking whether the model's place on the AA scale can be stated at all: the methodology page should show how AA sets up its runs, including the prompts and the settings they use. Still working on the verification-and-report part: I'm noting that the bit-width estimate from file size is approximate, and that quantization levels like Q4_K_M land near 4.8 to 4.9 bits per weight, so the 5.5 GB example lines up with that. Still writing the JavaScript helper that wraps each value in a pending-verification span when it's missing. Still working on the JavaScript helper that builds the chart rows: when a value is missing, it shows a pending-verification label instead of a number, and the low-to-high band only appears on my own row. Still writing the JavaScript notes array, with the caveats about AA scores and their versions. Now I'm finishing the JavaScript that builds the report body: the notes list is set, and the cover section is being assembled from the title, subtitle and the draft badge. Still writing the KPI cards: the official index, the estimate with its range, and the positioning summary. Still writing the HTML template: the model overview table and the GGUF version matrix are next, with each file's quant type, size and bit-width filled in from the data. Still finishing the HTML report body: the GGUF version table and the AA sub-score section are being filled in from the data, with a pending-verification mark wherever a field is missing. Still building the report's sources section: I'm looping over the source list to render each reference as a numbered item. Still writing the report's sources and footer: I'm rendering each reference as a numbered list item and closing with a note that unverified figures shouldn't be cited. Still writing the bar rows: each peer's fill is sized against the shared scale, and my own row gets a class that marks it as the target, so the bar styling stays consistent. I'm checking the reference list and the remaining wording: the peer models are grouped in the comparison, and the quantization claim stays hedged, since smaller models generally degrade more under aggressive quants and reasoning tasks are more sensitive. I'm adding a line to the 预期管理 section: with no measured data, the sensible prior is that the model sits among similarly sized open models rather than the frontier tier, but that's only a size-based guess and shouldn't go into the report as a finding. Now I'm spelling out how the AA sub-scores tend to behave: knowledge breadth, agentic tasks and long-context evals usually track parameter count more closely, while a sub-10B model can sometimes come near the frontier on narrow items like competition math, so I'll need to check which AA version, which sub-scores, and who ran the test before drawing any conclusions. Wrapping up the last piece: I'm adding a note on expectations, that a 9B model's composite AA score is unlikely to match the frontier tier, and that the details worth verifying are which AA version, which sub-scores, who ran the evaluation, and whether the tested model was quantized. # SenseNova-Loop-9B-Preview-GSQ-RCO-GGUF 调研：能确定什么、怎么预估、以及 PDF 模板先说明两条边界。这样做出来的报告不会排版精美，数字却是编的。 1. **我这边无法联网检索，训练数据里也没有这个模型的可靠信息。** 它可能发布在我的知识截止之后，也可能是第三方转制或非官方命名。所以我给不出它的真实 AA 分数，也不会编一个。 2. **"AA 智力"需要先厘清口径。** Artificial Analysis Intelligence Index 是 AA 用统一设置自己实测、再合成的分数。 - AA 没收录的模型，不存在"官方 AA 分"。 - AA 一般测原精度或服务商部署的版本，通常不会给社区 GGUF 量化文件单独打分。 - 网上"AA 预估 XX 分、对标某前沿模型"之类的说法，基本都是第三方估算，必须追问出处。下面是我能可靠提供的内容：名字拆解与核实方法、站得住的预估流程、对标组建议，以及一个可直接导出 PDF 的模板。数据未核实时，模板会自动加"草稿"水印。 ## 一、从名字能读出什么 | 片段 | 能确定的 | 需要核实的 | |---|---|---| | SenseNova | 商汤科技"日日新"大模型品牌 | 发布者是商汤官方组织，还是第三方转制 | | Loop | 无法确定。可能指循环/递归深度架构（同一组层重复计算多轮），也可能只是方法代号 | 模型卡；GGUF 元数据里的 `general.architecture` | | 9B | 约 90 亿参数 | 是稠密模型还是 MoE（总参数 ≠ 激活参数） | | Preview | 预览版 | 正式版发布后，能力与分数可能变化 | | GSQ-RCO | **不是** llama.cpp 标准量化类型（标准命名如 Q4_K_M、Q8_0、IQ4_XS） | 可能是自定义量化/校准方案或训练方法的缩写，看模型卡 | | GGUF | llama.cpp 生态的模型格式，Ollama、LM Studio 可加载 | 若是新架构，主线 llama.cpp 是否已支持 | **核实技巧：** - **看元数据。** 在 HuggingFace 文件页点开任一 `.gguf` 文件，可以直接看到： - `general.architecture`：判断 Loop 是不是新架构； - 各张量的量化类型：判断 GSQ-RCO 底层实际是什么。 - 本地也可以 `pip install gguf` 后运行 `gguf-dump 文件名.gguf`。 - **估算有效位宽：文件大小(GB) × 8 ÷ 参数量(B)。** 例如 9B 模型的 5.5 GB 文件约为 4.9 bit/权重，相当于 Q4_K_M 档。用它可以反推每个版本的实际压缩程度。 - **区分官方版与转制版。** 转制版的分数不能直接套用到官方版上。 ## 二、"AA 智力预估"怎么做才站得住 1. **锁定版本。** 到 artificialanalysis.ai 的方法论页面，确认当前指数版本、包含哪些子项、如何加权。AA 会不定期升级版本（增删评测集、调整权重），分值刻度随之变化，跨版本的分数不能放在一起比。 2. **先查官方收录。** 如果 AA 已收录，直接用官方分，并记下查询日期。 3. **未收录才自测。** - 用 BF16 或 Q8_0 版本，尽量按 AA 公布的设置跑各子项。 - 跑不了的子项标注缺失，不要拿其他评测替代凑数。 4. **给区间，不给点值。** - 缺失子项越多、自测设置与 AA 差异越大，区间就越宽。 - 只覆盖部分子项时，写"部分子项估计"，不能写成"AA 指数 XX 分"。 5. **量化版单列。** Q4 及更低位宽对小模型的长链推理和数学通常影响更明显。各 GGUF 版本应分别测，至少标注"预计低于原精度，幅度待测"。 ## 三、对标哪些模型 "对标"建议拆成两件事，在报告里分开写： - **综合定位**：预估分在 AA 同版本榜单上落在哪两个模型之间。模板会自动算出这两个"上下邻居"。 - **单项亮点**：哪几个子项接近哪个前沿模型。宣传里的"对标 GPT-x"多数属于这一类，不等于综合实力相当。 **参照组**以 AA 当前在榜模型为准。以下是我知识范围内的例子，现在很可能已有更新版本： - **同量级**：Qwen3-8B、NVIDIA Nemotron Nano 9B v2、GLM-Z1-9B、MiniCPM 系列。gpt-oss-20B 略大，但常被拿来做本地部署对比。 - **前沿锚点**： - 闭源：GPT-5 系列、Gemini 3 Pro、Claude Opus 4.5 / Sonnet 4.5、Grok 4； - 开源：DeepSeek V3.2、Kimi K2 Thinking、Qwen3-235B-A22B、GLM-4.6。 **预期管理：** - AA 指数里有几类子项通常更依赖参数规模： - 知识广度（如 HLE）； - 智能体任务（如 Terminal-Bench Hard、τ²-Bench）； - 长上下文（如 AA-LCR）。 - 按以往榜单规律，10B 以下模型在单项（尤其数学竞赛题）上接近前沿并不罕见，但综合指数追平前沿梯队会是非常反常的结果。 - 在没有实测数据之前，合理的先验是"落在同量级开源模型区间"。这只是基于规模的先验，不能当结论写进报告。 - 看到"9B 对标前沿"的说法，优先核实四件事： - 哪个 AA 版本； - 哪些子项； - 谁测的； - 测的是不是量化版。 ## 四、PDF 报告模板保存为 `report.html`，只需修改顶部的 `DATA`： ```html 模型调研报告

``` **导出 PDF：** 1. 用 Chrome 或 Edge 打开 `report.html`。 2. 按 `Ctrl+P`（Mac 为 `⌘P`），目标选"另存为 PDF"。 3. 在"更多设置"里勾选"背景图形"。只要还有未核实字段，封面就会标"草稿"，每页带水印。这是有意设计的，防止未经核实的数字被截图传播。 --- 如果你把以下材料贴给我（复制文字或截图都行），我可以帮你核对口径、算出各版本的有效位宽、填好 `DATA`，并写出结论段和对标分析： - 模型卡全文； - 文件列表（文件名 + 大小）； - AA 页面上相关模型的分数。

SenseNova-Loop-9B-Preview Analysis