Channel
@witcheergrimoire
On this record: Growth · Engagement · Reactions · Posts · Citations · Cite this entry
231subscribers
+34 since we began measuring on 8 August 2026
Risers and fallers across the register · movement among entries of Under 1,000.
| Telegram ID | -1003666474428 |
|---|---|
| Type | Channel |
| Username | @witcheergrimoire |
| Created | Between 1 December 2025 and 31 May 2026 — estimated from Telegram’s id allocation, not measured. How this range is calculated. |
| First recorded | 8 August 2026 |
| Last confirmed live | 17 September 2026 |
| Measurements held | 7 |
| Confirmed unchanged | 1 time, most recently 17 September 2026 |
| On Telegram | t.me/witcheergrimoire |
| Measured (UTC) | Subscribers | Change |
|---|---|---|
| 17 Sept 2026, 09:19 | 231 | +5 |
| 8 Sept 2026, 19:15 | 226 | +1 |
| 30 Aug 2026, 18:15 | 225 | +3 |
| 22 Aug 2026, 17:54 | 222 | +19 |
| 15 Aug 2026, 13:08 | 203 | +6 |
| 8 Aug 2026, 11:35 | 197 | no change |
| 8 Aug 2026, 09:30 | 197 | first reading |
20 posts held, back to 26 June 2026 — the reader has not yet reached the start of this channel’s public history, so older posts may sit further back, unread. Read across 1 page of Telegram’s post history, 20 posts per page.
Nothing published in the last 30 days. ERR and ER are rolling 30-day measures, so there is nothing to compute — we hold 20 posts for this entry, the most recent from 6 August 2026. An engagement rate over an empty window would be a number about nothing.
63 reactions across 19 posts, in 6 distinct kinds. The most used accounts for 38.1% of them.
| Reaction | Count | Share | Share, drawn |
|---|---|---|---|
| ❤ | 24 | 38.1% | |
| 👍 | 21 | 33.3% | |
| 🔥 | 15 | 23.8% | |
| 👏 | 1 | 1.59% | |
| 💯 | 1 | 1.59% | |
| 🤝 | 1 | 1.59% |
No sentiment is inferred, and none should be read in. This table is ordered by count and by nothing else. Emoji do not carry stable meaning across languages or communities — 🙏 is thanks in one channel and mourning in another — so we publish which ones were pressed and how often, and pass no judgement on what an audience meant by them.
Precision. Telegram publishes reaction counts per emoji and short-forms each one — 4.34K, 1.2M — so any single kind at or above 1,000 reaches us at three significant figures, and only counts below 1,000 are exact. The shares above are ratios of those figures and carry the same error. This is also why the total here can differ slightly from a reaction total printed elsewhere on the page: both are sums of the same rounded parts, taken over samples with different edges.
Coverage. Reactions were read on 19 of the 20 sampled posts in this sample. Summed by Telegram’s own count on each post — not by adding up the per-emoji breakdown above — those same posts carry 63 reactions in total: the kind of figure the paragraph above means by “a reaction total printed elsewhere on the page”.
Measured over the 20 most recent posts we hold, published 26 June 2026 to 6 August 2026, using the newest reading held for each. Telegram Stars are excluded: they are a payment, not a reaction, and they have their own section.
🔬 reasoning on/off on a 2.6B model I benchmarked LFM2.5-2.6B on native tool-use: 30 tasks, reasoning on vs off, one RTX 5090, greedy. first pass: 96.7% both legs, McNemar p=1.0. a perfect null. except the "off" leg still logged ~700 reasoning tokens per task, which should have been near zero. the cause: LFM2.5's GGUF chat template hard-forces <think> on every assistant turn and reads none of enable_thinking / reas…
👍1
Signed witcheer
real NVFP4 on a 5090: Blackwell's 4-bit float lands on the frontier NVFP4 is NVIDIA's native FP4 format for Blackwell. I already measured the honest K-quant frontier of Qwen3.6-27B (the quant-tax ladder), so a real NVFP4 GGUF of the same base drops straight onto it as one more level, same harness (MMLU/GSM8K/HumanEval, greedy, think-off), one RTX 5090. it lands on the frontier. composite 92.11, between Q5 and Q6, w…
❤2
Signed witcheer
the quantization tax on Qwen3.6-27B, measured side by side the folk rule is: run the biggest GGUF quant that fits your VRAM, lower bits means a worse model. I ran the whole K-quant ladder (Q8_0 down to Q3_K_M) on one RTX 5090, same model and same harness every rung, scoring MMLU + GSM8K + HumanEval alongside decode speed and file size. down to Q4 the tax is inside the noise: - Q8_0: 92.3 quality, 52.8 tok/s, 29.0 …
❤8
Signed witcheer
touched grass a bit end of July, coming back tomorrow with new benchmark, and tests! I’m currently eyeing on 1 or 2 DGX Sparks, need to figure out the investment but it seems more and more worth it since the latest Deepseek release
👍5🔥4👏1
Signed witcheer
fourteen nights of a benchmark that is supposed to find nothing since 9 july a cron job on the 5090 wakes at 06:45, clones llama.cpp master fresh, builds it for sm_120, runs a smoke benchmark, and goes back to sleep. nobody watches it. fourteen runs so far, across thirteen distinct upstream commits. every one built. every one passed the smoke test. mean 888.22 tok/s range 886.76 - 890.08 (3.32 tok/s, 0.374%…
👍3🔥1
Signed witcheer
Photo, posted without a caption
Signed witcheer
drafter training, epoch end! 8 nights, 54,000 steps, ~55 GPU-hours, every night rc=0, one RTX 5090. epoch 1 of the first EAGLE-3 draft head for Hermes-4.3-36B (Seed-OSS) is complete, and the nightly smoke retires. the release bench takes over: cuda graphs on, 3 configs, 4 workloads, 8 fixed prompts each, baseline 66 tok/s everywhere. • tree-3-4-8: 1.38x prose, 1.38x code, 1.57x repetitive, 1.29x chat. wins every wo…
👍4🔥1
Signed witcheer
nvfp4 on consumer blackwell, round 2: Unsloth's Qwen3.6-27B on one RTX 5090 the model card says "2.5x faster" (no baseline stated) and "runs on 24GB". measured on vLLM 0.21.0, sm_120. ~~~ three walls before the first token: 1. the checkpoint quantises lm_head to FP8; vLLM builds lm_head unquantised and dies on the orphan scale. open upstream, reproduced on v0.22: github.com/vllm-project/vllm/issues/44081. fix: de…
❤3🤝1
Signed witcheer
2-bit KV cache in july 2026: your 4090 can run it, your 5090 cannot. OSCAR (arXiv 2605.17757) stores most of the KV cache at 2 bits, keeps a small bf16 sink and recent window, and hands back ~8x of your KV memory. I spent today trying to run it on the RTX 5090. the gap is kernels, not architecture. sglang's INT2 prefill calls fa3, built for sm80/86/89/90a: A100, 3090, 4090, H100. the 5090 (sm_120) is not on the lis…
👍2
Signed witcheer
Photo, posted without a caption
👍2
Signed witcheer
training a speculative-decoding draft head for Hermes-4.3-36B on one 32GB card EAGLE-3 drafts train on the teacher's hidden states: 3 aux layers + the final one, ~41KB per token at hidden 5120. the standard offline recipe caches those to disk, and 54K samples comes to ~4.5TB. I have 570GB free, so offline is dead on this box. online mode recomputes activations per batch instead: zero storage, teacher resident in VR…
💯1
Signed witcheer
the last modality landed: video with synchronized audio, on one consumer card. LTX-2.3 is Lightricks' ~19B dual-stream DiT: 14B generates video latents, 5B generates audio latents alongside them, so the sound matches the motion. one pass, one mp4 out (h264 + AAC 48kHz). the "22B" in the name is marketing. I ran the distilled two-stage pipeline (8 steps at half resolution, then a 2x upsample + 3 refine steps) on one…
❤1
Signed witcheer
Showing the 12 most recent of 20 posts we hold for @witcheergrimoire. View and reaction counts are the latest single reading for each post, not a live figure, and a recent post is still accumulating both. A view count marked ≈ was rounded by Telegram before we ever saw it — t.me prints views in full below 1,000 and to three significant figures above, so ≈1,200,000 means somewhere between 1,150,000 and 1,249,999. Unmarked counts are exact. Text is reproduced from the public post preview and truncated for length.
Republished by
Channels on the register that have forwarded this channel's posts into their own feed.
Built only from forwarded posts we have actually read, on both sides. Coverage is early and deliberately incomplete: a missing link means we have not read the post that would prove it, never that the relationship does not exist. Counts are distinct forwarded posts observed, so they only ever go up as we read more.
Named by 1 registered channel — every channel on the register whose own posts have named this one, by its current username or any other username it currently holds, merged from two separately captured readings of the same fact so a namer caught by only one of them is not missed and a namer both caught is not counted twice. A username this channel has since dropped is not matched — that handle may belong to someone else now, and crediting today’s namer to yesterday’s owner would misattribute it.
Named by
Channels on the register whose posts name this channel's handle.
A mention is a weaker signal than a forward and is counted separately for that reason — naming a channel is not republishing it, and a handle in a post body is easy to place deliberately. The post counts beside each row below are distinct posts in which the handle appeared, from posts we have read on both sides — the “Named by N registered channels” figure above is a different count, of distinct NAMING CHANNELS rather than posts, and is not the sum of the rows under it.
A live page changes as we take new readings, so a citation should name the measurement it is based on, not just the URL. The line below cites the subscriber count as measured 17 September 2026 — this entry's latest reading, not the date you are reading this.
“witcheer grimoire” (@witcheergrimoire), 231 subscribers as measured 17 September 2026. Telegram Register, tgregister.com/channel/witcheergrimoire.
Full measurement history, CC BY 4.0. Every reading this register holds for this entry, not just the latest one, as a dated, downloadable record: CSV · JSON. Free to use with attribution to tgregister.com. Each file carries its own generation timestamp, which is the figure to cite for exactly when the data was retrieved.